YT YouTube2Text

YouTube2Text → Transcripts → Build & Sell with Claude Code (10+ Hour Course)

Build & Sell with Claude Code (10+ Hour Course) — Transcript

by Nate Herk | AI Automation · 151,204 words · 20,888 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00I'm about to take you from a complete
0:01beginner to a pro cloud code user. Even
0:03if you've never touched the tool before,
0:04by the end of this video, you'll be able
0:06to build automations, websites, apps,
0:08[music] whatever you want. You'll even
0:09have your very own AI executive
0:11assistant. So, I have put a ton of time
0:13into making sure that this course is as
0:14comprehensive as possible, and I've laid
0:16it out in the exact order that I would
0:18have wanted to learn Cloud Code if I was
0:19starting over. So, we've got 24
0:21different chapters that are covered in
0:22this course. Let's take a quick look.
0:23I'm going to start off by telling you
0:25guys about the shift in the Agentic AI
0:27market and why you should be learning
0:28Claude Code. I'm going to help you guys
0:29get set up. We're going to go over the
0:30cloud code operations. We're going to
0:32talk about tokens and context when it
0:34comes to just dealing with AI in
0:35general. We're going to talk about
0:36cloud.md. You're going to build your
0:38first workflows. We're going to deploy
0:40those automations so that they actually
0:41can run 24/7. We'll talk about project
0:43architecture, the built-in commands,
0:45rag, building and deploying websites,
0:47APIs, and MCPs. We'll take a look at the
0:49Google CLI. I'll help you guys build
0:51your very own executive assistant. Then
0:53we're going to deep dive into skills,
0:54sub agents, agent teams, browser
0:56automations, permissions, context
0:58management, GitHub work trees. We've got
1:01some fun hacks for you guys and fun
1:02things that you can do with cloud code.
1:03And then finally talking about how you
1:05can actually monetize this new
1:06knowledge. So I don't want to waste any
1:08time. Let's just get straight into the
1:09course.
1:14All right. All right. So, before I have
1:15you guys open up Cloud Code and we start
1:16getting our hands dirty, I just wanted
1:18to sort of talk about the actual space
1:20and what this shift means and why it's
1:22so important. So, that's what we're
1:23going to be covering in this section.
1:25Check it out.
1:28Aentic workflows are not just [music] a
1:30trend. They're the future of the AI
1:31industry. More and more companies are
1:33making the shift to agentic workflows.
1:34And this is just getting started because
1:36it's estimated that the AI agentic
1:37market is going from about $7 billion
1:39this year to around 93 billion in the
1:41next couple of years. So, I can tell you
1:43right now that knowing how to build aic
1:45workflows is going to be one of the most
1:46valuable skills that you can have. So,
1:48in this video, I'm going to break down
1:49why you should be building aic workflows
1:51and then I'm going to actually build one
1:52live in front of you so you can see
1:54exactly how it works. And by the end,
1:55I'll show you how to actually sell these
1:57if you want to make some money with your
1:58skills. So, let's get into it. So,
1:59before we build anything, I want to show
2:01you why this all matters because it's
2:02not just hype. This is real money moving
2:04into real technology. Right now, the
2:06Aentic AI market is sitting at around $8
2:08billion. By 2030, that's expected to hit
2:1040 to 50 billion. That's not just a
2:12small jump. That's an entire industry
2:13being built in front of our eyes. And
2:15it's not just projections. About 25% of
2:17enterprises are already deploying
2:18Agentic pilots this year. And by 2027,
2:20that number will jump to 50%. So half of
2:23major companies will be running some
2:24version of Agentic Workflows within the
2:26next 2 years. And with that comes
2:27massive budget allocations, new security
2:29requirements, and a ton of new
2:30opportunities for people who know how to
2:32build these systems. So why is this
2:33happening now? What's driving the shift?
2:35It comes down to pretty much one thing
2:36which is companies are starting to hit
2:38that ceiling of what traditional
2:39automation can do and they're starting
2:40to realize they could move a lot faster
2:42with more agentic workflows. If you've
2:44been building workflows in tools like
2:45end to end or Zapier, you know the
2:47drill. You map out every step. You
2:48connect the different nodes or blocks.
2:49You handle the edge cases yourself and
2:51it works until it breaks because
2:52traditional workflows will break when
2:54they hit something unexpected. And when
2:55that happens, someone has to usually go
2:57in manually and fix that. And that's
2:58maintenance. That's time. That's
3:00ultimately money. Now, I do want to be
3:01real with you here because there's a lot
3:02of noise online about a dentic workflows
3:04that makes it sound like they're just
3:05completely magic and they fix themselves
3:07forever. And that is partially true, but
3:09only in a specific context, at least
3:10[music] right now. Cuz when you're
3:11actively working in a tool like Claude
3:13Code and you trigger a workflow yourself
3:15and say, "Hey, go research these
3:16competitors and build me a report." The
3:17agent is sitting right there with you.
3:19So, if something breaks, the agent can
3:20catch it mid-run. It can adjust its
3:22approach. It can update its tools and
3:23keep going. That self-healing piece is
3:25very, very real and it's incredibly
3:27powerful while you're building and while
3:28you're iterating. But once you deploy
3:30that workflow to run on its own on a
3:32schedule or triggered by a web hook or
3:33something like that, that is when you're
3:35deploying the code, you're deploying the
3:36tools, not the actual agent itself. So
3:38if you've seen my previous videos where
3:40we've used the WAT framework, we are
3:42basically deploying the W workflows and
3:44the T tools, but not the A agent. But
3:46I'll cover this more in depth later
3:48during the live build if you're
3:49confused. But what this means is that
3:50the self-healing ability ultimately goes
3:52away when the code is up in the cloud,
3:54you know, running automatically. And at
3:56that point, it does behave more like a
3:57traditional automation. But that's
3:58really a good thing because automations
4:00are predictable. They're deterministic.
4:02And those types are the best ones. So
4:03then where's the real advantage? Really,
4:05it sits in how you build. Traditional
4:07automation is like building a train
4:08track by hand. You're laying every rail,
4:10every switch, every connection all by
4:12yourself. Whereas with aentic workflows,
4:13it's like you're just telling a
4:14construction crew, "Hey, I need you to
4:16build a train track from here to there."
4:17And then they build it for you. Meaning,
4:19if they hit a problem during
4:20construction, they would figure it out.
4:21So you end up with a better train track.
4:23It's built faster with fewer mistakes
4:25because the agent handled the edge cases
4:26during the build process that you might
4:28have missed or not thought of. And then
4:30the idea is you battle test it before
4:31you ever actually deploy it. So then you
4:33have a lot of confidence that it will
4:34always work. So in our train analogy,
4:36before we deploy that train track, we
4:38would have like 10 different types of
4:39trains test drive on it. They would be
4:41different weights, different lengths,
4:42and maybe different wheels. And we'd
4:43want to make sure that our track can
4:45work for all different types of trains
4:46before we deploy it. And the reason this
4:48is actually possible now is because the
4:49technology has finally caught up. LM
4:51have gotten really reliable enough to
4:52use in production and we're not just
4:54playing around with chatbots anymore.
4:55These models can reason, they can make
4:57decisions and they can execute
4:58multi-step tasks with real consistency.
5:00On top of that, we've got things to use
5:01like skills or MCP or aa. We've also got
5:04infrastructure like trigger.dev, modal
5:06or versell that make deployment way
5:08simpler than it used to be. And most
5:09importantly, we've got tools like cloud
5:11code that make all of this accessible to
5:13non-developers. So, we can see that the
5:14market is absolutely shifting towards
5:16aic systems and the numbers back it up.
5:18But here's a question that's probably on
5:19your mind. Does this mean everything
5:20that I've learned about Naden or
5:22traditional automation is useless? Not
5:24even close. And let me explain why. The
5:26people who are going to struggle with
5:27agentic workflows are the ones who
5:28completely skip those fundamentals and
5:30jump straight into cloud code, tell it
5:31to build something, and have no idea if
5:33what is being built is actually good.
5:35They probably don't know what a web hook
5:36is or how APIs work. They won't be able
5:38to spot when the agent made a bad
5:39decision because they've never built the
5:40thing manually themselves. Now, that's
5:42not to say that a beginner can't learn
5:43cloud code. And because you actually
5:45understand how automations work under
5:46the hood, you can communicate precisely
5:48what you want much clearer. And you'll
5:50see once we hop into the live build just
5:51how important it is to actually be able
5:53to communicate really clearly. All
5:54right, so now you understand why agentic
5:56workflows are such a big deal. Now I'm
5:57going to show you how simple it is to
5:59build an agentic workflow using cloud
6:00code. And by the way, if you want to
6:02follow along with what I'm about to
6:03build, you can grab all the resources
6:04and files that you need completely for
6:06free in my free school community. The
6:08link for that is down in the
6:09description. All right, so we're now in
6:11Visual Studio Code, which is where we're
6:12going to actually use Cloud Code. Visual
6:14Studio Code is just an IDE or an
6:16integrated development environment. And
6:18that's where, like I said, we're going
6:18to be using Cloud Code. So, if you don't
6:20have this up and running yet, all you
6:21have to do is go to a browser and type
6:23in Visual Studio Code and download the
6:25right one for your operating system. And
6:26then once you open that up, it should
6:27look like this. The first thing you have
6:28to do is install the Cloud Code
6:30extension. So, you're going to go over
6:32here to the lefth hand side. You're
6:33going to open up extensions. And then
6:34you will see right here cloud code. Or
6:36if you don't, you will search Cloud
6:37Code. And once you pull that open, it'll
6:39ask you to install it. And then once you
6:41do, it'll basically just prompt you to
6:42sign in with your Claude subscription.
6:44Now, you do have to be on a paid plan
6:45for Claude in order to access Cloud
6:47Code. As you can see on the free
6:48version, you don't get Claude Code, but
6:50right here on the paid version, you do.
6:52So, you can start off with the pro plan
6:53at 17 bucks a month. And then, if you
6:56need to, if you keep hitting limits, you
6:57can upgrade to max, which honestly, you
6:59probably will. I'm on the max plan, and
7:00it is an amazing return on my
7:02investment. So, I would just go with the
7:03max. But anyways, you'll get
7:04authenticated, and then it will bring
7:06you back here. And now we can actually
7:07get going using cloud code and building
7:09workflows. So I'm going to close out of
7:10this screen. And what we're going to do
7:12now is we're going to open up a project.
7:14So on this lefth hand side, I'm going to
7:15go up here to explore. And this says you
7:17have not yet opened a folder. Open a
7:18folder. So essentially when we're in
7:19cloud code, we're going to be working
7:21within a certain folder. And that's kind
7:23of the way that I think of like this is
7:24the project that we're working on. So
7:25I'm going to click open folder and I'm
7:26going to open up a blank project. So you
7:28can see I'm in a folder called
7:30newsletters demo. And there's nothing in
7:31it. It's completely fresh. I'm going to
7:33click select folder. And now we can see
7:34we're in this project. I'm real quick
7:36just going to close out of this and
7:37close out of this just so we have a
7:39really clean interface to look at with
7:41not much going on and I can explain what
7:42we're actually about to do. So to make
7:44this as simple as possible in cloud code
7:45we have an agent and we have files.
7:47That's it. The lefth hand side is where
7:48we see those files. We'll see different
7:49workflows. We'll see tools. We'll see
7:51all these little things. And then on the
7:52right hand side we'll have the cloud
7:53code agent and that's where we talk to
7:55it. We plan with it. It asks us
7:56questions and it actually executes and
7:58writes the code or builds the workflow
8:00for us. So, if I switch back into Visual
8:01Studio Code and I double click right
8:03here and then I open up this button that
8:05says Claude Code, this is where we
8:06actually open up the actual Claude Code
8:08agent right there. So, I'll close out of
8:09this. You can see this is kind of what
8:10we're talking about now. Files on this
8:12side. There's nothing there yet. And the
8:13Claude Code agent right here. Now, what
8:14we have to do next is give Claude Code a
8:16cla.md file, which is basically just
8:18instructions for this specific project.
8:20And you can really just think of this as
8:21a system prompt. So that way when we the
8:24user send a message to our cloud code
8:25agent, it doesn't just process what we
8:27said and respond to us, but it also
8:28every time reads the claw.md file. So
8:30this is where you're going to put
8:31important things like how the folders
8:33are laid out. You know where to find
8:35your different files, what its end goal
8:36is, any frameworks that you might be
8:38using. So in this case, what we're going
8:39to be doing is we're going to be using a
8:40framework called WAT, which stands for
8:42workflows, agent, and tools. So, real
8:44quick, if you pop over to my free
8:46community and you go to the classroom
8:47and then you click on cloud code right
8:49here, you'll see the watclaw.mmd and you
8:51can go ahead and download this file
8:52right here. And once you've downloaded
8:53that file, you can actually just drag it
8:55over here to the lefth hand side and it
8:57should pop up as cloud.md. And if you
8:59wanted to, you could read through this
9:00entire basically system prompt to see
9:02what I'm telling it about how to build
9:03workflows, how to build tools, how to
9:05keep learning and how to, you know, set
9:07up its its folders and everything. But
9:09I'm not going to read that all out right
9:10now. What I'm going to do is just
9:11basically tell Claude Code to set up the
9:13project. read the claw.mmd file and then
9:15set up the project and the structure and
9:18then we'll start building workflows
9:19together. So I'm basically just going to
9:20shoot that off and it's going to go
9:21ahead and read that and get everything
9:22ready. So we'll see soon on the lefth
9:24hand side we've got all our different
9:25folders set up. But while it's going
9:26through and doing that let me just
9:27explain what these different things are.
9:30So agent that is the actual cloud code
9:32agent that we just talked to as you saw
9:33and the agent utilizes workflows and
9:35tools to help us automate things. So the
9:37first thing is workflows. These are
9:39markdown files which you just saw
9:41similar to the cloudMD and it looks like
9:43this. It's basically completely natural
9:44language. You could read through every
9:45line and understand exactly what's going
9:47on. It just uses things like pound signs
9:49and, you know, dashes and asterisks to
9:52separate like what's a header and what's
9:54bold to stress importance for the agent.
9:56Workflows are natural language
9:58processes, instructions. So, right now,
10:01let's just use an analogy of a recipe.
10:02The workflow is the recipe. So, you'd
10:04have a workflow for how to bake a
10:06chocolate cake. And when you want to
10:07bake that chocolate cake, it's going to
10:09tell you what to do in certain order.
10:11So, it's going to say, "Preheat the oven
10:12to this. Boil some water." I don't know
10:13why you'd boil water for a cake. Crack
10:15two eggs in a bowl, you know, measure
10:16out a cup of flour, whatever it is. And
10:18those are the tools. So, the tools are
10:19all of the ingredients, but without the
10:21structure of the workflow, saying use
10:23tool one, then tool 5, then tool 7, then
10:25tool 10. Without the order and the
10:27structure, the tools are useless. So,
10:29basically, the workflows tell the agent
10:30how to build the tools. And what's
10:31really cool about both of these is as
10:33they're being built and as they're being
10:34used, the agent will improve them over
10:36time if it makes mistakes or if it
10:38learns things. So that is why we use the
10:39WAT framework to build our workflows
10:41with cloud code. So now that that's
10:43done, you can see that this is finished
10:44up and it's basically said, okay, the
10:46project is set up. Here's a summary of
10:47the structure. Here's what I understand.
10:48We're going to be building workflows in
10:50this project likely around newsletter
10:51operations. WAT framework. I will act as
10:53the agent. I will read workflows. I will
10:55run tools. I will handle errors and
10:56improve the system. I've got Python
10:58ready to go. And I'm going to store
10:59secrets in thev. So this is where we're
11:02going to put our API keys rather than
11:04putting them straight into claw so that
11:05they could be exposed who knows where.
11:07Okay. So, what I'm going to do is go
11:08ahead and do a /cle just to get rid of
11:10this conversation and we can start
11:12fresh. And we're going to start to plan
11:14out this workflow that we want to build.
11:15Before we start planning, I'm going to
11:16switch this to plan mode. So, you can
11:18see we're in bypass permissions. You can
11:20go to ask before edits. You can go to
11:21edit automatically. But I want to go to
11:22plan mode. And it's really, really
11:24important, as we talked about earlier,
11:25to be able to communicate clearly what
11:27you want. And the cool thing about Cloud
11:28Code is when we give it a plan, even if
11:30it's pretty ambiguous, it will come back
11:31and say, "Okay, in order for this to be
11:32good, I need to know X, Y, and Z." So,
11:34I'm going to give it a fairly ambiguous
11:36prompt here. And then you're going to
11:37see it ask us questions and plan out
11:38this workflow for us. Hey Claude, I want
11:40to build a workflow which will basically
11:42be a newsletter automation. I want to be
11:44able to tell you that I need a
11:45newsletter about a certain topic and you
11:47will do research. You will structure it
11:50in HTML. You will make it look pretty
11:52and you will also create a few
11:53infographics to go with it. So help me
11:56figure out what text stack to use here
11:57and what else you might suggest that I
11:59haven't yet thought of. So I'm going to
12:00shoot that off. Now, whether we're in
12:02plan mode or, you know, bypass
12:03permissions, what happens is the agent
12:05starts thinking and it starts testing
12:06things out. So, it's thinking, it's
12:08reading through files, it has this
12:10little thing that will say computing or
12:11deciphering or wobbling or whatever it
12:13is, just a bunch of little silly words.
12:14Um, but that just basically shows you
12:16exactly what it's doing. All right, so
12:17we're just at the point where it's
12:18asking us some questions before it
12:20continues working on the plan. The first
12:21thing is for research, do you want to
12:22add an external search API to pull in
12:25data? So, what I'm going to do is say,
12:26yeah, sure. Let's just go ahead and do
12:28perplexity. for delivery. It asks us if
12:29we want to use Beehive or if we just
12:31want to send the HTML file for now. I'm
12:33actually just going to go ahead and say
12:34let's actually just send this over in
12:36Gmail. And now it asks us about brand
12:38assets, which is really cool. So, if we
12:40want to, we can send over some brand
12:42guidelines or logos and stuff like that
12:43to make sure that the newsletters are
12:45always formatted and they feel on brand.
12:47So, I'm going to go ahead and say yes, I
12:49will provide some brand assets. So, then
12:51what happens is it comes back with a
12:52final plan. Let me just zoom out a
12:54little bit so we can actually see that a
12:55little bit better. We'll go ahead and
12:56see what it came up with. So, newsletter
12:57automation workflow. We want to conduct
12:59research, generate HTML with polished
13:02visual design, create infographics to
13:03accompany the content. We've got a
13:05research layer. We've got the content
13:06generation. We got the infographics. So,
13:08it says that it could use data stats
13:10infographics or it could use SVG. Why
13:13not image generation? It's too
13:14unpredictable. It can't embed. I'm
13:16actually going to go ahead and say that
13:17I wanted to use nanobanana. So, I'm
13:19going to go ahead and type in here for
13:20the infographics. I want you to generate
13:22AI images using nano banana. You can use
13:24a platform called key.ai. So this just
13:26shows the importance of plan mode and
13:28reading through the plan so that you can
13:29make sure before it actually starts
13:31building everything you like what it's
13:32going to do. Okay. So now that new plan
13:34has been done you can see the text stack
13:36is going to be research with perplexity.
13:38The content will be written with claude.
13:39The infographics will be created with
13:41nano banana. We will write the email in
13:43HTML and then send that via Gmail. And
13:45it even comes up with a section here
13:47about things the user likely hasn't
13:49considered. So things like human review,
13:51subject line, metadata, brand
13:52consistency, all this type of stuff.
13:54Now, the last thing that I actually
13:55forgot to give it was my brand assets.
13:57So, what I'm going to do real quick
13:58before we accept or, you know, keep
14:00working on the plan, I'm going to create
14:01a new folder over here. I'm going to
14:02call this brand_assets.
14:05And you can see I dragged in two things.
14:06I dragged in AISPG, which is our logo,
14:09and I dragged in our brand guidelines.
14:10So, I want the newsletter to be
14:12formatted in this way. So, I'm going to
14:13click on no key planning, and I'm going
14:14to tell Claude that it needs to use
14:16those two assets. So, what's cool is
14:18that I can actually directly tag them.
14:20So I'm saying make sure the whole
14:21newsletter is branded based on my logo
14:23and brand guidelines. So for logo I'm
14:24going to do at and I'm going to type in
14:26AIS. You can see that it's going to show
14:27AIS PNG. And then here I can do at AIS
14:30and I'm going to click on brand
14:31guidelines. So now it's going to look at
14:33those exact two things and it's going to
14:35be able to make sure that the newsletter
14:36is branded. Okay. So this time the plan
14:38looks good to go and I'm just going to
14:39go ahead and auto accept and I'm going
14:40to turn on bypass permissions. So it's
14:42going to build everything out. It's
14:44going to put the different files that we
14:45need and then we should be able to
14:46basically just add our API keys and then
14:48test it. So, what it's doing now is it
14:49creates a to-do list. So, this is all of
14:51the things that it has to do and as it
14:53actually completes them, it crosses them
14:54out. So, it's really cool because you
14:55can work on something else on a
14:57different screen and just kind of check
14:58in on cloud code to see where it's at
15:00and if it needs any help. Now, you guys
15:01may be wondering about this bypass
15:03permissions mode. If you don't see this,
15:05you just have to go to your settings. In
15:06your settings, search for cloud code.
15:08And then right here, you'll be able to
15:09see allow dangerously skip permissions,
15:11which turns on allow bypass permissions
15:13mode. Okay, so that is finished up. It's
15:15telling us here's what it built. So it
15:17created two config files which we can
15:19see right here. It's got newsletter
15:20style which basically just shows like
15:22the colors and the text and the
15:23background and it's got recipients which
15:24is where we need to add who this is
15:26actually being sent to. So this is where
15:27we would add a huge list of you know our
15:29email list basically. Then it created 1
15:312 3 4 five different tools. If I open up
15:33those right here the tools are research,
15:35generate infographic, assemble HTML,
15:37send via Gmail and archive to sheets.
15:39And here is what all of those five tools
15:41do. And then of course we have the
15:43actual workflow right here which is our
15:44markdown file which basically shows step
15:46by step how to actually build the
15:48newsletter and what tools to use. And
15:50this is the complete natural language
15:51just explaining the process. So now that
15:53those have all been created the last
15:55thing that we have to do before we
15:56actually test it out is we have to give
15:57it credentials. So anthropic perplexity
15:59key.ai and then our Gmail. So what I'd
16:02do is I go to Plexity grab my API key. I
16:04would come into thev and then over here
16:06it's created these placeholders. So all
16:07that I have to do is paste in my API key
16:09right there and then hit save to make
16:11sure that all this saves. So I'm going
16:12to go ahead and do this now for my other
16:13API keys. Okay, so we've done everything
16:15that it told us to do. We've set up all
16:16our credentials. At least I hope we
16:18have. If we run into any errors, Cloud
16:19Code should fix it or tell us what to
16:21do. What I'm going to do now is just
16:22kick off a prompt. Write me a newsletter
16:24about Aentic AI. So I literally just
16:27said, write me a newsletter about Aentic
16:29AI. And that's it. What it's doing now
16:30is it's looking through the relevant
16:32workflows and tools, and it's going to
16:33figure out what to do. Here you can see
16:34it said, "I found the newsletter
16:36workflow. Starting with step one, I'm
16:37going to do some research." You can see
16:39after that it's going to plan and
16:40generate the infographics. It's going to
16:41write the newsletter content and then we
16:43actually have a human review point. So
16:44it's going to get subject line approval
16:46and then if it's approved, it'll go
16:47ahead and send the newsletter. So now
16:49your job at this point is just to watch
16:50it and to make sure it's doing
16:51everything right. And if it runs into
16:53issues, it should fix itself. But
16:54sometimes it may need you to help steer
16:56it in the right direction. The first
16:57test run is the only one where it's
16:58really like this because you have to see
17:00how it works. But then after that, you
17:01should be able to just trust that it's
17:02going to run pretty much perfectly every
17:04time. Here you can see we've already run
17:05into our first issue. There was a unic
17:06code encoding issue, but it's just going
17:08to go ahead and fix it. And that's great
17:09because I don't really know what this
17:10means at all. So I'm glad that it
17:12understands what to do. Nice. So you can
17:13see it planned out three infographics.
17:15It's got a market growth. It's got
17:16Gartner road map and it's got impact
17:18metrics. So here's a good example. It
17:20was trying to generate those
17:22infographics using key.AI, and it was
17:24getting an error. So what it did is it
17:25looked into the problem. It said, "Let
17:27me investigate to find the correct
17:28endpoint." It did some web searching. It
17:30looked through the docs. It did multiple
17:31searches as you can see. And it figured
17:33out that the endpoints have changed. And
17:34now it's able to switch the tool so that
17:36it works this time. There we go. So, it
17:38said, "I found the fix. Here's the right
17:39endpoint. Let me update the tool so that
17:41this doesn't happen again." And now it
17:42just went ahead and fixed the tool. All
17:44right. At this point, it did a human
17:45review step. And we could obviously say
17:47we don't want this if we don't want it.
17:48But for now, let's just go ahead and see
17:49what it wants. It wants us to approve a
17:51subject line. It asks us to choose which
17:53one. I'll go ahead and send five. And
17:54then we will see the final output. Okay.
17:57So, a few things happened and I'm
17:58actually glad they did so I can show you
17:59how you need to troubleshoot this. So,
18:01the first thing is we got the email, but
18:02the HTML is all messed up. It came
18:04through with a background color, but
18:05then all of this just is horrible. So,
18:07we're going to have it fix this. The
18:08second thing is I gave it the wrong
18:10Google sheet ID to archive to sheets
18:12because there was some sort of access
18:13issue. So, I'm going to go ahead and fix
18:15that sheet ID and I'm just going to use
18:16my natural language to tell it that this
18:18is horrible. I've updated the sheet ID.
18:20However, the actual email that I
18:22received is completely awful. I can't
18:24read any of it. It doesn't even make any
18:26sense. take a look at figure out what
18:27happened and try to send me it again.
18:29So, it's going to go ahead and diagnose
18:31what happened and then hopefully send us
18:32a better version. So, once again, it
18:34found the issue. It found out exactly
18:35how to fix it and now it's updating the
18:37workflow in the tool so it doesn't
18:38happen again. Now, of course, cloud
18:39code's not perfect. You guys can see
18:40that in this demo, but think about if
18:42you were doing this in something else
18:43like ended or something that's a bit
18:44more manual and you were running into
18:46these issues and you'd have to go back
18:47and fix all of the logic yourself and
18:49try to debug all this. I've literally
18:50just been telling it to fix it and then
18:52like doing other work or going in the
18:53other room and waiting for it to figure
18:55it out on its own. Okay, now it fixed
18:56everything. And if I go over to my
18:58email, we see this newsletter. What
19:00happens when AI stops waiting for
19:01instructions? We can see that we've got
19:02our logo up top. We've got AIS
19:04intelligence brief. It does think that
19:05it's June 2026, which is wrong, but we
19:07could obviously fix that very easily.
19:08But now we move into the actual
19:10newsletter. And keep in mind, this
19:11started with one prompt that said,
19:13"Write me a newsletter about Aentkai."
19:15That was it. Also, throughout the
19:16newsletter, pay attention to the fact
19:17that it's using our fonts. It's using
19:19our brand guidelines, our colors, all of
19:20that in this newsletter. So, the first
19:22section is about the market landscape,
19:23an explosion that cannot be ignored. I'm
19:25not going to go ahead and read all of
19:26this text. It would just take too long.
19:28A nanobanana AI generated image with
19:30text with graphics. And this infographic
19:32is also adhering to our brand
19:34guidelines. In section two, we have
19:35architecture. We've got a little bit of
19:36a quote here. And if we keep scrolling
19:38down, we've got some more statistics.
19:39We've got section three. We've got
19:41another quote. And we have another
19:42infographic. Once again, adhering to our
19:44brand guidelines and using a little logo
19:46up here as well. And that's pretty much
19:48how the rest of the newsletter goes.
19:49We've got section four. And we can see
19:50our third and final infographic that has
19:52a different version of the AIS logo as
19:54well as our brand guidelines. So this
19:56was literally iteration one. There's a
19:58lot of things that we can improve here.
19:59And all we would do is we'd open up
20:00Cloud Code and we'd ask it to make it
20:02better using natural language. We could
20:03actually make sure that every
20:04infographic it creates uses our actual
20:06logo rather than prompting some sort of
20:08AIS logo in there. It ends with some key
20:10takeaways and then we have it ends with
20:12some key takeaways. We have a call to
20:14action down here and then all of the
20:16sources we could actually click on and
20:17it would take us to that actual site
20:19where it pulled the data from. So that
20:21was version one of the newsletter and I
20:23think that that's pretty solid. Now the
20:24cool thing about these projects in cloud
20:25code is as you use them more they get
20:27better and better because every time I
20:28run this workflow it might find
20:30something else out and it will update
20:32its cloud MD, it'll update its
20:33workflows. It'll update its tools. As I
20:34give it more brand asset, as I give it
20:36more context and more knowledge, it just
20:38gets better and better. And then once
20:39you really trust the actual workflows
20:41and tools, that's when you go ahead and
20:43you come back into this. And then once
20:44you really trust the workflows and the
20:46tools that you've created using cloud
20:47code, you would basically take these two
20:48things and you would push those into
20:50like a GitHub repository and you'd sync
20:52those to something like trigger.dev or
20:53modal in order to actually have them run
20:55every single Monday at 6 a.m. or daily,
20:57something like that. I'm not going to
20:58dive into that in this video, but if you
21:00want to see one where I did, then I'll
21:01tag one right up here. So, what you guys
21:03just saw here was me using hardly any
21:05prompting, just using my natural
21:06language, giving it a few logos and
21:08colors, and then giving us a really,
21:10really good output for a newsletter.
21:12Now, one thing that we didn't cover in
21:13this video, but we will be covering a
21:14lot more in the future, is how you could
21:16actually make your workflows even better
21:17and better, and that's the idea of using
21:19skills. Whether it is a skill that you
21:21create yourself or whether it's a skill
21:22that someone else has already built. So,
21:24skills are basically just system prompts
21:25that you could load in when you need
21:27them. So, let's say you ask Cloud for
21:28help. Hey, can you design me a website?
21:30The agent will then check through all of
21:32the skills it has access to and it will
21:33see based on all of these skills, does
21:35my current request require this? So,
21:37it's almost like the same way it decides
21:38if it should use a tool or not. So, for
21:40example, there is a front-end design
21:41skill that makes Claude Code so much
21:43better at designing websites. And so, if
21:45I'm ever building a project where I need
21:47it to be able to build websites, I would
21:48tell it to always invoke the front-end
21:50design skill. And the reason I'm
21:51bringing this up is because you can
21:52create your own. So, what I might do in
21:54this version is once I realize what I
21:56really like about how it creates
21:57newsletters, I will tell it to turn that
21:59into a skill. So maybe it is the skill
22:01of making infographics look really
22:03really polished with the AIS logo in the
22:05top left corner and I could create that
22:07skill so that every time it needs to
22:08create a new infographic it reads that
22:10first and then it makes its outputs a
22:12lot more consistent. So I know that this
22:13seems a little bit intimidating at first
22:14but hopefully you guys realize after
22:16watching this how easy it was for me to
22:17actually do this once again with hardly
22:19any technical knowledge. We didn't set
22:20up any API calls. We didn't do anything
22:22like that. We just talked to it. But now
22:23the question is how do you actually turn
22:25a skill like this into income? So this
22:27is something that I see all the time. a
22:28business owner watches your YouTube
22:30videos or LinkedIn posts or whatever it
22:31is and sees flashy AI demo. Maybe that's
22:33a voice agent or a really cool chatbot
22:35or a crazy looking dashboard and they
22:37come to you or some sort of like, you
22:39know, AI agency and they say, "I want
22:40that." But when you actually sit down
22:42and you look at their business and their
22:43operations, that's not what they need at
22:44all. The real problem is that leads
22:46might be falling through the cracks or
22:47the onboarding is taking way too long or
22:49there's tons of manual data entry going
22:50on. Just think about it like plumbing.
22:52If you have a pipe that's clogged, it
22:53doesn't matter how much water you pour
22:55into it, it's not going to flow any
22:56faster if there's a clog. Most
22:57businesses are out here trying to put as
22:59much water into the pipe as possible,
23:01hiring more people, throwing AI at
23:02random problems. But what they actually
23:03need is someone who can come in, find
23:05the clog, and then clear that, and then
23:07start to add more water in. That's
23:08really the skill. And if you can cut
23:09through the noise and identify real
23:11constraints and unclog that pipe, that's
23:12worth way more than building some super
23:14flashy agent that looks cool, but
23:15doesn't actually move the needle. The
23:17build itself is also not what businesses
23:19are paying for because building is
23:21getting easier and easier every day,
23:22which is good news, but it also kind of
23:24brings about some panic because more
23:25people can spin up these automations
23:26much quicker and that's becoming a
23:28little bit more commoditized. So, if
23:29you're trying to compete on I can build
23:31AI automations, you're going to be in a
23:33race to the bottom. What you need to do
23:34is act as the doctor, not the
23:36pharmacist. I've used this analogy a lot
23:37on my channel. A pharmacist just fills a
23:39prescription that someone else wrote,
23:40but a doctor sits down with the patient,
23:42asks questions, runs diagnostics, and
23:44figures out what's actually wrong before
23:46anything is then prescribed. That's the
23:48difference between someone who just
23:49builds workflows and someone that
23:50businesses will pay serious money to
23:52work with. So, when you're talking to a
23:53business owner, you're not leading with
23:55I build aic workflows in cloud code.
23:57They don't care about that. You're
23:58leading with I can save you x amount of
23:59time per month. You're leading with I
24:01can save this process x percentage of
24:04errors. And that's exactly why you
24:05should not be pricing yourself hourly.
24:07Because if you can build something in 30
24:08minutes, that ends up saving the
24:09business, let's just say, 20 hours a
24:11week, that's not a 30-minute job. That's
24:12tens of thousands of dollars in value
24:14over the course of a year. So if you
24:16price yourself at an hourly rate, you're
24:17putting a ceiling on your income and
24:19you're completely ignoring the value
24:20that you're actually delivering. Now,
24:21hourly can be fine early on when you're
24:23just getting started and you're building
24:24trust and you need to get your first few
24:26wins. But once you can clearly show the
24:27ROI, the hours saved, the cost
24:29eliminated, the revenue generated, all
24:31that kind of stuff, then your pricing
24:32should really be reflecting that value,
24:34not your time. Trading time for money is
24:36not very scalable. So, here's a simple
24:37way to think about it. You sit down with
24:38a client and you figure out their
24:39processes, and you calculate that this
24:41system is going to save them $10,000 a
24:43month. Now, let's say you charge $5,000
24:45for that build. That should be a
24:46no-brainer for them. They make their
24:47money back in 2 weeks, and then
24:48everything else is just profit for the
24:50business. And it's also a great deal for
24:51you because that build might just have
24:53taken you a few days, maybe a few weeks.
24:55That is basically valuebased pricing.
24:57Everybody wins. Now, in terms of
24:58actually finding clients, I've done a
25:00full deep dive on that in another video,
25:01which I will go ahead and link right up
25:03here. But at a high level, the approach
25:04is simple. You don't need a huge
25:05audience. You don't need to start a
25:06full-blown agency. You just need to
25:07start conversations with the right
25:08people. You need to be transparent about
25:10what you're building and lead with how
25:11you can help them. Once you deliver the
25:12solution, you stick around because once
25:14that first system is running and they
25:15see the results, they're going to want
25:16more. They're going to want you to
25:17optimize the build. They're going to
25:18want you to expand on it. They're going
25:20to want you to help find new
25:21opportunities inside their business.
25:22That's how a $3,000 build turns into a
25:24$50,000 a year relationship. But the key
25:26there is that you have to be the one to
25:29track the metrics. You have to take
25:30ownership over that. You have to
25:31practively show them the value that the
25:32system is actually adding. That's super
25:34super important. And that's exactly the
25:35path freelancer to consultant to trusted
25:37partner. You're not just building
25:38workflows. You're becoming the person
25:40businesses rely on to make their
25:41operations smarter. So we just went from
25:43understanding what's happening in the
25:44agentic workflow market to actually
25:46building one live and seeing how to sell
25:48these systems for premium prices. Here's
25:49the thing though. This isn't just the
25:51future of automation. It's happening
25:52right now. Companies are already making
25:53the shift and the demand for people who
25:55can build these systems is only going to
25:57grow. So, if you want to dive deeper
25:58into this kind of stuff, I've got a
25:59community with over a quarter million
26:00members where I share templates,
26:02resources, and all the files from videos
26:04just like this one.
26:08Okay, so now we're going to start to get
26:09into it a little bit. So, the first
26:11thing I want to talk to you guys about
26:12is the different ways that you can
26:13actually use Cloud Code so that I can
26:15help you figure out which way is best
26:17for you. And then we're going to move
26:18into some foundational concepts like
26:20prompting, AI models, you know, tokens,
26:22context, windows, cloud.mmd, what does
26:24all that stuff mean. So that's what
26:26we've got coming in this chapter.
26:30One of the most common questions I've
26:32been getting lately is where should I be
26:33running cloud code? Whether that's in
26:35anti-gravity or VS Code or desktop app
26:37or in the terminal. And it's important
26:38to understand that they're all a little
26:39bit different, but they're all a little
26:41bit the same. So in this video, I'm
26:42going to break down the five best
26:43methods. For each of those, I'm going to
26:45show you what it looks like, what it
26:46does, the pros and cons, how you can get
26:48it set up, and who it's actually for, so
26:50that by the end of this video, you'll
26:51100% know which way that you should be
26:53running Claude Code. So, let's not waste
26:55any time and get straight into number
26:56one. All right, so number one, we have
26:59running Claude Code in the terminal.
27:00When you run in the terminal, this is
27:02typically what it looks like. And it may
27:03feel a little bit more intimidating if
27:04you've never used the terminal before,
27:06but it's important to understand that
27:07this is the foundation. So, even if you
27:09never used this directly, you need to
27:10understand that this exists because it's
27:12kind of the core. You open the terminal,
27:14you type in Claude, and then you
27:15instantly have Claude right there, and
27:17you start chatting. It can read the
27:18files. It can edit the folders that
27:20you're in and do work like Cloud Code
27:22should. Everything though is pretty much
27:24text. There's no buttons. There's no
27:26menus. It's not very visual. It's just
27:28text. It is a CLI. And the engine that
27:30powers this, every other Surface uses as
27:33well. So, the desktop app, the VS Code
27:35extension, they're all wrapping around
27:36the same engine. And moving on to pros
27:38and cons. The pros are that you have the
27:41most control. So running in the
27:42terminal, you're going to have the most
27:43amount of commands and the most amount
27:45of like hackability. A lot of times also
27:47features are coming first to the
27:49terminal and it works with any editor
27:50because it just needs a terminal window,
27:52which means if I'm using something like
27:53VS Code, I can use the extension if I
27:55want something more graphical and I can
27:57use the terminal in there if I want more
27:59power. The cons are that it's text only.
28:00So it's sometimes not great to
28:02understand like what's going on behind
28:04the scenes or what's going on with your
28:06files and folders and a bit of a steeper
28:08learning curve if you're not comfortable
28:09in a terminal. For instance, I don't
28:10love working in the terminal, but now I
28:12can do it and I understand it and
28:13sometimes I have to. So, who is this
28:15for? For people that already live in the
28:17terminal. They understand it. This is
28:18their home. If the word terminal makes
28:20you nervous, then just keep watching
28:21because there are other options that let
28:23you use cloud code and still achieve the
28:26power you need. So, real quick, I just
28:27wanted to show you what this could look
28:28like. I'm opening up my terminal that's
28:31going to pop up right here. You can see
28:32that right now I am in my kind of home
28:34directory in Nate H. If I wanted to get
28:37into a different project, I would just
28:38need to do a CD to switch into that
28:40folder. And so that's what I mean by
28:42command line interface. We switch into a
28:44folder with text rather than if I pulled
28:46up like my actual file explorer rather
28:48than being able to just switch folders
28:49over here or clicking on different
28:51buttons. So that's kind of the
28:52difference between a GUI and a CLI. But
28:54all you have to do to install it is go
28:55to cloud code docs and you can see this
28:57quick start and it basically just says
28:59make sure you have this stuff. And so
29:00you do need a paid cloud subscription.
29:02So either the pro, max, teams or
29:04enterprise. And then you basically just
29:06have to run one command. So whether
29:08you're Mac OS, Linux, Windows
29:10PowerShell, Windows command line, you
29:12just run this command and then it will
29:13take you through some onboarding. So it
29:14will say, "Hey, can you please log into
29:16your account? Can you do this? Can you
29:18acknowledge that this is blah blah
29:19blah?" And here's where you can see you
29:20basically would CD into your project.
29:22And then all you have to do is open up
29:24your terminal and type in Claude. And
29:26now after I trust this workspace, I have
29:28Claude and I can say hello and I can
29:30talk to it the same way that I would in
29:31anything else. I can hit question mark
29:33to see the shortcuts. I can start doing
29:35slash commands if I want and see front
29:37end design skill or I can change the
29:38model or I can even do something like
29:41ultrathink which you can't really see in
29:44other ways. And like I said, it just has
29:45the most power because I can also do
29:47something like customizing my status
29:48line, which means that I could basically
29:50customize what I see down here. So I
29:51could see the cloud info like model
29:53name, context usage, or I could even
29:55describe other information that I want
29:56to see about this session. And then if I
29:58want to exit out the session, I can just
29:59hit C twice and cloud code has now been
30:03terminated. And now moving on to number
30:04two, we have the desktop app. So this is
30:07what it looks like. In the desktop app,
30:08you have chat, you have co-work, and you
30:10have code. So if you're already using
30:11the desktop app for these things, then
30:13you're probably really used to the way
30:14that this looks and feels. And you also
30:16get a cool little animated crab that
30:18will do funny things right there. So
30:19this is a visual GUI first experience
30:21for people who don't want to touch a
30:23terminal. It's the same cloud code
30:24engine, of course, but it's just wrapped
30:26in a different type of interface with
30:27buttons and panels and visual feedback.
30:30So you basically just Google, how do I
30:31install cloud desktop? It's going to go
30:33over some system requirements and then
30:34you go to the downloads page and then
30:36you basically just go through the setup
30:37in order to actually download this for
30:39your operating system. So, what are the
30:40pros and cons? Well, we have visual
30:42differences because you can see the
30:43changes that Claude made line by line
30:45and accept to reject them. It's just a
30:46bit easier to actually tell what's going
30:48on in my mind. They also have a really
30:49cool built-in app preview. So, in the
30:51desktop version, if you're building an
30:52app or something like that, it'll start
30:53up the dev server and you can see it
30:55right in the app itself, which is kind
30:56of a cool little feature. You can have
30:58multiple sessions running in parallel,
30:59which you can do in all of these, but
31:01it's just a little bit different because
31:02these get isolated in their own branch
31:04and you can click through them pretty
31:05easily. And it probably is the easiest
31:07and least intimidating for non- terminal
31:09type of people. And you still get the
31:10same local file access. Now, for the
31:12cons, this is a bit more of a manage
31:14experience. It's less customizable. It's
31:16less hackable. The desktop releases can
31:18sometimes lag behind the CLI. And right
31:21now, it's just Mac and Windows only.
31:23Okay. So, here is typical Claude that
31:26you would see, but I've got chat, I've
31:27got co-work, and then I've also got
31:29Claude code. And so, this is what it
31:31looks like in the desktop app. You can
31:33see that we still have our same
31:34conversation feel. Up in the top right
31:36is where you can see we could start our
31:37actual preview if we have, you know,
31:39some sort of app that we're building or
31:40a website. But something else that we
31:42get in Cloud Code is that we get our
31:44scheduled tasks. And that's a really
31:45cool native feature that will run
31:47whenever we have this desktop app open.
31:49And we can have tasks scheduled for once
31:51a week, once a day, whatever we want. We
31:53also have the ability to kind of look at
31:55what we have for customization. So in
31:57this project, we can look at our
31:58connectors and our skills. And here's,
32:00like I said, the crab does some funny
32:02things every once in a while. You can
32:03change your chat model down here. You
32:05can still use different slash commands.
32:06As you can see, we've got access to all
32:08of these different commands. But there
32:09are probably some where you might try to
32:11invoke them. Like let's try if we try to
32:13use agents. And as you can see, I don't
32:15think it did anything. But if I open up
32:16the terminal and I do / aents, we have
32:19this actual command which lets us create
32:21new ones or you know play with other
32:22ones that we've already built in this
32:24project. So who is this for? This is if
32:26you want the power of cloud code but you
32:28like clicking buttons instead of typing
32:29commands. It's a really good place to
32:31start especially if you've already been
32:32using co-work and you want to start
32:33dabbling with cloud code. But just stay
32:35tuned because you might also like option
32:37four which is what I use every single
32:38day. And now moving on to number three
32:40we have the web. Now, this is a research
32:43preview, an experimental type of
32:45product, but it is pretty cool that you
32:46can access this on the web, so you could
32:48use it from your phone. So, there's no
32:49local setup needed, which is really
32:50nice. And it runs entirely on the cloud.
32:52So, you go to claw.ai, you open the code
32:54service, you connect your GitHub repo,
32:56and that's what it uses. So, it
32:58basically clones into your repo in a
33:00cloud environment, and it will do all
33:01the work for you. So, it can still
33:02manage those files and folders. You
33:04review the changes, and then you can
33:05create a pull request, and you can do
33:07all of that from your browser. And a big
33:08deal here is that it keeps running even
33:10if you actually turn off your computer
33:12because it's running on Anthropics cloud
33:14environment. So pros and cons, we've got
33:16almost no local setup, just a browser,
33:18just to add a GitHub account. It works
33:20from any device, whatever you need.
33:22Sessions persist, so Cloud keeps working
33:24while you're away, and it's great for
33:26kicking off long tasks and checking back
33:27later. So Boris Churnney, the guy who
33:29created Claude Code, did a tweet about
33:31how he actually uses it. And what he
33:33said here is that he runs five clouds in
33:35parallel in his terminal and he also
33:36runs five to 10 on claw.ai/code. So in
33:39the web. So really that's just to say
33:41that you don't have to use just one and
33:42stick to just one. But it's important to
33:44understand the differences. So right
33:45here you can see that I'm at
33:46claw.ai/code. I could here connect a
33:48GitHub repository as you can see. I
33:50could choose that over right here as
33:52well. I could be in testing or a
33:53different environment and I could start
33:55all my sessions off right here and then
33:56they would keep running. So you can see
33:58here I just started a new session. I'm
33:59actually not even in any repo right now.
34:01and I said hello. And you can see that
34:03it's going to give you a very similar
34:04feel to the desktop app version with our
34:07sessions on the left and then kind of
34:09our chat interface right here. So, who
34:11is this for? If you want to hand claw a
34:13task and walk away or if you want to
34:14code from your iPad. But of course, you
34:16could also use that new feature which
34:18was the remote control. So, you can kind
34:20of see how these all kind of blend
34:21together. They're similar, but they're
34:22different. All right, let's move on to
34:24number four, which is idees. And this
34:27has also caused a lot of confusion
34:28because there's multiple different IDEs
34:30out there. So this image right here is
34:32how I use it in Visual Studio Code. An
34:34IDE stands for integrated development
34:35environment and it's basically just a
34:37GUI for editing files and folders. So VS
34:40Code is an IDE. Cursor is an IDE.
34:41Anti-gravity is an ID. I don't know why
34:43I have cursor twice right here. But
34:45those are all IDEs, which is why I said
34:46that this has caused some confusion cuz
34:48people are like, why would you use this
34:49in VS Code over cursor or over
34:51anti-gravity? And then anti-gravity has
34:53its own agents. So people get confused
34:54about that too. But they're all just
34:56code editors. And Cloud Code has an
34:57extension that allows us to run Cloud
34:59Code inside of them. And I like it
35:01because I can see the files that I'm
35:03looking at. And Cloud can see what I've
35:04highlighted and I can see my project
35:06structure and my folders and files. I
35:08can drop things in and I can reorganize
35:09it. Now, some of these IDs do have their
35:11own built-in agents, which is why it
35:13might get a little weird because
35:14anti-gravity has some Gemini agents.
35:15Cursor has its own. VS Code even has its
35:17own agents, too. And there's lots of
35:19other extensions and different ways that
35:20you can customize your IDE. All right.
35:22Right. So, starting with the pros, we
35:23have zero context switching because
35:24Claude can see exactly what you see in
35:26the editor. You can also review changes
35:28using your editor's built-in tools,
35:29which is really nice. You have really
35:31quick access to opening up different
35:32sessions and using keyboard shortcuts.
35:34And like I said, I just feel very
35:36productive in there. And it also works
35:37in multiple IDEs as far as the cloud
35:39code extension. So, you can pick
35:41whichever editor that you prefer. Now,
35:42some of the cons are similar to some of
35:44the cons on the other ones. You don't
35:46get some of the advanced features that
35:47are CLI only, but I'm going to show you
35:49guys why that's not a big deal at all.
35:51Your experience is also tied to the
35:52performance of the editor. So if VS Code
35:54is slow or it's crashing, then that's
35:55going to make Cloud Code feel like it's
35:57crashing. But really, it's just your
35:58IDE. And each IDE has its own agent,
36:00which can be a little bit confusing. But
36:02if you're just using the Cloud Code
36:03extension, then it's not too bad at all.
36:04So when you open up VS Code, which is
36:06free to download, or anti-gravity or
36:08cursor, this is what it looks like. And
36:10you can basically open files, you can go
36:12into folders, you can clone a GitHub
36:14repo. But where the magic really comes
36:16in is when you go over to the lefth hand
36:17side and you open up your extensions.
36:20And this is where you would search for
36:22claude code for VS Code and you would
36:24install this. And this basically pops up
36:26a little button in the top right which
36:27says cloud code. And when I click on
36:29that now I have the ability to actually
36:31talk to a cloud code agent. And then it
36:33also gets pretty powerful because
36:34there's tons of other extensions you can
36:35use. So then when you actually open up a
36:37project it's really nice because you can
36:39see all of the folders and all of the
36:40files over here. And not only can you
36:42see them here, but you can open them up.
36:44So, if I click on my cloud MD, I could
36:46go ahead and read this right here as
36:47markdown. I can also see that these are
36:49new lines because they're green. And I
36:50can see over here that all of the things
36:52that are green or yellow haven't been
36:53yet pushed to my GitHub because they're
36:55either new or they have been edited. So,
36:57that visually helps me a ton. Also, when
36:59Cloud Code is making a plan, so in this
37:01session, I had it build a random plan.
37:02It pops that out right here as text. And
37:04what I can do is I can leave comments on
37:06specific elements if I want to correct
37:08it or have it, you know, fix some
37:10things. And we all know the importance
37:12of planning. So, this really helps me be
37:13able to make the plans better so that
37:15the automations or the skills end up
37:16being better in the end. Now, we also
37:19have some status indicators. So, at the
37:20top of this window, if we see that
37:22there's a blue dot, it means that
37:23Claude's waiting for us. And over here,
37:24if there's an orange dot, it means that
37:26Claude has finished up. If I want to, I
37:28can also have multiple different
37:29sessions running in different little
37:31areas. So, I could have, you know, split
37:32pane view where I have four windows and
37:34I can talk to all four agents. And then
37:36finally, we talked about how sometimes
37:38we need some features that don't allow
37:40us to work in VS Code. So, for example,
37:42if I try to run the / agents command
37:44right here, it says, hey, you need to do
37:45this in the terminal. So, that's fine. I
37:47go to the terminal, but then I can just
37:48bring it back in the session. So, now I
37:50have my cloud code terminal running
37:52right here, and then I also still have
37:54the ability to look at all of my
37:55different files and folders over here.
37:57So, it's really the best of both worlds.
37:58So, who is this for? If you want to
38:00spend your day in a code editor and you
38:01want cloud code right there next to your
38:03code, this is the way to go. It's really
38:05not that overwhelming. So, if you're
38:06currently deciding if you should switch
38:07from the desktop app to VS Code, I would
38:10probably go ahead and do that. And last
38:12but not least, we have number five,
38:13which is VPS. So, this means that you
38:16can actually run Cloud Code on some sort
38:18of virtual private server rather than on
38:20your local machine. So, similar to the
38:22cloud on the web, except for rather than
38:23anthropics cloud, you pay for some sort
38:26of cloud yourself. And the way that I
38:28like to do that is with Hostinger. And
38:29because it's running on a remote server,
38:30it can always stay on. So, why would you
38:32do this? The code and the services on
38:34there have to live somewhere. So if
38:36you're running Docker containers,
38:37databases, and automations, Cloud can
38:39sit right next to them at all times with
38:41access. And we have persistence, meaning
38:43you can start a task, you can close your
38:45laptop, and cloud keeps working on the
38:47server. So I actually have a VPS session
38:49of Cloud Code set up. And I've connected
38:51it to a Telegram bridge. So I can now
38:53talk to my Cloud Code wherever I am from
38:55my phone right here. And it can respond
38:57to me. It can look through my files. It
38:58can create files. It's basically cloud
39:00code in your pocket. And what else is
39:02cool is if you have it running on a
39:03virtual private server, you can SSH into
39:05that from your own terminal locally or
39:07from VS Code locally or even the desktop
39:09app. So you can pretty much get all the
39:11power. So the pros here it is it's next
39:13to your real infrastructure direct
39:14access to all of that. It's always on
39:16and you can work across any device. The
39:18cons are that you kind of need some
39:20basic server knowledge. Although what I
39:21did is I set up a cloud code project to
39:24help me spin up and maintain my VPS.
39:27It's also a little bit more setup
39:29friction than the others. And Claude
39:30does have real power if it's on a server
39:32and can see everything on the server. So
39:34you obviously want to be careful with
39:35permissions and things like that. So if
39:36you go to the link in the description,
39:38you can see that Hostinger has a plan to
39:40specifically help you launch Cloud Code
39:42VPS. You can get started for as little
39:43as six bucks a month. So that's pretty
39:45cool. When you come down and choose an
39:46option, you basically just get to
39:47compare these different options of RAM
39:49or CPU cores or bandwidth. I would
39:52probably just start with this one. And
39:53you can always scale up or scale down as
39:55needed. And if you have no idea what
39:56you're doing here, then basically just
39:57use the Kodi AI assistant and tell it
40:00about what you're trying to run and it
40:01will probably help you be able to make
40:03the right choice. And then once you get
40:04in there, you'll have a dashboard, which
40:05I'll show you guys. And you can also use
40:06that if you need to help scale down or
40:08up. And once you go through the setup,
40:09you will see that you can choose a
40:1124-month, 12-month, or 1-month plan. If
40:13you're on one of the annual plans, you
40:15can use code Nate Herk, and you can get
40:1610% off, which ends up saving you a ton
40:19of money. You can also get daily auto
40:20backups. You choose your location. And
40:22right here, you can basically launch
40:23this with an app. So right here they
40:25have cloud code which makes it super
40:27easy to set up and get going. And then
40:29once you're in your dashboard, you can
40:30see the CPU usage, the memory, the disk,
40:32all of this kind of stuff. And once you
40:34notice that this is getting a bit too
40:35high, then you can just scale up your
40:37plan with a click of a button. But
40:38basically all you need to do from here
40:40is SSH into your actual server and then
40:43install Cloud the same way you would do
40:44on a terminal or anywhere else. And then
40:46you have Cloud Code running there. And
40:48you can set up sessions that last
40:49forever. And you don't even have to
40:50actually SSH in. You could literally
40:52just hit this button right here, which
40:53brings up the actual server terminal. As
40:56you can see, we have this really ugly
40:57terminal looking thing right here. And
40:58then you can go to this doc, which
41:00basically shows you exactly what you
41:01need to do in order to actually install
41:03and log into cloud code. And then it
41:05looks the exact same as you would have
41:06used it anywhere else. So, who is this
41:08for? If you're running servers,
41:10deploying apps, or you want an always on
41:11AI assistant working on your code 24/7,
41:13then this is the setup. All right. Well,
41:15that is going to do it for today. Now
41:17that you guys have watched this
41:18breakdown, hopefully you understand
41:19exactly which way you want to use or
41:21which ways you want to use Cloud Code
41:23and now you're ready to actually start
41:24diving in.
41:28Okay, so what does Cloud Code actually
41:30do? Well, the short answer is that it
41:33can do basically anything, which is why
41:35it has been blowing up and which is why
41:37I'm having so much fun using it. But you
41:39can build apps, you can build websites,
41:41and you can build automations. You can
41:42even build like games. You can enhance
41:45your personal workflows. You can debug,
41:47refactor, and test and write code. You
41:50can create API integrations. You can
41:51generate documentation, handle git
41:53workflows, data analysis, and content
41:54pipelines. You can pretty much do
41:56anything even now that we have like
41:59browser use. So you can literally give
42:01Claude Code the access to open up a
42:04browser tab and look at it, screenshot
42:06it, screen record it, click on things,
42:08fill out forms. It's just really cool.
42:11And the one mindset shift that I really
42:12want you guys to adopt is genuine
42:15curiosity.
42:17Because when you first get into Cloud
42:19Code, obviously you'll be following this
42:20program, so I'm going to have things
42:21pretty structured, but it might just
42:23feel like
42:25this is a very like loose program. not
42:27not my course but like cloud code and it
42:30was built in a way that it's insanely
42:31customizable insanely hackable so that
42:34you have unlimited possibilities so when
42:36I say be genuine cur genuinely curious I
42:39just mean if you are confused about
42:41something ask it if you are wondering if
42:44something's possible ask it if you want
42:46to understand how you build something
42:47just ask it and you can ask follow-ups
42:49and follow-ups and follow-ups you can
42:50have it design exercises for you to
42:52understand concepts it's going to be
42:54your best friend because anything that
42:56is technical, it will know better than
42:58you. Whether that's an API or whether
43:00that's a tool or whether that's um some
43:03command line, you know, interface
43:05function that it's running, just ask it.
43:08So, what I really like to do is when I'm
43:10in my cloud code and I'm talking to it,
43:13you can see what it's thinking, right?
43:15So, I literally just like to read what
43:16it's thinking and I like to read the
43:18lines of command that it runs. And the
43:20more and more you read what it's doing,
43:22the more you understand it. So in this
43:25section, we're just going to get some
43:26initial kind of like setup and concepts
43:28talked about. So cost, prompting,
43:30permission modes, cloud models, tokens,
43:32context, windows, built-in tools, and
43:34cloud.md. So let's dive in. So the first
43:37thing is cost, right? You have to be on
43:38a paid plan. Um, you can start off with
43:41pro. It's 17 bucks a month build
43:42annually, but I would definitely
43:44recommend you just go for the max. If if
43:47you want to start with pro, you'll
43:48probably just hit your limits pretty
43:49quick and then you'll get frustrated and
43:50you'll either upgrade or you just have
43:52to wait an hour or a couple hours.
43:55And real quick, I wanted to talk about
43:56the ROI because yes, 100 bucks a month
43:59is an expensive subscription in the
44:01grand scheme of things, right? Like you
44:03have n bucks a month subscriptions, 29.
44:04100 seems like a lot and even if you
44:06push it to 200, but it's all relative.
44:09Think about the fact that a software
44:12engineer costs around 11,000 bucks a
44:14month or an annual salary of maybe in
44:16this range
44:18and you can get the output of a software
44:20engineer, a full-time software engineer
44:22for 100 bucks a month. That is crazy.
44:24And that's the way I want you to think
44:25about this. Not a subscription. Think
44:27about it as a software engineer on your
44:29laptop.
44:31It's a complete mindset shift, right?
44:33Because they can do everything that a
44:35developer does. Basically, Enthropic,
44:37the company that built cloud code, uses
44:39cloud code for everything. They built a
44:42tool called cloud co-work and they built
44:44that in like 10 days with a team of two
44:46or three developers only using cloud
44:48code. That would have taken months and
44:49that would have taken a massive team of
44:51developers. They're even right now using
44:54AI to write and review all their code
44:56because humans were missing things that
44:58the AI is catching. So, it's just like
45:00it's crazy crazy leverage. It's not
45:03marginal ROI. It is truly
45:05transformational ROI. [snorts]
45:08So prompting prompting matters, right?
45:10Because cloud code at under the hood is
45:12an AI model that can talk to your tools,
45:15that can do things, that has access to
45:16your files. And so prompting is so
45:18important because the quality of the
45:20clearness of your prompt and the context
45:23that you feed it is directly tied to the
45:26quality of output it gives you. So a
45:28vague prompt produces vague results.
45:30Garbage in, garbage out. Think of Claude
45:33as a brilliant contractor that you just
45:35hired. They have tons of skills, but
45:36they have never seen your project
45:38before. So, the more precisely
45:40precisely the more precisely that you
45:42explain what you want, including your
45:44context, your constraints, and your
45:45expected outcomes, the far better the
45:47result will be. And as you guys, you
45:49know, it's all about the reps. As you
45:50guys start to talk to Cloud Code more
45:51and more, you'll realize just how good
45:54it is. And one key thing there is using
45:58voice to text. So, just to show you guys
46:02what that means real quick, when I speak
46:04in natural language with my voice, I can
46:07speak way faster than I can type. And I
46:10speak way more, I guess, naturally.
46:13Like, when I speak, I feel like I
46:15actually unlock more thoughts because
46:16I'm not having to be restricted by my
46:18fingers. I'm restricted by just oxygen,
46:21I suppose. So, I use this tool, which um
46:24there will be a link for in the
46:27description. Just go ahead and check out
46:28the voice to text tool down there or I
46:30can go like this. You see this little
46:31thing at the bottom of my screen. I hold
46:33this and now I can basically just talk
46:35and I can even hit the space bar. So now
46:37my hands can be off and I can just be
46:39talking. So I can be brain dumping and I
46:41can be talking about whatever and
46:42telling Cloud Go, "Hey, I want you to
46:43build this." And then boom, all of a
46:45sudden my words have just appeared. So
46:48that has been a major major productivity
46:50boost, major unlock for me. And if you
46:52guys don't have that tool, I would a
46:54thousand% recommend that you get that
46:56tool and you use that tool. Okay, let's
46:58keep on moving here. [snorts]
47:01So, bad versus good prompt example. Bad
47:04would be something like build me a
47:05website for my dog walking business.
47:06It's going to look generic. Good would
47:08be create me a simple landing page for a
47:09dog walking business called Happy Paws.
47:11It should have a hero section with a
47:12headline, a list of three services with
47:13prices, and a contact form at the
47:15bottom. Use a clean, modern style with a
47:17blue and white color scheme. Now, the
47:19cool thing about this is this still
47:21honestly isn't a great prompt, but what
47:23you can do is you can feed that prompt
47:25right here into Cloud Code. You can
47:27utilize plan mode, which you guys will
47:28understand in just a bit, and you can
47:30let Claude code ask you the difficult
47:32questions that it needs. So, pretend
47:33right now you you want to build a
47:35website. You are talking to the world's
47:37best website designer.
47:39You give that person this prompt. That
47:42person is going to come back to you with
47:43tons of questions. Okay. Um, what's the
47:45this? What's this? What's this? Tell me
47:46more. And you can let them adopt that
47:48role of asking you questions. So it
47:50makes your job so easy. You can say
47:52things like you should be 95% sure
47:54before you move on or ask me any
47:56questions to make sure you understand
47:57me. And that will literally make sure
48:00that cloud code doesn't start building
48:01anything until it completely understands
48:02your request. So permission modes, we've
48:06got three kind of main ones that I'm
48:08going to cover. We've got plan mode,
48:10which means Cloud can help you plan. It
48:13thinks, it can read things. It can do
48:15web search, but it won't actually build.
48:18We have accept edits, which means that
48:19cloud can read things, write files, edit
48:21files without asking. But it still has
48:24to ask permission for bash commands,
48:25which means kind of like actually taking
48:27action. And we have bypass permissions
48:29mode, which basically means you talk to
48:31Claude, it can go do whatever it wants,
48:34full autonomy.
48:36Now, this is beneficial sometimes when
48:37you have come to a plan that you like
48:39and then you shoot it off in permission
48:41mode, in bypass permission mode, because
48:42you don't want it to babysit it. So
48:44there's some value there.
48:47Now, here's what they look like in cloud
48:49code. You can see at the bottom I'm
48:50switching between plan, bypass, ask
48:54before edits, and you literally just
48:55toggle that with a button. So this is
48:58the GUI, and we can just toggle it. You
49:00can notice that this kind of looks like
49:01the way that, you know, maybe Claude in
49:03the browser looks or chatbt in the
49:05browser looks. So that's your different
49:06modes. Super easy. Now we have models.
49:10So the number that's attached to the
49:13back, don't worry about that. Right?
49:14Like right now we're on Opus 4.6, SA
49:164x6. But these are the model families in
49:19Claude. So we have Claude Haiku, Claude
49:21Sonnet, and Claude Opus.
49:23They have different strengths. They have
49:25different weaknesses, and they have
49:26different costs. And it's important to
49:27understand. So Haiku is the fastest. It
49:30is the simplest, and it's the most
49:31lightweight, lightweight, and it's also
49:34the cheapest. So it makes sense, right?
49:35Sonnet is kind of like the balanced
49:37version. So, daily coding, it's very
49:38balanced, it's fast, and it's mid-range.
49:40And Opus is a heavy reasoning model. So,
49:43it's probably going to take the longest,
49:44and it's probably the smartest, but it's
49:46also the most expensive. Now, I will be
49:48very upfront with you guys. I use Opus
49:51for everything. The only time I will
49:53switch to Haiku is maybe if I have
49:54certain sub agents or I have tons and
49:56tons of tokens to process. But when
49:58you're just starting out, just keep it
50:00on default or just keep it on Opus just
50:02to get a feel for it, right? So here are
50:05the models in the same interface. You
50:07can type /model and it lets you choose
50:09between default set or opus, right? And
50:13it's just as simple as literally you saw
50:15at the beginning of this clip. All I did
50:16was I typed /mod and it popped up with
50:18this option to change my model. So very
50:21easy.
50:22Now what do the numbers mean? I'm not
50:24going to spend too much time here, but
50:25it's basically just like the version. So
50:27cloud [clears throat] 3 was the
50:28original. Cloud 3.5 had improvements.
50:30Cloud 4 was introduced. Cloud 4.5 had,
50:32you know, more improvements. Cloud 4.6
50:34Six has more improvements. So the models
50:36just increase the number, right? So use
50:40the model slash command. A good
50:41strategy. Use Sonic for 80% of work.
50:43Switch to Opus for complex architecture
50:45decisions or tricky bugs. Then switch
50:46back. Kind of like best practice. But
50:48like I just told you guys, I just like
50:50to use Opus, right? So what is a token?
50:54A token is a unit. And that is basically
50:56how you know we read in words and
50:57letters. AI reads in tokens. So it's not
51:00exactly one word. You know, a token is
51:02not exactly one word. It's roughly three
51:04to four characters. And I say roughly
51:05because sometimes you'll notice like
51:07this punctuation mark like this comma is
51:09one. This period is one. Um here we have
51:12in and that's two. And that also
51:13accounts for the spaces as you notice.
51:16This is also one. So it's like it's not
51:18also it's not completely consistent. But
51:20roughly
51:22one token is about 75% of a word. So,
51:28this is important to understand because
51:30tokens cost money.
51:32And we have something with all of our AI
51:34models called a context window. The
51:37context window is Claude's working
51:38memory. So, just imagine that Claude has
51:40a notepad. And when you're talking to
51:42it, it writes down everything that you
51:44said and everything that it said back to
51:45you in this notepad. And when the
51:47notepad fills up, it's bad, right? Like,
51:49it's going to start to um lose track of
51:53what's going on. And it also might just
51:54have to reset.
51:56So things that go into this context
51:59window are the system prompt, your
52:01tools, your cloud.MD files, MCP servers,
52:04all of your conversation history, all of
52:06the files, everything that Claude said,
52:08everything that you said. And the con
52:10standard context window is about 200,000
52:12tokens at least at the time of recording
52:14this. The models are getting better
52:16every day. So it's important to know
52:18that and it's important to contextualize
52:20that one huge huge principle of cloud
52:22code once we get more advanced is
52:24minimizing token usage and keeping you
52:27know they call it context management. So
52:28that will be covered later in the course
52:29but for now don't worry too much about
52:32that. Just be aware of that concept.
52:34Okay.
52:36So here's why it matters. It fills up
52:40and as it fills up you consume more
52:42tokens which means you're going to hit
52:43your limits. You also have this dilemma
52:45or this problem of being lost in the
52:47middle. Meaning information at the start
52:49of the conversation and at the end of
52:51the conversation get prioritized because
52:53stuff in the middle can tend to get lost
52:54because there's just so much data,
52:56right? And then of course you have your
52:58cost. More tokens is more computational
53:01cost and um hitting your limit faster
53:04because your subscription is based on
53:06tokens.
53:08So here is what we also call context
53:10rot. So as you have more tokens build up
53:13in a session, we significantly see or we
53:16see a significant drop in the quality
53:18and accuracy of the LLM. So if your LLM
53:21all of a sudden, if Cloud Code all of a
53:22sudden starts making things up, then
53:24maybe compact your token window or maybe
53:26open up a new session because as you can
53:28see as there's more tokens the quality
53:30just kind of takes a steep drop off and
53:32this is seen across all the models. This
53:34is a handdrawn sketch, right? But this
53:35is basically how it works across all the
53:37models. Context rot is a real thing with
53:39AI.
53:41So, here are some helpful commands that
53:43help you with some context stuff. So,
53:45slashcontext, it shows you the current
53:47token usage breakdown. Slash compact
53:50compresses the conversation and saves
53:51key information so you can keep going as
53:53if you still have all that history
53:55there. Slashcle just wipes everything
53:57and starts fresh. And SLR goes back to
54:00an earlier point in the conversation
54:01after a code change. There is something
54:03that Cloud Code does called autocompact,
54:06which means that it'll automatically
54:07compress your conversation once you hit
54:08that limit. So, when you're first
54:10getting started, don't stress too much
54:12about context Windows. It'll
54:15autocompact, but just something, like I
54:17said, I wanted you to keep in mind. So,
54:19here's an example. I'm typing in /clear,
54:21and that clears the conversation as you
54:22saw. And now I'm typing in /context, and
54:25it's pulling up all of the different
54:26things in this project that are eating
54:28up token usage. So, here you can see,
54:31oops, didn't mean to go forward. I meant
54:32to pause. Here you can see that in this
54:35session, I'm using model cloud is 4.6
54:38and my tokens are at 225,000
54:42or sorry 22,000 out of 200,000. So I've
54:46already eaten up 11% of my context
54:48window. If I was to then scroll down,
54:50you can see that there are other things
54:52in here that are eating tokens. So I
54:53have um you know all of these system
54:55prompts, these system tools, these MCP
54:57tools. I have MCP servers that are
54:59taking up some context and I've also got
55:00like skills and agent files, right? So
55:03the more things you have in your
55:04project, the more tokens that are going
55:06to be consumed.
55:09Okay, so now I wanted to talk about
55:12built-in tools in Cloud Code. Now, when
55:16you're using Cloud Code, it'll basically
55:18tell you which tool it's using, and you
55:20don't have to memorize these at all.
55:21They're very intuitive, right? Like
55:22read, write, edit, bash, glob, they're
55:24pretty much intuitive. Bash means you're
55:26running a shell command in your
55:28terminal. So, if you open up your
55:29terminal and type something, that's
55:30that's a bash. A glob is looking for
55:32things. A GP is also searching for
55:34things. An ls is listing files that
55:36exist. Web fetch is getting content from
55:39a URL. Web search is searching the web
55:41for a URL. And this is just a quick
55:43breakdown so you're familiar with tools.
55:46Um, now like I said, I off the top of my
55:49head wouldn't be able to tell you what
55:50they all do. Well, maybe I would now
55:53because I've just read them so many
55:54times, but you don't need to memorize
55:56these. The reason I bring this up is
55:57because when you hop into Cloud Code and
55:58you see these words, you're probably
56:00going to think to yourself, "What in the
56:01world is a glob? What is a GP?" So, I
56:04just wanted to get you familiar with the
56:05fact that this just means it's looking
56:06through files or it's executing
56:08commands.
56:09So, here's an example, right? I said,
56:11can you create me an image using my AIS
56:13logo? Um, it's computing. So, it has
56:15these little things that will pop up. It
56:17executed a skill called generate image.
56:19We can see it's thinking. We can see
56:20it's searching. It's using the glob. We
56:22can see that it found the logo. We can
56:25see that once again it's computing or
56:27it's finagling or whatever words. Here's
56:29a bash command. And this bash command
56:31was running a script to go generate
56:33that, you know, AI generated image. So
56:37that's just a quick preview on what it
56:39actually looks like and why I wanted to
56:40bring up those different built-in tools.
56:44Okay, so claw.md.
56:47This is something you're going to hear
56:48so many times. If you've ever built an
56:51AI agent before or a chatbot or
56:52anything, you've given it a system
56:54prompt. Cloud.md is just a system
56:57prompt. That's all. The reason it's
56:58called claude is because we're using
57:00claude code and MD means markdown. So
57:02I'll show you guys an example of what
57:03markdown looks like in a sec. It's not
57:05scary. It's not code. So this system
57:08prompt gives claude code instructions
57:10about the project. Every single time
57:12before like when you send cloud code a
57:14message, every single time before it
57:15reads your message, it reads its claw.md
57:18file first. So it's a system prompt.
57:22Now, there's kind of three main layers
57:24that you want to put in your cloud.MMD
57:26because if you think about tokens and
57:28you think about what I just said, if
57:29your cloudmd file is huge, that means
57:32every single time you ask it a question,
57:34it reads that whole thing and consumes
57:35tokens really fast. So, keeping your
57:37cloud.MD file lean is important. So, we
57:40want the what, which is your text stack,
57:42your product structure, any key packages
57:44or skills. We have the why, which is the
57:47purpose of each component. And we have
57:48the how, which is how you want Claude to
57:51work. Now, the good thing is Claude is
57:54really good at building CloudMD files,
57:57right? So, here is an example of the
57:59beginning of my cloud.MMD file for my
58:01executive assistant. You are Nate Herk's
58:03executive assistant. Your job is to help
58:05him spend less time on operations,
58:06people management, and admin so he can
58:08focus on learning AI tools and making
58:09YouTube videos. This is his number one
58:11priority. Everything else supports it,
58:12right? Very clear goal. Now I also give
58:15it other information so it can read
58:17context about me, about work, about my
58:19team, about my priorities and then I
58:21list my tools and then I list you know
58:22other you know project management
58:24frameworks or anything that it needs to
58:25know anything that's really important.
58:27[snorts]
58:28Now we have a cool command called /init
58:30which basically initializes your project
58:32by scanning your codebase and creating a
58:33cloud.MD file. So if you opened up a
58:36project and you don't have a cloudmd
58:38just run slashinit and it will make you
58:40one. So that's a pretty cool command.
58:43Okay, that is going to do it for now.
58:46Hopefully I didn't bore you guys too
58:47much. You're done with slides. Let's go
58:49actually get our hands a little bit
58:51dirty with cloud code and start building
58:53your first agentic workflows. So, super
58:55excited. I'll see you guys over there.
58:57Aenic workflows are changing how we
58:58build AI automations. But not only that,
59:00they're also changing the entire
59:02industry with more businesses [music]
59:03investing in Aentic AI to improve their
59:05workflows. So, if you're looking to
59:06start building AI automations to make
59:08your life easier or you want to make
59:09money building these for businesses,
59:11then this is [music] the best place to
59:12start. I've been building AI automations
59:13for a little over a year now and I've
59:15already helped thousands of people build
59:16their first AI automation. So I do know
59:18how intimidating everything can look at
59:19first. So that's why in this video I'll
59:21be telling you everything you need to
59:22know and then I'm going to show you how
59:23to build your first agentic workflow
59:24from absolutely zero. My job is to make
59:26this as easy to understand as possible
59:27for you. So let's get into it. All
59:29right. So before we actually build
59:30anything, let's make sure that we're on
59:31the same page about what aentic
59:33workflows actually are. If you've been
59:34building traditional automations, you
59:36know the drill. You use a tool like make
59:37or n. You drag a node onto the canvas.
59:39You configure it. You connect it to the
59:41next one. You make sure the right
59:42variables are passing through. You test
59:43it. You add another node. And you keep
59:45going. And when you hit an error, which
59:46you will, you read the error message.
59:48You figure out what went wrong. You fix
59:49it. You test again. And you repeat until
59:51it works. You're basically building the
59:52whole thing manually. And if it breaks
59:54again later, you're the one who has to
59:55go back in there and fix it yourself.
59:57Now, this was a huge leap in the AI
59:58automation space because it
59:59significantly lowered the barrier to
1:00:01entry and it allowed anyone from any
1:00:02background to learn these tools and
1:00:04build some really powerful automations
1:00:05in a matter of days. But agentic
1:00:07workflows completely flip that whole
1:00:08process because instead of telling the
1:00:10system how to do something step by step,
1:00:12you're just telling it what you want and
1:00:13then the agent figures out the rest. So
1:00:15think about it like hiring a really
1:00:16talented developer. You don't sit there
1:00:18explaining the code or walking them
1:00:19through the logic line by line. You walk
1:00:21in and you explain the problem. You
1:00:22describe the outcome you want and then
1:00:23you ask, okay, what else do you need
1:00:25from me? So that's what makes it aic.
1:00:26The system reasons, it adapts, it asks
1:00:28clarifying questions when it needs to.
1:00:30It makes decisions. It fixes itself when
1:00:32something breaks. And it does research
1:00:34all to make your job as easy as
1:00:35possible. Now I do believe that
1:00:36traditional automation with a tool like
1:00:38Naden isn't going anywhere. It's still
1:00:40perfect for repetitive predictable
1:00:41tasks. And there are two terms that we
1:00:42use in the AI automation space.
1:00:44Deterministic and non-deterministic.
1:00:46Deterministic means predictable. And in
1:00:48automation, predictable is beautiful.
1:00:49Boring is beautiful because you know
1:00:50exactly what's going to happen every
1:00:52single time the automation runs.
1:00:53Non-deterministic means that given an
1:00:55input and you don't know exactly what
1:00:56the output will be. There's variability,
1:00:58there's judgment, there's AI, and AI is
1:01:00non-deterministic. So our job as AI
1:01:02automation builders is to make a
1:01:03non-deterministic process as
1:01:05deterministic as possible because
1:01:07typically business processes are pretty
1:01:09deterministic or at least as
1:01:10deterministic as they can be. So that's
1:01:12exactly where agentic workflows shine.
1:01:14They unlock tasks that are too variable
1:01:16for traditional automation stuff that
1:01:17needs judgment calls at every step just
1:01:19to be a little bit more dynamic. Maybe
1:01:20research, maybe content creation,
1:01:22customer support, lead genen. These are
1:01:24messy processes that can involve a lot
1:01:25of moving pieces. So with aentic
1:01:27workflows, we can handle that
1:01:28variability and the system actually gets
1:01:29better over time instead of just setting
1:01:30it and forgetting it or having to go
1:01:32manually in and make improvements by
1:01:34yourself. There's a reason why so many
1:01:35builders right now are shifting to tools
1:01:37that are a little bit more genetic like
1:01:38cloud code or anti-gravity because they
1:01:40fix a lot of the common struggles with
1:01:41traditional automation. There's no more
1:01:43finding and fixing errors manually. No
1:01:45more setting up API calls yourself. No
1:01:46more manually connecting to MCP servers.
1:01:48No more getting stuck on the logic. So
1:01:50here's a really simple way to think
1:01:51about that evolution. Let's say you
1:01:52wanted to get to a carnival across town
1:01:53and you know roughly where it is, but
1:01:55you still need directions. Traditional
1:01:56automation is like using a paper map and
1:01:58a compass and you're looking at the
1:01:59street names and you're trying to figure
1:02:00out your own routes. You're choosing the
1:02:02streets to walk down and you can get
1:02:04there and you will get there. It just
1:02:05takes a little bit more effort and if
1:02:06you make the wrong turn, you'd have to
1:02:07figure that out and course correct. But
1:02:09with Aentic Workflows, that's just like
1:02:10pulling out your phone, googling for the
1:02:12carnival, and then it basically gives
1:02:13you this blue line and all you have to
1:02:14do is follow it. And if you go off the
1:02:15path, it will like recalculate and it
1:02:17will make sure that you go back to the
1:02:19actual outcome that you're looking for.
1:02:20So in both scenarios, you can get to the
1:02:22same destination, but it's just a
1:02:23completely different experience getting
1:02:24there. So let's talk about what's
1:02:26actually happening under the hood and
1:02:27what you need to know in order to build
1:02:29an agentic workflow. So we're using a
1:02:30framework called WAT. You could just hop
1:02:32into Cloud Code right now and start
1:02:34talking to the agent and honestly it
1:02:35would do just fine. But without
1:02:36structure, things get messy fast. Think
1:02:38about it like a school locker. If you
1:02:39just threw every piece of paper, every
1:02:41homework assignment, every note from
1:02:42every subject into a locker with no
1:02:44organization, you could get straight
1:02:46A's, but it would be tough because you'd
1:02:47be digging through piles of paper. you'd
1:02:49probably forget things or lose things
1:02:50and that's why you would have finders,
1:02:52shelves, folders, notebooks. Structure
1:02:54makes everything easier. So, it's the
1:02:55exact same thing here. We need to tell
1:02:56cloud code how to stay organized and
1:02:58that's why we do that using our
1:02:59framework called WAT. W stands for
1:03:01workflows, A stands for agent and T
1:03:03stands for tools. Each piece of that
1:03:05framework has its own job. So, let me
1:03:07break that down. All right. So, first we
1:03:08have the workflows which are the
1:03:10instructions. These are instruction
1:03:11files that are written in markdown,
1:03:12which is basically just natural
1:03:14language, but it uses things like pound
1:03:15signs and asterisks so that the agent
1:03:17knows what are the headers, what are the
1:03:18subheaders, what's bold, what's
1:03:20important, stuff like that. So, I'll put
1:03:21a quick example right up here on the
1:03:22screen if you've never heard of
1:03:23Markdown, but just know it's super
1:03:24simple and you could go read that and
1:03:26you would not be confused at all. So,
1:03:27think of a workflow like a job
1:03:28description or an SOP, just a process.
1:03:30It tells the agent what to do. For
1:03:32example, we might have a workflow called
1:03:33competitor analysis. The workflow tells
1:03:35the agent to research businesses, then
1:03:36gather data from five competitor
1:03:38sources, then analyze those findings,
1:03:39and then analyze our business, and then
1:03:41create a PDF report. So, it's just
1:03:42basically a process, a sequence of
1:03:44steps. They're guidelines. The agent
1:03:45then uses these guidelines to figure out
1:03:47how to achieve that end goal. And here's
1:03:49the cool part. As the agent works and it
1:03:50gives you outputs, you can say, I liked
1:03:52this, but I didn't like that, or go
1:03:53ahead and change this, and it'll
1:03:54actually update its workflow file so
1:03:56that next time it calls on the workflow,
1:03:57it will do better. Now, the A stands for
1:03:59agent, which is the coordinator. This is
1:04:00the actual AI. This is claude code
1:04:02itself. This is the brain. It reads your
1:04:04workflows and it reads those
1:04:05instructions and it looks at what tools
1:04:07it has available. And then it makes
1:04:08decisions about which tool to use and
1:04:09when. And if something breaks, it will
1:04:11handle the error. It will research it.
1:04:12It will figure it out and it will adapt
1:04:14for you. So really just think of this as
1:04:15like a project manager. You hand them
1:04:16the instructions and they will delegate
1:04:18tasks to the right people. Except for
1:04:19not people. It's more so they delegate
1:04:21tasks to the right tools and workflows.
1:04:23So you don't have to figure out the
1:04:24sequencing or the logic. Cloud code does
1:04:26it. And then the T are the tools which
1:04:27are kind of the workers or the actions.
1:04:29Tools are Python scripts that actually
1:04:30do the work. And this is where the ugly
1:04:32code lives. But don't worry, you don't
1:04:33have to touch it. Each tool will have
1:04:35one specific job. The workflow is a big
1:04:36process. A tool is just one specific
1:04:38action like scraping a website or
1:04:40generating a PDF. The agent then calls
1:04:42these tools when it needs to based on
1:04:43what it says in the workflow
1:04:45instructions. So for research workflow,
1:04:47your tools might be one to scrape a
1:04:48website, one to analyze findings, and
1:04:50one to generate a PDF. And here's the
1:04:51best part. These tools also get
1:04:53automatically built by cloud code, and
1:04:54if they fail, they get automatically
1:04:56updated and fixed by cloud code. They're
1:04:58super modular, so you can call a tool
1:05:00with a different workflow if you want to
1:05:01later. So, how do these three layers
1:05:02work together? Let me just show you how
1:05:04this connects. Let's say that we give an
1:05:05agent a task like research company X's
1:05:07pricing and then create a PDF report for
1:05:09me. The agent reads the workflow, the
1:05:11instructions, looks at the available
1:05:12tools and decides the sequence. So,
1:05:14first it would call something like a web
1:05:15search tool to find the relevant info.
1:05:17Then it would call something like a
1:05:18scrape website tool to pull content from
1:05:20those URLs. Then it calls the analyze
1:05:22finding tool to synthesize everything.
1:05:23And finally, it could call the generate
1:05:25PDF tool to create that branded report.
1:05:27Now, the whole time it's reasoning and
1:05:29it's making decisions based on what you
1:05:30told it to do in the workflow and you're
1:05:31not mapping it out step by step. The
1:05:33agent handles the logic and updates it.
1:05:35All right, so that's what I wanted you
1:05:36guys to understand about Agentic
1:05:37Workflows. All right, but before we
1:05:38continue, if you want to follow along
1:05:39with the video, you can download this
1:05:41resource I'll be using in my community.
1:05:42Once you're there, you just need to look
1:05:43for the post with this video and you'll
1:05:45find it attached as a markdown file.
1:05:47Now, let's get back to building your
1:05:48first agentic workflow. Okay, so we got
1:05:50all of that stuff out of the way. The
1:05:51first thing I need you to do is go to
1:05:53Google or a browser and type in VS Code
1:05:55or Visual Studio Code and go ahead and
1:05:57download this. This is where we're going
1:05:59to be using Cloud Code. So once you
1:06:00install that, it's going to look like
1:06:01this when you open it up. It's just kind
1:06:03of like a welcome onboarding screen.
1:06:04What I'm going to do is just break
1:06:06everything down as far as what you
1:06:07actually need to click on, what you need
1:06:08to know because there's a lot of buttons
1:06:10in here and it's probably a new
1:06:11interface which makes it overwhelming,
1:06:12but it's going to be simple. You'll see.
1:06:13So before we do anything else, we have
1:06:15to actually install the Claude Code
1:06:17extension. You can see up here I've got
1:06:18this little button where if I click on
1:06:19it, it opens up Claude Code and we get
1:06:21the little crab and we can now talk to
1:06:22Claude Code. So, you're not going to
1:06:24have that by default. The way you get
1:06:25that is you go over to the lefth hand
1:06:26side to the menu bar and you're going to
1:06:27click on extensions. You are then going
1:06:29to search for claude code. So, if you
1:06:30just type in claude, it should pop up
1:06:32over here and you'll click on it and
1:06:33then you just have to install this
1:06:35extension. Now, once you install this,
1:06:36it will prompt you to sign in with your
1:06:38anthropic or your claude subscription.
1:06:40And so, you do have to be on a paid
1:06:42cloud subscription in order to use cloud
1:06:43code. You can see here on the free
1:06:45version you don't have it but pro or max
1:06:47or the higher max you will have cloud
1:06:48code with opus 4.5. So once you're on a
1:06:51pro or max plan then you will come back
1:06:52into VS Code. You'll sign in with that
1:06:54and you should be all set to start using
1:06:56cloud code right here. So we've got that
1:06:57configured. Now what I want you to do is
1:06:58click on the button up in the top right
1:07:00and close out of this window and you
1:07:01should be able to see that you have
1:07:02cloud code right here. So now in order
1:07:04to really use cloud code we have to be
1:07:06in some sort of project. So, if I go
1:07:08over to the lefth hand side and I go all
1:07:10the way up to the explorer, you can see
1:07:11that it says no folder open, which means
1:07:13basically we're not in a project. So,
1:07:15you're going to go ahead and click on
1:07:16open folder and you can see what I did
1:07:17is I created a folder right here called
1:07:19first agentic workflow and it's
1:07:20completely blank. So, open up a blank
1:07:22folder or go create a new one and then
1:07:24select it. So, this is what your screen
1:07:25should now look like. You've got your
1:07:26folder on the lefth hand side with no
1:07:28files in there. You've got these other
1:07:30panels on the right and what we're going
1:07:31to do is close out of the VS Code agent
1:07:33and then we're going to open up Cloud
1:07:34Code and then we're just going to get
1:07:35rid of the welcome VS Code screen. So
1:07:37what we have here is files on the lefth
1:07:38hand side and this is where we're going
1:07:39to see any folders that we create, any
1:07:41of the files that Claude actually makes
1:07:43for us. And then right here is where we
1:07:45can actually talk to Claude code and
1:07:46just think of this as your typical chat
1:07:48GBT interface, your Gemini interface or
1:07:50of course your Claude interface. This is
1:07:51where the agent lives and then this is
1:07:52where the files live. So this is where
1:07:54we'll see the workflows and tools as we
1:07:55mentioned. So you remember earlier I
1:07:57talked about how we had to make sure
1:07:58that our agent understands our structure
1:08:00just like it wouldn't want to throw
1:08:01notes and random stuff in a locker. We
1:08:03have to give it structure. So, what
1:08:04we're going to do is we're going to give
1:08:05it this file that's called a claw.md
1:08:07file. And if you want to get this, I
1:08:08will have it available for download in
1:08:10my free school community. So, this is
1:08:12basically the onboarding document. We're
1:08:14catching the agent up to speed as far as
1:08:15how do we want to work. So, you can see
1:08:16what we're doing is we're explaining
1:08:18you're working inside the WAT
1:08:19frameworks, agents, tools. So, then we
1:08:21go ahead and explain. Layer 1 is the
1:08:23instructions. Layer two is you. Layer
1:08:25three are the tools. I'm not going to
1:08:27read out this whole markdown file line
1:08:28by line. You guys can access it like I
1:08:30said, but I'll hit on a few of the
1:08:31important things. So, we go over how to
1:08:33operate. So we tell it first look in
1:08:34your existing tools. Then you learn and
1:08:36adapt when things fail. So when you hit
1:08:38an error you read it, you fix the script
1:08:40and you reset. So for example, if you
1:08:41get rate limited on an API, you would
1:08:43dig into the docs. You would see if you
1:08:44could discover a batch endpoint. You
1:08:46would refactor the tool to use it,
1:08:47verify it works, and then update the
1:08:48workflow so that error never happens
1:08:50again. And then of course, you want to
1:08:51keep the workflows current. We explain
1:08:53the self-improvement loop. We explain
1:08:54the file structure, which is going to
1:08:56look like this. These are the different
1:08:57folders we're going to have. We'll have
1:08:58one for temporary files. We'll have one
1:08:59for tools. We'll have one for workflows.
1:09:01And then we'll of course have some other
1:09:02different files in here as well. So
1:09:04anyways, that is our cloud.MD file. So
1:09:06what I'm going to do is I'm going to
1:09:07drag it over here into the lefth hand
1:09:09side because this is where we have our
1:09:11project files and folders. So I drop it
1:09:12in there. You can see it opens up over
1:09:14here. We could also read it right there,
1:09:15but I'm just going to go ahead and close
1:09:16out of that. And now we have claw.md set
1:09:19up right here. So what you can do now
1:09:20that we have cla.md is you could do a
1:09:22slash command, which is /init. And that
1:09:24basically just initializes the
1:09:25environment. But we could also just do
1:09:26this in natural language. So, I'm going
1:09:27to go ahead and say, "Hey, Claude, I
1:09:29just dropped in a claw.md file that
1:09:30explains how I want you to work in this
1:09:32project. Go ahead and initialize the
1:09:34project and get everything set up and
1:09:35ask me any questions if you have any."
1:09:36So, when I shoot that off, what you're
1:09:38going to notice is that we can see
1:09:39everything that Claude's doing. We're
1:09:40going to see its thinking. We're going
1:09:41to see its thoughts. We're going to see
1:09:42what it's doing. So, in this case, it
1:09:43literally says, "Okay, I'll read the
1:09:44claw.md file to understand your project
1:09:46requirements, and then I'm going to get
1:09:47everything set up." And then, what's
1:09:48cool is we can actually see what it's
1:09:50doing. So, we can see that it read the
1:09:51file. We can see now that it
1:09:52understands, and now it's going to
1:09:53create a to-do list and start to make us
1:09:55those folders. So, temporary tools,
1:09:57workflows, and you can see on the lefth
1:09:59hand side, it actually just built those.
1:10:00Now, one thing you may have noticed
1:10:01right here is that when we're talking to
1:10:02Claude, yours might look a little bit
1:10:03different because you may be looking for
1:10:05this bypass permissions mode. When we
1:10:06talk to Claude code, we can either use
1:10:08bypass permissions, we can use ask
1:10:09before edits, we can use edit
1:10:10automatically, or we can use just plan
1:10:12mode. So, if you want to be able to get
1:10:13bypass permissions mode, you have to go
1:10:15to your settings, and then you're going
1:10:16to type in cloud code, and then you just
1:10:18have to allow dangerously bypassed
1:10:20permissions. Now, yes, I know it sounds
1:10:21dangerous because the word dangerous is
1:10:23explicitly in there, but it's not too
1:10:24bad. It's really just more so if you
1:10:26give it a huge task and you don't do any
1:10:27planning and you don't know what it
1:10:29could do, it's just going to go execute
1:10:30everything without asking. So typically
1:10:31the flow that we like to follow is use
1:10:33plan mode, have it build out a really
1:10:34nice plan, ask you questions, and then
1:10:36once you're confident in it, say, "Yep,
1:10:37go ahead." And you turn on bypass
1:10:39permissions, which you guys will see me
1:10:40do that exact thing when we start
1:10:41building this workflow. So it did ask
1:10:42some questions. Do you want me to
1:10:43continue with the straightforward
1:10:44initialization? Do you want any Python
1:10:46packages? Do you want to do a git
1:10:47repository? Are there any specific tools
1:10:49or workflows? Right now, we're not going
1:10:50to worry about that. All I wanted to do
1:10:51was just get this folder structure set
1:10:53up. As you can see, we've got workflows,
1:10:55nothing in there. We've got tools with
1:10:56nothing in there. And we've got a
1:10:57temporary folder with nothing in there.
1:10:58So, I'm just going to go ahead and do
1:10:59/clear, which is just going to reset our
1:11:01conversation. All right. So, let's talk
1:11:02about the actual workflow that we want
1:11:03to build today. What I want to build is
1:11:05a competitor research workflow. And I
1:11:07want the deliverable to be branded PDFs,
1:11:09meaning I want to give Cloud Code my
1:11:11logo, my brand guidelines, and
1:11:12information about my business. And then
1:11:14it has to go research competitors. It's
1:11:16going to create, you know, like maybe a
1:11:17SWAT analysis or opportunities for us or
1:11:19tracking what they're doing really well.
1:11:20and then it's going to report all back
1:11:22with a PDF that once again is branded.
1:11:24So that's basically what I'm going to
1:11:25start with because I know what I want,
1:11:26but I don't know maybe the tools we're
1:11:27going to use or the exact structure. So
1:11:29I'm going to switch over here to plan
1:11:30mode and I'm just going to say exactly
1:11:32that. Hey Clude, so I've got an idea for
1:11:34a workflow that I want you to build. I
1:11:36basically at the end of it want a PDF
1:11:38and I want it to be branded. So I want
1:11:39to be able to give you my company logo
1:11:41and my company brand guidelines and the
1:11:43whole PDF output should have my logo on
1:11:46there and have our colors and our
1:11:47typography and stuff like that. But what
1:11:48I want you to do is it's basically a
1:11:50competitor analysis and research
1:11:52workflow. So I want to also give you
1:11:54information about my business that you
1:11:55need to save. And based on that
1:11:56information, you need to go find
1:11:58competitors and you need to find me
1:12:00areas to improve my business. Maybe see
1:12:02what's working well for them and just
1:12:03build me out a good way for me to keep
1:12:05tabs on the market and what's going on
1:12:06with my competitors. Yeah, that's kind
1:12:08of what I'm looking for. So help me
1:12:09build a plan for this workflow. And once
1:12:10again, of course, you can ask me any
1:12:12questions that you have if you're
1:12:13confused. All right, so that was my
1:12:14request. You can see it was all natural
1:12:15language. It's very simple. It's
1:12:17probably the way I would just speak to
1:12:18like a human. And what it's going to do
1:12:19now that it's on plan mode is it's going
1:12:20to think. It's going to look at some
1:12:22stuff. It might even do some initial
1:12:23research in order to help build a plan.
1:12:25And then what it's going to do is it is
1:12:26going to actually ask us questions. Now,
1:12:28I know that this might seem a little
1:12:29intimidating, but really when I was
1:12:30learning Claude Code, the way that I did
1:12:32it was I would ask it a question and
1:12:33then I would just read every single line
1:12:35of what it's doing. If any of these
1:12:36tasks or glob pattern like what is that?
1:12:38If anything confused me, I would just
1:12:40say what is this? What did you do here?
1:12:42Why did you do that? So, it's really
1:12:43about if you're genuinely curious and
1:12:44you just read and pay attention, you
1:12:46will pick this up really, really fast.
1:12:48So, you can see it first of all explored
1:12:49existing workflows and tools and then
1:12:51it's looking at branding and PDF
1:12:52capabilities. This was basically just
1:12:53seeing if it could find any tools in the
1:12:56folder for Python. It's reading other
1:12:57files in our environment to see what's
1:12:59going on. And now you can see that we
1:13:00are in the question phase. So, it's got
1:13:02four questions to start. First one is
1:13:03discovery. How should competitors be
1:13:05identified for analysis? I can either
1:13:07provide a list. It can autodiscocover
1:13:08based on my business info or hybrid. You
1:13:10know what? So, let's just go ahead and
1:13:11try autodiscocover based on my business
1:13:13info. Then it asks us what business
1:13:15information should the workflow collect
1:13:16and save about your company. Company
1:13:17description and value prop, product,
1:13:19service, and pricing, target market, and
1:13:21customer segments, key features. Let's
1:13:22just do all of that. We wanted to get as
1:13:24much information about us as possible.
1:13:25For analysis, it asks what aspects of
1:13:27competitors should be analyzed. We've
1:13:29got products, services, features,
1:13:30pricing and business model, marketing,
1:13:32messaging. I want to analyze all of
1:13:33this. Why would we not? So, I'm going to
1:13:34choose all of it. For branding, it says,
1:13:35"Can I use the existing branding assets
1:13:37I found in your YouTube analysis
1:13:39project?" And this is basically because
1:13:41I've got a big project that this first
1:13:43agentic workflow project sits in and it
1:13:44can search through those as well. But
1:13:46I'm going to assume that this is like
1:13:47you guys setting up a first workflow. So
1:13:48I'm going to say no, which is I will
1:13:50provide different branding. So I'm going
1:13:51to go ahead and submit those answers.
1:13:52It's going to take those adjust the plan
1:13:53a little bit and it may come back with
1:13:55more questions. It may not. So we'll
1:13:56see. And this is pretty cool because it
1:13:57said I have all the information I need.
1:13:59Let me launch a plan agent to actually
1:14:01design the implementation approach. It's
1:14:02asking us how often do we want to run
1:14:04this. I'm just going to go with monthly
1:14:05for now. For the output, it asks if we
1:14:06want anything else besides a PDF. That's
1:14:08pretty cool. But I'm just going to go
1:14:09PDF. And then for the budget, what's
1:14:11your comfort level with API cost for
1:14:13this workflow? I like how it's showing
1:14:14us different approaches here. And I'm
1:14:15just going to continue to go with the
1:14:16recommended approach. And we're going to
1:14:18go ahead and do this middle one. All
1:14:19right. So, at this point, it actually
1:14:20finished the plan. So, if I scroll all
1:14:22the way back up to when it started
1:14:23telling us, you can see this is super
1:14:25comprehensive. So, I'm still scrolling.
1:14:27Okay. Competitor analysis workflow
1:14:28implementation plan. Build a monthly
1:14:30recurring competitor analysis system
1:14:31that automatically discovers
1:14:32competitors, researches their offerings,
1:14:34generates a fully branded PDF report
1:14:35with actionable insights. So user
1:14:37requirements would be discovery,
1:14:38business info, analysis, output,
1:14:40frequency, budget. We've got the
1:14:42architecture. We've got the text stack.
1:14:43So it's going to be using cloud sonnet.
1:14:45It's going to use firecraw and
1:14:46perplexity. It's going to use sonnet.
1:14:48And it's going to use report lab for pdf
1:14:49generation. So it'll probably prompt us
1:14:51to go grab an API key there. And it's
1:14:53also going to generate charts using
1:14:54mattplot liib, which I believe is a
1:14:56python extension or plugin. And it's
1:14:57going to help with charts. It's also
1:14:59going to add some things in our folders
1:15:00over here. So you can see it's going to
1:15:01add a new folder called brand assets,
1:15:03and that's where we will upload our logo
1:15:04and our brand guidelines. And it's
1:15:06basically planning to create a few
1:15:07different things. It's going to create
1:15:08some files in the temporary folder. As
1:15:10you can see, it's going to create a
1:15:11workflow called competitor analysis. And
1:15:13it's going to create these five tools.
1:15:14So, collect business info, discover
1:15:16competitors, research competitors,
1:15:17analyze competitors, and generate
1:15:19competitor PDF. So, exactly like I said,
1:15:21it's going to create this workflow. It's
1:15:22going to create these tools, and then we
1:15:24should be good to run it. Now, one thing
1:15:26I noticed is it has a brand
1:15:27configuration file, and this basically
1:15:29made up our brand information, and it
1:15:30would probably want us to come in here
1:15:32and choose, you know, a name and maybe a
1:15:34logo path. But what we're going to do is
1:15:36we want to actually just drop in those
1:15:37files for it. So that's something that
1:15:38we will have to change, but we'll just
1:15:39keep going for now. We can see that it
1:15:40decided how to handle edge cases like
1:15:42competitor websites block scraping or
1:15:44insufficient competitors found, rate
1:15:46limiting, invalid brand assets, and data
1:15:49completeness issues. It's also giving us
1:15:50a cost breakdown, which is pretty cool.
1:15:52So the first run will be about a dollar
1:15:54and a half. It's also going to be doing
1:15:55subsequent runs on a 30-day cycle, which
1:15:57will be a huge cost savings because it's
1:15:59going to cache some of the data. And
1:16:00then if you're adding some new
1:16:01competitors, it'll be maybe another 50.
1:16:03So anyways, that is the end of the plan.
1:16:05We could basically go ahead now and auto
1:16:06accept or we could keep planning. And I
1:16:08do want to say no keep planning because
1:16:09there's one thing I want to change,
1:16:10which was our brand assets. So what I'm
1:16:12going to do is I'm going to take a logo,
1:16:14drag it into the lefth hand side. Take
1:16:15the brand guidelines and drag that into
1:16:16the lefth hand side. And you can see
1:16:18that we have these things up here that
1:16:19pop in and we can actually see them.
1:16:21That plan looks good. The only change I
1:16:23want to make is about the brand
1:16:24guidelines and the assets. So I just
1:16:26dropped you in two files, AISpng.png
1:16:29and AIS brand guidelines.png.
1:16:32Those are the ones that I want you to
1:16:34use to create the branded PDF. So look
1:16:36at those, extract the information out of
1:16:38them and make sure that the logo and the
1:16:39colors and everything appear on the
1:16:41final output PDF. And if you need to,
1:16:43you can throw those into a folder in
1:16:44this project to keep things organized.
1:16:46So this is awesome. It said, "I found
1:16:47your logos and I'm creating a brand
1:16:49asset guide." So we've got the logo and
1:16:51then it also extracted our colors and
1:16:54typography and now it's going to update
1:16:55the plan to use those assets. And now
1:16:57since that has been changed, we're going
1:16:58to go ahead and auto accept the plan.
1:17:00Hopefully it will get working for us. Of
1:17:02course, it's going to create a to-do
1:17:03list and then it's just going to start
1:17:04building all these different scripts,
1:17:05whether that's a workflow or a tool. And
1:17:07then it's going to test the workflow.
1:17:08And like I said, it'll probably have to
1:17:10come back and ask us for an API key or
1:17:12something like that. All right, so the
1:17:13workflow is ready. We can see that we
1:17:14have our branded assets set up. We've
1:17:16got the workflow completed. We've got
1:17:17the Python tools completed. We've got
1:17:18some setup files. So, we also have a
1:17:20readme that we could open up, which
1:17:21should basically just tell us pretty
1:17:23much how this actual workflow works. So,
1:17:24that's pretty cool. Now, we do have to
1:17:26go get two API keys to start. We need an
1:17:28entropic key and we need a firecrawl
1:17:30key. So it does actually tell us that
1:17:31here again we have to install
1:17:32dependencies. We have to set up those
1:17:34API keys and then we have to run the
1:17:36workflow. So first of all to install the
1:17:37dependencies I'm just going to say can
1:17:39you do this and then paste in exactly
1:17:41what it gave me. And so it's interesting
1:17:42it asks us to do that when it could have
1:17:44just done it itself. As you can see it's
1:17:45able to just run that command for us. So
1:17:47once it did that now it says okay you
1:17:49have to go create two API keys. So we
1:17:51need to create av file which is just
1:17:53going to be a copy fromv.ample.
1:17:55So I'm actually just going to run this
1:17:56command and it's going to copy that file
1:17:58for us. And now we have an actualv file
1:18:00right here. And so you can see it says,
1:18:01"Okay, cool. Now we need your enthropic
1:18:03key, your firecall key." And if I open
1:18:04up the it gives us those placeholders.
1:18:06So basically all we have to do is go
1:18:08grab those keys and then put them in
1:18:09here instead of these placeholders that
1:18:11you can see right there. So first let's
1:18:12go to Enthropic. I'm going to go to my
1:18:14cloud developer platform. I'm going to
1:18:16create a new key. This one's going to be
1:18:17called competitor analysis demo. We're
1:18:20going to have this key right here. Copy
1:18:22that and go into VS Code. Paste it in
1:18:24there. And then the next one we need to
1:18:25get is the firewall key. So, you can use
1:18:27the link in the description to go to
1:18:28Firecrawl. You can actually get 10% off
1:18:30and 1,000 free credits if you use code
1:18:32Nate and use the link in the
1:18:33description. But I'm going to go to my
1:18:34dashboard here. And then all I have to
1:18:36do is grab my API key from right here,
1:18:38paste it into this section. And then
1:18:40what you have to do is make sure you
1:18:41save this file. So you could do crl S or
1:18:43you could just go file, save. But now
1:18:45that that file has been saved, we
1:18:46actually should be good to go ahead and
1:18:48run the workflow to see if it works. So
1:18:50before I do that, you can see that we
1:18:51have this little thing down here, which
1:18:52is context. So 23% of your context
1:18:54remaining until autoco compact. So,
1:18:56usually when this goes over 60%, I
1:18:58usually just like to clear because
1:18:59there's this thing called context rot,
1:19:01which basically means the more and more
1:19:02you use one conversation, the worse the
1:19:04model kind of gets. So, we're going to
1:19:05clear the conversation. We're going to
1:19:06go ahead and ask it to run competitor
1:19:08analysis and then we're going to go
1:19:09ahead and see what happens. Now, the one
1:19:11thing I did notice is that we still
1:19:12haven't given it a ton of information
1:19:14about our business. So, I am a little
1:19:16confused why it hasn't asked about that,
1:19:17but we will see what happens. I'm going
1:19:18to keep it on bypass permissions mode to
1:19:20just see what it does. And I'm going to
1:19:21ask it to generate a competitive
1:19:22analysis. and I'm going to give it a
1:19:24really small amount of information about
1:19:26our business. Hey Claude, I need you to
1:19:27help me run a competitor analysis. My
1:19:29business is called Get Leads with AI and
1:19:32we basically help you scrape leads,
1:19:34build lead lists, and do personalized
1:19:36outreach at scale using AI. And we're
1:19:38starting to see a lot of competitors pop
1:19:40up. So, I want to understand our
1:19:41opportunities and what we need to be
1:19:43doing better. All right. So, what
1:19:44happened there is I just shot off a
1:19:45prompt and I didn't explicitly say like,
1:19:48"Hey, go use your competitor analysis
1:19:50workflow." But what it's going to do is
1:19:51it's going to think about what we have.
1:19:53So you can see that it just searched
1:19:54through our workflows. It searched
1:19:55through our tools. It said we already
1:19:56have a competitor analysis workflow set
1:19:57up. Let me read how this works and let
1:19:59me just go do it. So let's just see what
1:20:00it comes up with. I'll let you guys know
1:20:02if we have any questions. Otherwise,
1:20:03I'll check in with you guys when we get
1:20:04that output. Okay. So here we are
1:20:06getting some questions. So this is the
1:20:07part where it realized it didn't
1:20:08actually have enough information about
1:20:09us yet. So what's your primary target
1:20:11for get leads with AI? We're just going
1:20:13to go with um we'll just say marketing
1:20:15agencies. For pricing, I'm going to go
1:20:16with credit based usage. What's your key
1:20:19differentiator? I'm actually going to go
1:20:20ahead and say other and I'm just going
1:20:21to say our key differentiator is an
1:20:23all-in-one platform. But make sure
1:20:24you're saving all of this information
1:20:26that I'm telling you about my business
1:20:27somewhere in this project so I don't
1:20:29have to tell you it again. And so I'm
1:20:30pretty sure it would have done that
1:20:31either way because that's kind of the
1:20:32whole point here. But just to make sure
1:20:33that it does it for the sake of the
1:20:34example, I wanted to show you guys you
1:20:36have the ability to just tell it to do
1:20:37things. So got it. I'll save the
1:20:39information. So that's all you have to
1:20:40do. And now it's going to continue on
1:20:41with its to-do list. And actually it
1:20:42does come up with some more questions.
1:20:44So we're a single all-in-one product and
1:20:45our price range, let's just say 200 to
1:20:47500 a month. Now this is good. I know
1:20:49what you guys may be thinking is that's
1:20:50a lot of questions. Well, the thing is
1:20:52as you use it more and more, it gets
1:20:53smarter and smarter because each time
1:20:55you use it, you know, it has more
1:20:56information and you you give it
1:20:57feedback. So, yes, the initial setup may
1:20:59seem like a lot, but think about the
1:21:00questions it's asking and think about
1:21:02how good it's going to get now that it
1:21:03has all this info. So, you can see what
1:21:04it did is it created this file right
1:21:06here called business profile JSON. And
1:21:08this is where it decided to store all of
1:21:09the information about our business. And
1:21:11now, if we ever tell it something else
1:21:12and it needs to add like a new memory or
1:21:14fact about us, it will just go ahead and
1:21:15update this JSON file. Here's a great
1:21:17example of it fixing itself. So, it
1:21:18basically went ahead and started looking
1:21:20up for competitors and it found an
1:21:21error. So, I see there's a uni-ode
1:21:23encoding issue with the script on
1:21:24Windows. Let me fix that. It reads how
1:21:26to fix it and then it goes ahead and
1:21:27fixes it because there were some emoji
1:21:29characters or something like that. And
1:21:30then it said, "Let me update it." And
1:21:31what it's doing now is it's actually
1:21:32changing the script and changing the
1:21:34tool to make sure that that error
1:21:36doesn't happen again. It's also now
1:21:37created a new file called competitor
1:21:39list. So, it was able to do research and
1:21:40find different competitors like Apollo,
1:21:42Outreach, Clay, Instantly, Lemlist. And
1:21:45now if it ever needs to save more
1:21:46information about different competitors,
1:21:48it will just put it here. All right,
1:21:49looks like it's finishing up right now.
1:21:50So it found some key insights. You're
1:21:52positioned as a mid-market blah blah
1:21:53blah. Your eight main competitors, what
1:21:55you're doing right, critical gaps you
1:21:57need to address, top three
1:21:58recommendations. Let's see what those
1:21:59are. Add white labeling, introduced $99
1:22:01to $149 starter tier, and double down on
1:22:04build for agency's positioning. So it
1:22:05created three different files or sorry,
1:22:07four different files. It created the
1:22:08business profile, which we looked at. It
1:22:10created the analysis, competitor data,
1:22:12and the PDF report. Wow. So it created a
1:22:14new folder called competitors and it
1:22:15made an individual file for every single
1:22:17one of our competitors. So that's really
1:22:19cool. We can actually see a lot of data
1:22:20about them now. It created a folder
1:22:22called analysis history. So this is
1:22:23where we can see pretty much all of the
1:22:25data that it ran and got for this
1:22:26specific run. And now of course it has
1:22:28the PDF. So let's check out the final
1:22:30output. All right. So here it is. Now I
1:22:31can definitely say that these are my
1:22:33colors and the typography. So that's
1:22:34good. But I don't see the logo. And I
1:22:36really I think it's just because it is a
1:22:38white PNG logo. So I bet that it's up
1:22:40here. I just think that we probably
1:22:41can't see it. But anyways, we'll see if
1:22:42we can fix that. For now, we've got
1:22:44executive summary. We've got business
1:22:45profile. We've got competitive landscape
1:22:47with feature analysis. We've got
1:22:48competitor profiles, so high threat,
1:22:50medium threats. We've got all these
1:22:52different companies with strengths and
1:22:53our advantages. And then we also have
1:22:55our strategic recommendations at the end
1:22:57that we saw earlier. So, really the
1:22:59problem with this report is that our
1:23:00logo isn't visible because it's white.
1:23:03You can also see that it said that that
1:23:04run costed $143. So, not too bad. But
1:23:07what I'm saying now is that's great, but
1:23:08we can't see the charts or logos. I'm
1:23:10assuming because they're the same color
1:23:11as the background. Investigate and fix
1:23:12these issues. Now, typically I would put
1:23:14this in plan mode and go back and forth
1:23:15a little bit again, but for the sake of
1:23:17the demo, I want to see how good it's
1:23:18able to do when we just let it run with
1:23:20a super super vague request, as you can
1:23:22see. All right, so here's the thing. It
1:23:23said the PDF generator has several
1:23:25problems and listed those out. And now
1:23:26it's going to go ahead and fix those
1:23:27issues. So once again, this is just me
1:23:28telling you guys about you have to run
1:23:30the workflow a few times to discover
1:23:32those holes. And once you discover those
1:23:33holes, it'll fix them. And then you'll
1:23:35get to a place where you have more of a
1:23:36battle tested workflow. Okay, so it
1:23:38regenerated the PDF. And once again, we
1:23:40talked earlier about the caching. It's
1:23:41saving all the data it had already. So,
1:23:43it doesn't have to do a new search,
1:23:44which is really good. And in the future,
1:23:45it will still do current research, but
1:23:47it already has the business profile
1:23:49about all of our competitors, and it's
1:23:51already researched them. So, now it just
1:23:52has to see if there's anything new. But
1:23:53anyways, let's open up the new report
1:23:54and see how it looks. All right, so
1:23:56here's the new report. Okay, so we've
1:23:57got competitive intelligence report, get
1:23:59leads with AI, today's date, and now we
1:24:00do see the logo. So, executive summary,
1:24:02we can see this business profile once
1:24:04again, competitive landscape, competitor
1:24:06profiles. However, there we go. But we
1:24:08can finally see a pricing analysis chart
1:24:09which looks pretty solid. Cool. So, at
1:24:11this point, it would just be a matter of
1:24:12making tweaks cuz obviously this isn't
1:24:14perfect. There's some things we might
1:24:15want. We might want more details cuz
1:24:16like up here, you know, it's pretty it's
1:24:18not super super wordy and super
1:24:19detailed. So, maybe you like that, maybe
1:24:20you don't. At this point, you've got
1:24:22enough info and you've got every tool
1:24:24you kind of need. And you just go back
1:24:25and forth and ask for feature
1:24:26enhancements. And once again, you can do
1:24:27that all with completely natural
1:24:29language. But you can see I didn't have
1:24:30to go look at any API documentation. I
1:24:32didn't have to figure out how to prompt
1:24:33something to run a competitive analysis.
1:24:35I didn't have to go figure out how to
1:24:36generate these PDF reports or charts. It
1:24:38handled all of that for me. So, I hope
1:24:39you guys were able to follow along and I
1:24:40hope you're excited to go build your
1:24:42first agentic workflow. So, as you can
1:24:44see, building your first agentic
1:24:45workflow is actually so simple. But if
1:24:47you still have doubts about what you
1:24:48need to do or where to start, you can
1:24:49always join my community. In here, you
1:24:50can get every single resource that I've
1:24:52ever used in my YouTube videos. All the
1:24:54templates, the workflows, the prompts,
1:24:55the files, all completely free. Also, if
1:24:57you ever have a problem making something
1:24:58work, you've got tutorials in there on
1:25:00different topics, and the support of
1:25:01over a quarter million people willing to
1:25:03help you. The link for this is down in
1:25:04the description.
1:25:07All right. So, even if you don't know
1:25:08how to code or if you've never touched
1:25:10an IDE before, you're going to be just
1:25:12fine. IDE stands for integrated
1:25:13development environment, and you don't
1:25:14even need to know that it's said for
1:25:16that. So, what we're going to do is
1:25:17we're going to get into cloud code and
1:25:18I'm going to walk through everything
1:25:19that you need to know because it can be
1:25:21a little intimidating, but I'm going to
1:25:22show you exactly what you need to look
1:25:23at and what you don't need to look at.
1:25:25And by the end, it's going to be so much
1:25:26easier than you probably thought. So,
1:25:27the first step is you need to go to
1:25:28Google, search for Visual Studio Code,
1:25:30and then just download this. It's
1:25:31completely free to download. And this is
1:25:32where we're going to be using Claude
1:25:34Code. Once you open that up, this is
1:25:35what it's going to look like. And the
1:25:36first thing that I want you to do is go
1:25:37over to the lefth hand side and click on
1:25:39extensions. And once you get in here,
1:25:40just search for Claude Code. And then
1:25:42when you click on that, it's going to
1:25:43allow you to install the Claude Code
1:25:45extension for VS Code. And that's how we
1:25:47actually use it. So if you don't have a
1:25:48Claude plan, you are going to have to go
1:25:50get on a paid plan for Claude. You can
1:25:52start at 17 bucks a month. And this
1:25:54actually allows you to get Claude Code.
1:25:56As you can see, includes Claude Code
1:25:57with Opus 4.5. So you do have to be on a
1:25:59plan. And then once you open up the
1:26:00extension in here, it will prompt you to
1:26:02log in with that email that you have
1:26:03that plan for. And then it will
1:26:05basically sync it everything over here
1:26:06and you'll be able to use it. So the
1:26:08next step then is to open up a project.
1:26:10So on the lefth hand side, instead of
1:26:11clicking on extensions, you're going to
1:26:13click on explorer. And this says you're
1:26:14not in a project yet, you don't have a
1:26:16folder open, you need to open one up. So
1:26:18I've got a folder right in here called
1:26:19Agentic Workflows Demo. And that's the
1:26:21one that I'm going to open. If you don't
1:26:22already have one made, just go ahead and
1:26:23create one first and then you can open
1:26:25that up. And so you'll see if I click
1:26:26into this one, there's nothing in here.
1:26:27It's a completely blank project. So, I'm
1:26:29going to select that folder. And now we
1:26:30have this right here. So, this is our
1:26:32file explorer. This is where we can see
1:26:33Aentic Workflows demo. And then on the
1:26:35right hand side, what I'm going to do is
1:26:36click on this button up here, which
1:26:38looks like the Claude logo, and it says
1:26:39cloud code open. So, I open that up. I'm
1:26:41going to close out of this main window.
1:26:42And now, what we have is cloud code,
1:26:44which kind of looks like a chat GBT or a
1:26:46regular claude interface where we can
1:26:47talk with our coding assistant. So, this
1:26:49is what your screen should look like.
1:26:50Once you get here, let's talk about what
1:26:52comes next. Okay, the environment that
1:26:54we're currently working in, cloud code
1:26:55within VS Code. On the lefth hand side,
1:26:57we've got our files. So in ours right
1:26:58now we have one called agentic workflows
1:27:00demo but there's no other files in
1:27:01there. This is where cloud code will
1:27:03actually build workflows for us and
1:27:04build files and things like that and
1:27:06we'll see them populate on the lefth
1:27:07hand side. Now on the right hand side
1:27:08this is where we have our chat interface
1:27:10with the agent itself. This is where we
1:27:11do our planning. This is where it asks
1:27:13us questions and this is where it
1:27:14actually executes actions and once again
1:27:16we'll be able to see all of that live.
1:27:17So now I wanted to tell you guys about
1:27:19the framework that we're actually using
1:27:20today to build our agentic workflows.
1:27:22It's called WAT which stands for
1:27:23workflows agent and tools. So the agent
1:27:26itself is cloud code. That's who we talk
1:27:28with. That's what the AI brain uses to
1:27:30build workflows and tools. The workflows
1:27:32are going to come in a format called
1:27:34markdown, which just looks like this.
1:27:36It's natural language. It has headers
1:27:37and it has bullet points and bold font
1:27:39just to make it easier to read, but it's
1:27:41literally just a natural language
1:27:42document. And then the tools come in
1:27:44Python. So this is the logo. It'll be a
1:27:46py file, which I'll show you guys. And
1:27:48this is the ugly stuff. This is where we
1:27:50actually have code that I don't really
1:27:51want to look at. You guys don't want to
1:27:52look at, but luckily we don't have to.
1:27:54So, what's the difference between these
1:27:55workflows and tools? Well, workflows are
1:27:56processes and tools are actions to take.
1:27:59So, let's go back to our analogy of
1:28:00like, you know, food and maybe making a
1:28:03cake. So, when you want to make a cake,
1:28:04you've got a recipe and then you've got
1:28:05a bunch of ingredients and you have to
1:28:06figure out what to do with them. So,
1:28:07basically, the agent is a chef and the
1:28:09chef needs to make a cake. The chef is
1:28:11going to either read a pre-existing
1:28:13workflow, which is a recipe to how to
1:28:15make the cake, or the chef is going to
1:28:16build its own recipe. And within the
1:28:17recipe, it'll say, you know, like crack
1:28:19two eggs into a bowl, add a cup of
1:28:20flour, whatever. Those are the tools. So
1:28:22eggs are tools, flour is a tool, sugar
1:28:24is a tool. And so that's how the chef,
1:28:26the agent uses a combination of recipes,
1:28:29workflows, ingredients, tools in order
1:28:31to make something, which is either a
1:28:32cake or an agentic workflow automation.
1:28:34So now that you guys understand that
1:28:36framework, we need to make sure that cla
1:28:37code understands that framework. So what
1:28:38I'm going to do is I'm going to drag in
1:28:40a file. And this file will be available
1:28:42for download in my free school
1:28:44community. The link for that is down in
1:28:45the description. And this is our
1:28:46claude.md file. So every time that you
1:28:49set up a new project in cloud code, you
1:28:50have to give it a cloud. and file. They
1:28:52won't always be the same, but when
1:28:53you're building agentic workflows and
1:28:55you're using a WAT framework, you can
1:28:56just use this and copy and paste it
1:28:58every single time. This is basically
1:28:59telling Cloud Code how to work. This is
1:29:01its job instructions and description.
1:29:03So, if you were to go get a job at a
1:29:04grocery store on your first day, they
1:29:05wouldn't just let you loose. They would
1:29:06say, "Hey, we're going to get you
1:29:07onboarded. Here's what you do. Here's
1:29:09what you wear. You know, here is
1:29:11specific tasks you do." So, here we're
1:29:13telling the agent, you're working inside
1:29:14the WAT framework, which stands for
1:29:16workflows, agents, and tools. We have
1:29:17three layers. The first one is
1:29:18workflows, which are the instructions.
1:29:20The second one is agents, which is you,
1:29:22the decision-maker. And the third one is
1:29:24tools, which are the executions. So we
1:29:26talk about why this matters. The AI
1:29:27tries to handle every step directly,
1:29:28accuracy drops fast. So if each step is
1:29:3190% accurate, then you're down to 59%
1:29:33success after just five steps. So
1:29:34basically, we're just explaining why
1:29:36we're doing it like this. We then talk
1:29:37about how to operate. So you look for
1:29:39existing tools first. You learn and
1:29:40adapt when things fail. You keep your
1:29:42workflows current, blah blah blah. We've
1:29:44got a self-improvement loop. And then
1:29:45we've also got a file structure. So,
1:29:47like I said, Claude Code when we're
1:29:48working with it is going to create
1:29:49files. It's going to create tools. It's
1:29:51going to create maybe temporary docs to
1:29:52look at notepads. And when it does this,
1:29:54it adds them on the lefth hand side. So,
1:29:56if we don't tell Claude how to organize
1:29:57its files, it's going to get messy quick
1:29:59to the point where I don't understand
1:30:00where things are, and neither does
1:30:01Claude code. So, we're just giving it a
1:30:03nice structure for workflows, tools,
1:30:05temporary files, things like that. And
1:30:06so, obviously, you guys can read this
1:30:07whole thing. I don't want to spend time
1:30:09reading this line by line. But really
1:30:10the moral of the story here is this
1:30:12helps Cloud Code understand the
1:30:13framework we want to use, how to build
1:30:15workflows, so that when I'm talking to
1:30:16it, we're on the same page. All right,
1:30:18so what we want to do now is have Cloud
1:30:19Code read that and then set up our
1:30:21actual file structure. But before that,
1:30:22I wanted to show you guys one thing at
1:30:24the bottom when you're talking to Cloud
1:30:25Code, which is the mode. So you can be
1:30:27on ask before edits, you can be on edit
1:30:29automatically, you can be on plan mode,
1:30:31and you can be on bypass permissions. So
1:30:32all of that just gives Cloud Code a
1:30:34different level of autonomy. If you
1:30:35don't want it to do anything and you
1:30:36just want to make a plan, you start with
1:30:37plan mode. And this is really important.
1:30:39and I'll show you guys how this works.
1:30:40If you wanted to ask for edits, you can
1:30:41have it do that. If you want it to just
1:30:42edit automatically, you can have it do
1:30:44that. Or if you want to have it just
1:30:45bypass permissions and just completely
1:30:47go, then you choose that. Now, if you
1:30:48don't see this option in yours, you have
1:30:49to go to your settings, and then you'll
1:30:51type in Claude code, and then you will
1:30:52enable allow dangerously skip
1:30:54permissions. And I only really like to
1:30:55do this if I'm sitting next to Claude
1:30:57watching it work. And if I realize it's
1:30:59going off track, I can just kind of poke
1:31:00it and steer it back in the right
1:31:01direction. So, now that we've covered
1:31:03that, I'm going to be on bypass
1:31:04permissions mode, and I'm going to say,
1:31:05"Hey, Claude, I just gave you a claw.md
1:31:08file. I want you to go ahead and set up
1:31:09this project so that we're ready to go.
1:31:11We're ready to build a workflows
1:31:12together. So, as I shoot that off, you
1:31:14can see that this is very similar to
1:31:15chatbt except for in here, we can see
1:31:17everything that it's thinking and doing.
1:31:19So, let's just start from the top. It
1:31:20says, "Okay, cool. Let me read the
1:31:21cloud.MD file to understand the project
1:31:23setup requirements." And then we can see
1:31:24it actually did this action. It read it
1:31:26read this file that lives here. And now
1:31:27it said, "Okay, cool. I understand the
1:31:29WT framework. Let me check what already
1:31:30exists in the project and then I'll set
1:31:32up the required structure." As you can
1:31:33see, it then goes to list the current
1:31:35project contents. And there's nothing.
1:31:36The product is empty except for the
1:31:38claw.md file. Let me set up the right
1:31:39structure. And so you can see it creates
1:31:41a to-do list. It thinks, it searches, it
1:31:43updates the to-dos, and it basically
1:31:44goes through this step-by-step process
1:31:46that you can have full visibility into
1:31:48and see what it's doing, see how it's
1:31:49thinking until we're done. So now it
1:31:50looks like the to-dos are pretty much
1:31:51all the way done. And it's going to come
1:31:52back and say, "Nice, we're all set up.
1:31:54What do you want me to do next?" And you
1:31:55can actually see in real time on the
1:31:56lefth hand side, we now have different
1:31:58files. We've got our temporary folder,
1:31:59we've got our tools folder, we've got
1:32:00our workflows. Obviously, there's
1:32:02nothing in these yet besides just some
1:32:03read me and some basic stuff, but that's
1:32:05why we gave it this folder structure so
1:32:07that it keeps it organized and it
1:32:08doesn't just throw a bunch of things in
1:32:09a random order. Cool. So, it says ready
1:32:10to build a workflow. Let me know what
1:32:12you want to accomplish. Awesome. So,
1:32:14we're all set up. Let's actually start
1:32:15talking about the workflow that we want
1:32:16to build. So, what I've got is this list
1:32:18of remote jobs. So, I searched for
1:32:20social media. There's 622 remote jobs.
1:32:22And let's say I want to apply to all of
1:32:24these. Well, that would be really tough
1:32:25to log all of these manually. And
1:32:26there's multiple pages. There's 21 pages
1:32:28of these jobs. So, what we can do is we
1:32:30can have Claude Code look at this stuff,
1:32:32get all the jobs we need, and then put
1:32:33it into an Excel sheet for us. And for
1:32:35this, we're going to be using a tool
1:32:36called Firecraw. Firecrawl lets us do
1:32:38tons of different actions, and it lets
1:32:40us basically take a website like
1:32:41McDonald's right here. I can drop in the
1:32:43URL, and I can ask for a scrape, and
1:32:45it's going to go ahead and grab all of
1:32:46the information from that website. So,
1:32:48in this case, I just requested markdown,
1:32:50and it just pulls back all of the text
1:32:51from the website. As you can see, it is
1:32:53a lot more powerful than that. It can
1:32:54turn websites into LLM ready data.
1:32:56Whether that's scraping the data,
1:32:57getting screenshots, mapping the data,
1:32:59crawling the data, searching,
1:33:01extracting. There's a lot of different
1:33:02things that we can do with firecrawl.
1:33:04Now, the thing is we want to just say,
1:33:05"Hey, cloud code, use firecrawl. Just go
1:33:07after it. Use whatever the different
1:33:08tools that firewall offers in order to
1:33:10accomplish the job that I've got for
1:33:12you." And so, we do this using a
1:33:13framework called MCP, which stands for
1:33:15model context protocol. Now, I know that
1:33:17this may just sound like some tech
1:33:18jargon or some gibberish. So, let's try
1:33:19to contextualize this a little bit.
1:33:20Think about Gmail, for example. In
1:33:22Gmail, you've got an action to send an
1:33:24email. You've got an action to draft an
1:33:25email. You've got an action to get a
1:33:27bunch of emails. There's so many
1:33:28different tools within that tool. So,
1:33:30MCP basically says, "Okay, cool. The
1:33:31agent is going to figure out how to use
1:33:33all the tools, when to use all of them,
1:33:35what parameters to fill over, all that
1:33:36kind of stuff so that you, the human,
1:33:37don't have to think about that. So, if
1:33:38we go back to our example of making a
1:33:40cake, let's say we realize, okay, so for
1:33:42this cake, we need eggs, flour, and
1:33:44frosting." Okay, well, how do we do
1:33:45that? Well, let's just give our agent
1:33:47access to the supermarket MCP and say,
1:33:49"Okay, whenever you need a new
1:33:50ingredient, just go to the supermarket
1:33:52MCP, grab what you need. I don't really
1:33:53care. Just figure it out and then come
1:33:55back with the right ingredients." Rather
1:33:56than saying, "Okay, cool. So, like eggs,
1:33:58let's go to the egg store. Flour, let's
1:34:00go to the flower store. Frosting, let me
1:34:01go to the candy store." We're just going
1:34:02to get everything in one spot. And
1:34:04that's the power of MCP. So, in
1:34:05Firecalls Docs, you can see that they
1:34:07have an MCP server, which is amazing.
1:34:08And this lets us get stuff like web
1:34:11scraping, crawling, searching, all this
1:34:13kind of stuff that we want in any of the
1:34:14tools that we want to use. So right
1:34:15here, I can click on running on cloud
1:34:17code. And this shows us how to add the
1:34:19firewall MCP server using the Cloud Code
1:34:21CLI. So what I'm going to do is go ahead
1:34:23and copy this message right here. I'm
1:34:25going to come into Cloud Code and I'm
1:34:26going to clear out this conversation
1:34:27history. And I'm going to say, "Hey
1:34:28Claude, I want you to help me install
1:34:30the Firecrawl MCP server. You need to
1:34:32install it using this command in the
1:34:34Cloud Code CLI." And then I paste in
1:34:36that command. And what you'll notice
1:34:37here is that it's prompting us for our
1:34:39API key. And so an API key is basically
1:34:41like a password. And I don't actually
1:34:42want to give it to cloud code. I don't
1:34:43want my API key to be stored in the
1:34:45conversation history of cloud code. I
1:34:47want to just put it into the file or
1:34:49into the project locally myself just to
1:34:51do it a bit more secure. So I'm going to
1:34:52say go ahead and get this initialized,
1:34:54but I'm not going to give you my API key
1:34:56directly. I'm going to put it into the
1:34:58EMV file. So help me get that set up as
1:35:00well. So I shoot off that message. It's
1:35:02going to think about how to actually
1:35:03help me set all this up. So it's
1:35:04checking in on the existing file. It's
1:35:06checking in on our project configuration
1:35:07and then it's going to help us actually
1:35:08do this. Okay, so it said that I've
1:35:10added the API key placeholder to yourv
1:35:12file. Now you just need to add it there.
1:35:13So thev file is on this lefth hand side.
1:35:16I'm going to open that up and you can
1:35:17see now it says cool firecall mcp server
1:35:19put in your API key right here. So I'm
1:35:21going to delete this. I'm going to go
1:35:22into firecrawl and I'm going to go to my
1:35:23dashboard. So if you haven't already
1:35:25sign up for firecrawl, you can get
1:35:26started for free and you can get 500
1:35:27credits right away which is more than
1:35:28enough to play around with. And then you
1:35:30can see right here API key. So I'm going
1:35:31to copy this. I'm going to paste it
1:35:33right here into the actual MV file. And
1:35:35then I'm going to go to file and save
1:35:37this to make sure that it actually gets
1:35:38saved on our project. And then you can
1:35:40see it says then run the MCP add
1:35:42command. And it gives me all this
1:35:43reasons and I actually don't understand
1:35:45what this means. So I'm just going to
1:35:46ask it to see if it can do it itself. I
1:35:48don't exactly understand how to do that.
1:35:50I have added my firewall API key to thev
1:35:53file. Would you be able to actually just
1:35:55run this command to make sure we can
1:35:56install the firewall MCP server? Okay,
1:35:58so it went ahead and installed the MCP
1:36:00server. Now, something I did want to
1:36:01bring up is that when it actually runs
1:36:03that command with this bash operation,
1:36:05it does put in the API key right here,
1:36:07which technically will be stored in
1:36:08conversation history. So, in this case,
1:36:10we're fine because this is a free key.
1:36:11It doesn't have much access, and I'm
1:36:13probably just going to rotate it right
1:36:14after this video. So, what you would
1:36:15want to do in an environment where you
1:36:17have a key that has a lot of risk is you
1:36:19would want to have Claude Code just walk
1:36:20you through how you can run it in your
1:36:22own terminal in order to make sure that
1:36:23Claude never actually touches the key.
1:36:25But still best practice always store
1:36:26them in av way if you ever are pushing
1:36:28something to a public repository
1:36:30anywhere or someone gets access to your
1:36:32files then that's all going to be
1:36:33encrypted. So now we are pretty much set
1:36:35up for actually starting to build this
1:36:37agentic workflow. So I'm going to do a
1:36:38/cle start a new conversation and what
1:36:40we're going to do is like I said earlier
1:36:42we're just going to explain in super
1:36:43clear natural language what we want. So
1:36:45I'm going to go ahead and switch this to
1:36:46plan mode which I would always recommend
1:36:47doing before you actually start building
1:36:49an agentic workflow. I'm going to go
1:36:50back into this tab and I'm going to grab
1:36:52the URL from this page where we have 622
1:36:55job opportunities here for social media.
1:36:57Coming back into Cloud Code, I'm just
1:36:58going to start talking to it. So, I'm
1:36:59going to paste in this URL and I'm going
1:37:01to say, "Hey Claude, I just gave you a
1:37:03URL for a website that has a bunch of
1:37:05job opportunities. There are about 622
1:37:07job opportunities here, but they're
1:37:09spread across different pages." So,
1:37:10there's like 21 total pages. And what I
1:37:12want you to do is go ahead and scrape
1:37:13all those for me, and I just want you to
1:37:15put those into an Excel sheet so that I
1:37:17can actually look through them, you
1:37:18know, do things with them. make sure
1:37:19you're getting all of the relevant
1:37:20fields that I may want. So, that's kind
1:37:21of my overall plan. Let me know if you
1:37:23can help me make that project
1:37:25requirement more robust and you can feel
1:37:27free to ask me any questions that you
1:37:28may have for me to make sure that we can
1:37:29build out this workflow in a really high
1:37:32quality way. So, just shot that off in
1:37:33plan mode. So, it's going to do a lot of
1:37:35thinking. It's going to reason about
1:37:36like the way that we should actually do
1:37:38this. Hopefully, it understands that it
1:37:39can use Firewall's MCP server. As you
1:37:41can see, it's searching that right here.
1:37:42And then what it's going to do is it's
1:37:43going to come back to us with tons of
1:37:44questions, I'm sure. And then it's also
1:37:46going to come back to us with a plan. So
1:37:47once I'm able to approve that plan, it
1:37:49will start actually building the
1:37:50workflows and the tools for us. All
1:37:52right. So here we are with our first
1:37:53round of questions. Do you want me to
1:37:54also scrape each individual job detail
1:37:56for more complete info like company
1:37:58name, full description, benefits or just
1:38:00the listing? So for the for right now,
1:38:02we're just going to go with the actual
1:38:03listing. For the output location, can we
1:38:05So output location, it says where should
1:38:06I save the Excel file? I'm just going to
1:38:08go ahead and do a local in the temporary
1:38:10folder which we've created right over
1:38:11here. So I'll choose that. And then for
1:38:13filtering, do you want any filters
1:38:14applied or should I grab all 622 job
1:38:17posts? Now, what I'm going to say here
1:38:18is other and I'm going to go ahead and
1:38:20grab all the jobs. But I'm doing this as
1:38:22a demo for how to build an agentic
1:38:24workflow. So just go ahead and grab only
1:38:25200 for now just to prove that this
1:38:27concept works. So I've submitted those
1:38:29answers and now you can see it's going
1:38:30to keep on going with its plan. So we
1:38:32just got back this plan. I'm going to go
1:38:33ahead and give this a quick review, but
1:38:34also for the sake of the demo, I want to
1:38:36see how well it did on the first shot.
1:38:38So realistically, what I would do is I
1:38:40would go ahead and read this whole
1:38:41thing, and if there were any adjustments
1:38:42to be made, I would go ahead and make
1:38:43those. But what you can see is it gave
1:38:44us a pretty comprehensive plan of what
1:38:46it's going to do. It's going to create a
1:38:48tool called scrape daily remote. It's
1:38:49going to create a workflow called scrape
1:38:51job listings, and then it's going to
1:38:52actually execute that scrape and get us
1:38:54all this information. So I'm going to go
1:38:55ahead and say yes and auto accept. So it
1:38:57just spun up that to-do list. It's going
1:38:58to start going, and I'll check in with
1:38:59you guys when that's done. All right,
1:39:00that just finished up. We can see that
1:39:02we got 209 done. We have different
1:39:04metrics here. We have different
1:39:05locations and also created a tool and a
1:39:08workflow. So if I open up this folder,
1:39:09we can see we now have a scrape daily
1:39:11remote jobs tool. And in the workflows,
1:39:13we now have a scrape job listings
1:39:15workflow. So basically meaning next time
1:39:16we ask it to scrape jobs, it's going to
1:39:18be able to do it better and it has more
1:39:20direction because it's already done it.
1:39:21And if it has any mistakes, it will
1:39:22update the workflows and the tools so
1:39:24that the next time it's even better. But
1:39:26let's go ahead and take a look at the
1:39:27output. So it said that it stored it in
1:39:29the temporary file. So right here, temp,
1:39:31and we've got social media jobs Excel.
1:39:33All right. So, this is the Excel sheet
1:39:35that it created for me. I'm going to go
1:39:36ahead and zoom out a little bit so we
1:39:37can see, but it's got like different
1:39:38filters on here already, which is pretty
1:39:40cool that it did all this itself. We've
1:39:42got job title, we've got job type,
1:39:44position, location, experience,
1:39:45category, salary, description, summary,
1:39:48tags, we've got the actual URL, and
1:39:50that's pretty much it. And like it said,
1:39:51it was able to get us, I think it said
1:39:52209. So, yep, this is 209 total job
1:39:56postings just like that. So, we're going
1:39:57to go ahead and try a different use case
1:39:59now that uses the Firewall MCP. But what
1:40:01I wanted to show you guys is down here
1:40:03we have context. So it says 45% of your
1:40:05context is remaining until it will
1:40:07autoco compact. And so you guys have
1:40:08might have heard of something called
1:40:09context rot. It basically just means the
1:40:11more context that you have in a
1:40:12conversation history with an AI model,
1:40:14the worse it kind of gets. So typically
1:40:16whenever my context gets over 60, I'll
1:40:18probably just compact it or reset and
1:40:20then keep going. So what I'm going to do
1:40:21is I'm just going to go ahead and click
1:40:22this right now and it's going to compact
1:40:24our workflow and all of our conversation
1:40:25history so we can keep going, but it
1:40:27still remembers the important things
1:40:28that we've done. All right, so that's
1:40:29been compacted. And so it basically just
1:40:30summarized everything that we've done.
1:40:32So let's try something else now. Let me
1:40:33look at this same website, but now
1:40:36there's no search filter. So there's
1:40:37214,000 jobs. I'm just going to take the
1:40:39URL and I'm going to go into cloud code
1:40:41and say, "Hey Claude, I've got this URL
1:40:43that I need help with. I want you to
1:40:45basically be able to scrape this. I want
1:40:47jobs that are sales opportunities or
1:40:50sales jobs and I want to look just in
1:40:52Europe. Scrape all of this. I want to
1:40:54get like 500 jobs back and put it in a
1:40:56nicely formatted Excel sheet for me."
1:40:58Now, this time what I'm going to do is
1:40:59we're just going to go ahead and do
1:41:00bypass permissions. And like I said,
1:41:01normally you want to go on plan mode.
1:41:02You want to ask questions, but I just
1:41:03want to show you guys what this might
1:41:05look like with a pretty vague prompt and
1:41:07just letting clog code go after it. And
1:41:09hopefully it's able to do a better job
1:41:10now because it understands how it can
1:41:12scrape job listings and it has this tool
1:41:14and it may even have to create a new
1:41:15tool. So, let's just kind of let it run.
1:41:16I'm going to analyze what it's doing and
1:41:18then I'll report back once we see what
1:41:20it actually ended up doing. And start
1:41:21off by saying, cool. I have a scraping
1:41:23tool from before that I can use. So,
1:41:24here's my plan. I'm going to scrape the
1:41:25sales jobs. I'm going to filter for
1:41:27Europe and then I'm going to export it
1:41:28all to an Excel sheet. It found 409
1:41:31total sales jobs and now it has to
1:41:33filter for Europe. And when it did that,
1:41:34it basically found there's a limitation.
1:41:36There's only 52 sales jobs in Europe,
1:41:38but there are 409 total. So, let me
1:41:40check if including worldwide would help
1:41:42get closer to the goal. So, what's going
1:41:43on here? It has our natural language
1:41:45request, which was that we wanted 500
1:41:47sales jobs in Europe. And it realized,
1:41:48okay, this actually isn't going to work.
1:41:50Let me brainstorm and see what else I
1:41:52can do. And now you can see what it did
1:41:53is it asked us a question because it
1:41:55wants to help us reach our end goal
1:41:56better. So it says, "Do you want me to
1:41:58expand the search to get more jobs?" We
1:41:59could either keep the 55 sales jobs in
1:42:01Europe, we could broaden it to all job
1:42:02types, or we could also do US-based
1:42:05sales jobs. So I'm just going to go add
1:42:06US jobs as well. And now it should keep
1:42:08going. And hopefully now you can also
1:42:09see why I wanted to give it a temporary
1:42:11folder because in this operation where
1:42:12it's running into a few issues, it's
1:42:14creating some other temporary files like
1:42:16all sales jobs, sales jobs raw. It's
1:42:18also created three different Python
1:42:20tools that are temporary tools just
1:42:22because it knows that it needs some help
1:42:23filtering things out. But now it was
1:42:25able to find 372 sales jobs and it was
1:42:28saved to sales jobs Europe and US. So it
1:42:30says that we've got 372 sales jobs. I'm
1:42:33going to open up that Excel sheet and we
1:42:34can see if we scroll all the way down,
1:42:36we should have gotten 372. Perfect. And
1:42:38this is also similar because we can
1:42:40filter up here with all of these
1:42:41pre-made filters that it put in. We've
1:42:43got job title. We've got all this
1:42:44information. And it said that it added a
1:42:46region column right here. And this is
1:42:47where we could get rid of US and mixed
1:42:49and worldwide. And we now should see
1:42:51that we've only got about different
1:42:52actual rows. Yeah, 49. And just as one
1:42:54final test, let's see what happens if it
1:42:56gets a crazy type of request that we
1:42:58haven't really prepared it for. So right
1:42:59now, this is good at scraping jobs from
1:43:01a given URL. What if I just said, "Hey
1:43:02Claude, I'm looking to reach out to tons
1:43:05of dentists. Can you find me dentists in
1:43:07the United States? And give me their
1:43:09contact information so that I can
1:43:11basically just build up a lead list of
1:43:12dentists that I can contact. I want this
1:43:14to be in an Excel sheet. I'm going to
1:43:15shoot that off. Once again, this is in
1:43:17bypass permissions mode. So, we're going
1:43:18to see what it does. We're going to see
1:43:19if it uses fire crawl. We're going to
1:43:20see if it thinks about, hey, actually, I
1:43:22can't use firecraw. I need to get access
1:43:23to some sort of like, you know, lead
1:43:25generation API or lead list API. We're
1:43:27going to see what it does here. It said,
1:43:28I can help you build a dentist lead
1:43:30list. Let me first check what tools and
1:43:31workflows exist. Then it says, let me
1:43:33check the available APIs. So you have
1:43:34fire call available and I can build a
1:43:36similar tool to your existing job
1:43:38scraper but for dentist leads instead.
1:43:39All right, so look at this. It used
1:43:40firecol to search for dentist directory
1:43:42and then it started scraping those
1:43:44sources. Once again, the ADA site uses
1:43:46JavaScript to load results dynamically.
1:43:48So the static scrape doesn't work. Let
1:43:49me try a different method. So it found
1:43:50out yellow pages works well. There are
1:43:523,000 dentists just in New York City.
1:43:54And now I'm going to create a scraping
1:43:55tool so I can actually do all of this.
1:43:57Okay, so it looks like this is finishing
1:43:59up here. And what you guys can see is
1:44:00that it created a new workflow which is
1:44:02scrape dentist leads. And it created a
1:44:04new tool which is scrape dentist leads.
1:44:06All right, we ran into another issue.
1:44:07Only two dentists were found. The
1:44:09parsing might need adjustment. Let me
1:44:10check what was captured and then refine
1:44:12the reg x pattern. So look how awesome
1:44:14this is. It found the issue right here.
1:44:15And now what it's doing is it's fixing
1:44:17the tool. So it's updating the tool so
1:44:18that it doesn't actually run into that
1:44:20issue again. Okay. I mean look at this.
1:44:21It said done. I've scraped 120 unique
1:44:23dentist leads from four major cities.
1:44:25Here are the cities we got. It includes
1:44:26all of this data which is awesome. And
1:44:28then it says for future scrapes, I've
1:44:30also created a reusable tool that you
1:44:31can run any time. So, I'm going to open
1:44:33up that Excel sheet right here. And we
1:44:34can see that we do indeed have all of
1:44:36these different dentists here. It even
1:44:37formatted the Excel sheet a little bit,
1:44:39but we have phone number, we have
1:44:40address, we have city, state, zip code,
1:44:43websites, specialties, and we get the
1:44:45actual listing URL as well. So, this is
1:44:47incredible if you think about the fact
1:44:49that I didn't know what tools to use at
1:44:51all. I could have put this in plan mode
1:44:52and I could have said, "Hey, this is the
1:44:54workflow I want. Ask me questions, do
1:44:55research, figure out the best approach,
1:44:57and then I could have, you know, went
1:44:58back and forth with a little bit and
1:44:59this scrape might have even been better.
1:45:01But on the very limited amount of
1:45:02information I gave it, it still gave us
1:45:04a really good output that would have
1:45:05taken me so much longer to get manually
1:45:07or building it in an end because the
1:45:09truth is a lot of us know what we want.
1:45:10We know the end result, but we don't
1:45:12exactly know the exact tech stack and
1:45:13all of the different things that we need
1:45:14to get that end result. So why not let
1:45:16an AI agent with a really smart brain
1:45:18like Opus 4.5 figure that out for us,
1:45:21look at five different approaches, and
1:45:22then pick the best one. And the cool
1:45:24part is you're not just limited to one
1:45:25agent. You could open up five different
1:45:26agents in here. As you can see, we could
1:45:28just keep stacking agents on agents. And
1:45:30then what I could do is I could just
1:45:31tell all of them to try a different
1:45:32method. So I could have four different
1:45:33workflows running and then I could test
1:45:35all four at the same time and whichever
1:45:36one gives me the best result, I would
1:45:38just delete all the other agents and
1:45:39then stick with that main workflow. So
1:45:40before we wrap up here, I actually
1:45:41wanted to just contextualize one more
1:45:43time what's going on. So let's take a
1:45:44look at that first workflow we did. This
1:45:46one was called scrape job listings. This
1:45:48one says the objective is to scrape job
1:45:50listings from daily remote.com based on
1:45:51a search term. The required inputs are
1:45:53search term, max pages, output path. The
1:45:56tools to use are just this one, the one
1:45:58that we created called scrape daily
1:45:59remote jobs. And then it goes through
1:46:01the exact steps and the exact outputs,
1:46:03edge cases, error handling, all of the
1:46:05stuff that I didn't tell it to do. It's
1:46:06because of the framework. And it's
1:46:08because it understands how to fail
1:46:09safely. So that every single time
1:46:10whenever we say, hey, I want you to
1:46:12scrape leads from daily remote. It just
1:46:13invokes this workflow, which inside of
1:46:15it invokes the tool. So it's the exact
1:46:17same thing that just happened for
1:46:18scraping dentist leads as you can see.
1:46:20So that's how simple it is to build an
1:46:21agentic workflow. But before you go to
1:46:23actually build one by yourself, you do
1:46:24need to understand the mistakes that
1:46:26most people are making right now.
1:46:27Because understanding how to think about
1:46:28these systems is what's actually going
1:46:30to make it valuable to you. So the first
1:46:31mistake is not being clear enough about
1:46:33the actual goal. You can't just say, "I
1:46:34need a lead scraper for LinkedIn."
1:46:35That's way too vague. Agent will have no
1:46:37idea what kind of leads you want, what
1:46:38industry, what role. It'll just start
1:46:39pulling random profiles. Obviously, it
1:46:41can ask you questions, but you do need
1:46:42to be specific about the problem that
1:46:44you're actually trying to solve. And so
1:46:45what you're going to want to do, as you
1:46:46saw a little bit in the demo, is put the
1:46:48agent in plan mode and say something
1:46:49like, "Hey, here's a rough idea of what
1:46:50I want. help me turn this into an actual
1:46:52solid PRD or project requirement doc.
1:46:55The agent then will have to brainstorm
1:46:56and it will reason and it will think and
1:46:57it will maybe even do research for you
1:46:59and it will ask you all the right
1:47:00questions so it knows exactly what to
1:47:01build. Just like the way if you wanted
1:47:03to give an actual human software
1:47:04developer specs for an app or for a
1:47:06workflow or whatever it is, you would
1:47:08have to give them enough information so
1:47:09that they could actually build that.
1:47:10You're totally allowed to treat the
1:47:11agent like the expert. You're just the
1:47:13manager to make sure that you keep it on
1:47:14the right path. So mistake number two is
1:47:16not defining what done looks like.
1:47:18Agents need to know when to stop. If you
1:47:19don't give them the clear finish line,
1:47:20then they may over complicate things or
1:47:22break things or keep researching or keep
1:47:24looping, keep iterating, and they might
1:47:26just keep wasting time when the answer
1:47:27was actually simple. I've definitely
1:47:28seen agents overcomplicate a lot of
1:47:30things. So, instead of saying, "Search
1:47:31for LinkedIn profiles of CEOs at tech
1:47:33companies," which is pretty open-ended.
1:47:34Say something like, "I need exactly 75
1:47:37LinkedIn profiles of CEOs at tech
1:47:38companies. Put them in a spreadsheet
1:47:40with their name, company email, their
1:47:41link to their profile, and once you have
1:47:4375, you're done." It's a clear input and
1:47:45a very clear output, and that's how
1:47:46you're going to get consistent results.
1:47:47So now let's talk about why agentic
1:47:49workflows are just better. First, no
1:47:50more debugging loops. With traditional
1:47:52workflow automation, you'd build
1:47:53something, you'd run it, and then there
1:47:54would be some edge cases that you didn't
1:47:55think of, and that would break the
1:47:56system. So then you'd spend the next
1:47:58hour reading through the logs, looking
1:47:59at the error messages, looking at the
1:48:01execution data, and trying to figure out
1:48:02what went wrong and why. With agentic
1:48:04workflows, the agent basically handles
1:48:05all of this for you without even asking.
1:48:07You saw it earlier in the build. As the
1:48:08agent was working, it would run into
1:48:09some sort of roadblock or it would hit
1:48:11an error and it would just say, "Okay,
1:48:12this is what happened. Let me think
1:48:13about what I could do differently." And
1:48:14then I fix it. And then I update my
1:48:16workflows and my tools so that it
1:48:17doesn't happen again. It's basically
1:48:18self-healing. And that's a massive
1:48:19timesaver because this means that I can
1:48:21have an agent on my right monitor
1:48:23building stuff and then on my left
1:48:24monitor I can just be doing different
1:48:25work or maybe even watching a YouTube
1:48:27video or catching up on my favorite show
1:48:28and I've got cloud code right here
1:48:29building things for me and all I have to
1:48:31do is sit here and make sure I can poke
1:48:32it in the right direction if I need to
1:48:34every once in a while because at the end
1:48:35of the day it is AI and it is not a
1:48:36deterministic so it might veer off the
1:48:38path a little bit. Second is natural
1:48:40language control. With tools like ended
1:48:42in you had to pretty much learn every
1:48:43node. You had to know what each one did,
1:48:45when to use each one, and what all of
1:48:46the different parameters or settings
1:48:48meant. If you wanted to connect to an
1:48:49API, you had to read the API
1:48:50documentation. You had to find the right
1:48:51endpoint. You had to structure your JSON
1:48:53correctly. You had to set up the
1:48:54authentication. And that could be a lot,
1:48:55especially when you are new to the
1:48:56space. With the Genic workflows, you
1:48:58just explain what you want, and the
1:48:59system will look at all the tools
1:49:00available, whether it has an MCP server
1:49:02or not, or whether it just has to look
1:49:03through and research the API
1:49:04documentation on its own. And this is
1:49:06absolutely beautiful. Third, it gets
1:49:08smarter over time. So, I know we've
1:49:09talked about this a lot, but it's just
1:49:10so cool. If you wanted to update an
1:49:11automation in the past, you had to go in
1:49:12and you had to change the nodes and you
1:49:14had to configure it manually. With
1:49:15Agentic Workflows, every time the agent
1:49:17runs into an issue, it learns and it
1:49:19updates. Now, there is one important
1:49:20caveat that I wanted to talk about,
1:49:21which is the difference between
1:49:22automations that you trigger yourself
1:49:24versus automations that run on a
1:49:25schedule. So, if I'm sitting at my desk
1:49:27using Claude Code and I say, "Hey, you
1:49:28know, we just had a call. Go ahead and
1:49:29write up a proposal for client B."
1:49:31That's a human triggered event in this
1:49:32case and the agent's right there with
1:49:33me. So, I can watch it, I can talk to
1:49:35it, and that's how it's able to
1:49:36self-heal in real time. But if you want
1:49:38something to run on a schedule like
1:49:39every morning at 6 a.m. or maybe an
1:49:41event trigger like whenever someone
1:49:42submits a form on your website or
1:49:43something like that, that's actually
1:49:44going to be you deploying that code, not
1:49:46the actual agent. So the agent would
1:49:48deploy its workflows and tools, but not
1:49:50itself, not the cloud code model that
1:49:51lives in VS Code. And the agent is what
1:49:54actually makes the workflows and tools
1:49:55self-healing. So you're not deploying
1:49:56that. But anyways, I'm not going to dive
1:49:58deep into that right now. That's a whole
1:49:59other video. You can also check out this
1:50:01video which I will link right above up
1:50:02here where I go into pretty much that
1:50:04whole process of building an automation
1:50:05in cloud code and then actually
1:50:06deploying it just so you guys can see
1:50:07what that actually looks like. So look,
1:50:08I know that this might feel a bit
1:50:09overwhelming at first. The space is
1:50:11moving really fast, but the reality is
1:50:12that this is just the beginning. We're
1:50:13definitely headed towards like fully
1:50:15autonomous workflows, agents managing
1:50:16other agents and systems that improve
1:50:18themselves while you sleep. And the ones
1:50:19who understand how to make them faster
1:50:20will be ahead in this automation market.
1:50:22So I don't want you to worry if you're
1:50:23still learning how to make automations
1:50:24or you're still learning end to end.
1:50:26Your job isn't over and I think that's a
1:50:27really good place to start. But now
1:50:28we're just moving from builders to
1:50:30architects. The key thing that matters
1:50:31is how you adapt to new challenges. And
1:50:33if you want to make it easier, you can
1:50:34check out my free community with over
1:50:36200,000 AI builders like you. And I put
1:50:38everything that we talked about today
1:50:39into a completely free resource guide
1:50:41you can access in that community. Link
1:50:42for that is in the description.
1:50:46Cloud Code has been allowing me to build
1:50:48things that used to take me hours in
1:50:50just minutes. So that's exactly what I'm
1:50:51going to be teaching you guys today.
1:50:53Even if you don't know how to code and
1:50:54even if you've never touched an IDE
1:50:56before. IDE stands for integrated
1:50:58development environment, but if you
1:50:59didn't know that, it's still completely
1:51:00fine. It's crazy how fast the technology
1:51:02is evolving every single day. What used
1:51:04to take people this long with manual
1:51:06code was significantly reduced when
1:51:08Eniden came out because we could drag
1:51:09and drop nodes and build workflows that
1:51:11way. And now that has once again been
1:51:12significantly reduced with the release
1:51:14of things like cloud code and
1:51:15anti-gravity. Now, I'm not out here
1:51:17saying that Naden is dead or that cloud
1:51:19code completely replaces any. They're
1:51:21slightly different. But I am going to
1:51:22show you how easy it is to build
1:51:23automations with cloud code today. If
1:51:25you've never touched Claude Code before
1:51:27or even watched a video about it, you're
1:51:28in the right spot because my job is to
1:51:30make confusing things as simple as
1:51:32possible. So, in today's agenda, I'm
1:51:33going to be going over the interface,
1:51:35what do you need to know, because
1:51:36there's a lot of stuff, but I'm just
1:51:37going to tell you what's actually
1:51:38important to understand. We're going to
1:51:39go over the framework that we use to
1:51:40actually build automations. I'm going to
1:51:42talk about planning and the importance
1:51:43of clear communication. We're going to
1:51:45talk a little bit about the superpowers
1:51:46that you can give cloud code like MCP
1:51:48servers and skills. We're going to talk
1:51:50about testing and how you actually
1:51:51optimize your workflow. and then talk
1:51:53about deployment, which means actually
1:51:54kind of turning it on or pushing it into
1:51:56production. And I'm not just going to be
1:51:57talking throughout all of this. I'm
1:51:59actually going to build a full workflow
1:52:00in front of you guys and deploy it by
1:52:02the end. So after this video, you'll
1:52:04have everything that you need to go
1:52:05build your first automation in Cloud
1:52:06Code. And you're going to see how easy
1:52:08it really is. All right, so we're just
1:52:09going to jump right into it. This is the
1:52:11interface. We're going to be using
1:52:12Visual Studio Code, which has been
1:52:14around for a long time. And if you go to
1:52:16Google and type in VS Code, you can just
1:52:18go ahead and go to this link and just
1:52:19download it. It's free to download. And
1:52:21then in here is where we're going to
1:52:22actually be using Claude Code. So this
1:52:24is what it should look like. What we're
1:52:26seeing here is just kind of the welcome
1:52:27page. You can see we can open new files,
1:52:29new folders. We can do some of these
1:52:30walkthroughs. But what I'm going to do
1:52:32here is I'm going to go over to this
1:52:33lefth hand side and click on extensions
1:52:36and just type in Claude Code. And then
1:52:38you'll see right here that this
1:52:39extension pops up which lets us use
1:52:42Claude Code inside of VS Code. So what
1:52:44you're going to do here is go ahead and
1:52:45install it. You could also do this in
1:52:47anti-gravity or in cursor or somewhere
1:52:49else or you could even use the cloud
1:52:50code kind of app by itself and install
1:52:52that locally, but wherever you choose to
1:52:54use it, you're going to log in and then
1:52:56we'll get started. I'm just using VS
1:52:58Code in today's tutorial. It'll prompt
1:53:00you to sign in with your Anthropic
1:53:01account and then you'll be all set. Now,
1:53:03in order to access Cloud Code, you do
1:53:04have to be on a paid plan of Claude. As
1:53:06you can see, if you're on the 17 bucks a
1:53:07month plan with Pro, you get Claude
1:53:10Code. Um, but you will probably find
1:53:12pretty quick that you'll want to upgrade
1:53:13to Max or the the higher version of Max
1:53:16because you'll be doing a lot of
1:53:17automations in there and you don't want
1:53:18to hit your limit and then have to
1:53:20upgrade. But you could always start on
1:53:21pro and then upgrade later. So once we
1:53:23got that extension installed, I'm just
1:53:24going to go ahead and click on this
1:53:25button in the top right which looks like
1:53:27the Enthropic logo and I'm just going to
1:53:29open up Claude Code. I'm going to close
1:53:30out of this window and now you can see
1:53:33that we have basically a chat GBT like
1:53:35looking interface where we have Claude
1:53:37code right here. So on the lefth hand
1:53:39side, instead of looking at the
1:53:40extensions marketplace, we're going to
1:53:41click on this button up at the top that
1:53:43says explorer. And what it tells us
1:53:44right here is that you have not yet
1:53:46opened a folder. So it prompts you to
1:53:47open a folder. So before we go ahead and
1:53:49open one up, let's talk about why and
1:53:51what we're looking at. So this is kind
1:53:53of the environment that we're looking at
1:53:55right now. We've got our files on the
1:53:56lefth hand side and this is where we're
1:53:57going to actually build our project, our
1:53:59system prompts, our workflows, our
1:54:01tools. And then on the right hand side,
1:54:03we have the agent. So this is where we
1:54:05talk to Claude Code. We have it help us
1:54:07with a plan. It asks us questions and
1:54:09then it actually executes on those
1:54:10actions. So lefth hand side is files,
1:54:12right hand side is the agent. It's going
1:54:14to be super simple and I'm going to show
1:54:16you how we can keep our file structure
1:54:17really clean so it doesn't get
1:54:19overwhelming and confusing on this lefth
1:54:20hand side over here. So whenever you're
1:54:22in cloud code, you have to be working
1:54:24inside a project and that's why it
1:54:25prompts you to open up a folder. So what
1:54:27I'm going to do is in my documents, I've
1:54:29got a folder called agentic workflows
1:54:30and I've got a bunch of ones that I've
1:54:32been playing around and testing with.
1:54:33But I'm just going to go ahead and open
1:54:34up a new blank folder for today's video.
1:54:36I'm going to go ahead and call this one
1:54:38YouTube
1:54:39analysis. And then I've created that
1:54:41folder. So now when I go back into cloud
1:54:44code, I'm just going to open up that
1:54:45folder. Cool. So I just opened it up and
1:54:47it changed what we were looking at over
1:54:48here. On the right hand side, we've got
1:54:50like VS Codes agent. So I'm not going to
1:54:52worry about that and just close out of
1:54:53that. And then on the lefth hand side,
1:54:55you can see we're now in the YouTube
1:54:56analysis folder, but there's nothing in
1:54:58there yet. So once again, I'm just going
1:55:00to reopen Cloud Code. close out of this
1:55:02one. You can see you can have multiple
1:55:04different files open on the right hand
1:55:05side. So if you wanted to have like five
1:55:07cloud code agents running or you wanted
1:55:09to look at five different files or
1:55:11system prompts, you could do so. But
1:55:12right now we're just going to keep it
1:55:13open to one. So the first thing that we
1:55:15need to do is we need to give cloud code
1:55:17a system prompt for this project. And
1:55:19that's the first thing that you should
1:55:20do whenever you open up a new project in
1:55:22cloud code. And we call this system
1:55:23prompt a claude.md file. MD just
1:55:26standing for markdown. So I'll show you
1:55:27guys that in a sec. But without a system
1:55:29prompt, it's like we have an NN AI agent
1:55:31like an expert copywriter and we don't
1:55:33actually give it a system prompt in
1:55:35here. So without a system prompt, it
1:55:36wouldn't actually really be an expert
1:55:38copywriter. It would be super generic.
1:55:40It wouldn't understand the tools it has,
1:55:41the product that we're trying to sell or
1:55:43where the documents live and what those
1:55:45look like. So that leads me into the
1:55:47next part of the video, which is talking
1:55:48about the framework, which is how we
1:55:50actually build these automations. So
1:55:52here's a really, really simple
1:55:53visualization of what we're actually
1:55:55doing here. We've got our agent which is
1:55:57claude code and the agent is going to
1:55:59help us build workflows. Workflows
1:56:01meaning processes, SOPs, instructions of
1:56:04what we actually want to do. And inside
1:56:06those workflows, we're going to give it
1:56:07access to tools. And tools means
1:56:09actually executing actions. So send
1:56:11email would be a tool. Research a
1:56:13YouTube channel would be a tool. So it's
1:56:15really similar to the way that we have
1:56:16workflows and tools in Nitn. Here you
1:56:18can see is an edit in workflow for a
1:56:20daily news summary. And inside the
1:56:22workflow, which is a specific set of
1:56:23instructions in a specific order. So,
1:56:25it's a deterministic process. We have
1:56:27different tools. We've got a tool here
1:56:29for Tavali to do research. We've got a
1:56:30tool here for an AI agent to do the
1:56:32newsletter writing. And we've got a tool
1:56:34at the end to send a Gmail message. So,
1:56:37hopefully that all makes sense. It's
1:56:38going to be really simple. We're going
1:56:39to have a folder for workflows. And in
1:56:41there will be all of our processes.
1:56:42We're going to have a folder for tools
1:56:44and in there will be all of the actual
1:56:45things that it can execute. And then the
1:56:47agent basically helps us set up those
1:56:48tool files and workflow files and then
1:56:50execute those actions. So, I'm going to
1:56:52do is drag in this claude file. And you
1:56:54can see it's a claude.md. This could be
1:56:56called agents.mmd, gemini.mmd, whatever
1:56:59you want. In this case, we're using
1:57:00claude code, so I'm calling it
1:57:01claude.md. But let me go ahead and
1:57:04expand this one and let's briefly read
1:57:06through it so you understand exactly
1:57:08what I just talked about with the
1:57:09workflows, agents, and tools. So this is
1:57:11the agent instructions for this specific
1:57:13project. You're working inside of the
1:57:15WAT framework, which stands for
1:57:17workflows, agents, tools. This is a
1:57:18three-layer framework and it basically
1:57:20separates concerns so that the
1:57:21probabistic AI handles reasoning while
1:57:23deterministic code actually handles the
1:57:26execution and that is what makes these
1:57:27systems actually reliable. So like I
1:57:29said layer 1 is the workflows the
1:57:31instructions. So these are markdown SOPs
1:57:34stored in the workflows folder which
1:57:36will be created in a sec. Each workflow
1:57:37defines the objective, the required
1:57:39inputs, which tools to use, expected
1:57:41outputs and how to handle edge cases.
1:57:44It's written in completely plain
1:57:45language the same way that you'd brief
1:57:46someone on your team. And by the way,
1:57:48when I say markdown, it basically just
1:57:50means this structure. This is a markdown
1:57:52file right here where we have like
1:57:53headers and subheaders and bold font and
1:57:55things like that. Layer two is the
1:57:57agent. So this is the actual cloud code
1:57:59agent that we talk to. This is your
1:58:01role. You're responsible for the
1:58:03coordination between workflows and
1:58:04tools. You read the relevant workflow.
1:58:06You run tools in the correct sequence
1:58:08and you handle failures. You ask
1:58:09clarification questions when needed.
1:58:11Layer three, we have the tools, and
1:58:13these are actually going to be Python
1:58:14files. So right here, you can see claude
1:58:16is a markdown file. So it's cloud.md. We
1:58:18said that our workflows were going to be
1:58:20markdown files. So it will be like um
1:58:22scrape website.mmd.
1:58:24But then in the tools which we will have
1:58:25another folder for over here. We're
1:58:27going to have tools that are going to be
1:58:29py. So a python file. So in this case we
1:58:31can see there's an example tool called
1:58:33scrape single site. py which would be a
1:58:36python script that would execute an
1:58:37action. These can be API calls, data
1:58:39transformations, file operations,
1:58:41database queries. And a lot of times in
1:58:43these tools, we'll need an API key, but
1:58:45we're not going to actually store them
1:58:46in the tool code logic itself because if
1:58:49that got exported or we pushed that onto
1:58:51the web, then our API keys would be
1:58:53exposed. So, we're going to handle
1:58:55secrets by storing them inv files. You
1:58:58don't have to understand exactly what
1:58:59that means or how that works right now.
1:59:00We'll show you. So, then we talk a
1:59:02little bit about like why this matters,
1:59:03how to operate. So, you look for tools
1:59:05first. You learn and adapt when things
1:59:07fail because these agentic workflows are
1:59:09basically self-healing. So, as we're
1:59:10going through and building this
1:59:12workflow, you will see that it says,
1:59:13"Okay, I ran into an error here. Let me
1:59:15figure out what happened and let me fix
1:59:16it." So, fix the script and retest
1:59:18document what you learned. So, if it ran
1:59:20into an error and it fixed it, it will
1:59:22go ahead and change the workflow file so
1:59:23it doesn't run into that error again.
1:59:25So, an example could be you get rate
1:59:26limited on an API, you dig into the doc,
1:59:28so you do research, you discover a batch
1:59:31endpoint, you refactor the tool to use
1:59:33it, you verify that that works, and then
1:59:35you update the workflow so it never
1:59:36happens again. This is once again where
1:59:38we talk about that self-improvement
1:59:39loop. We talk about the file structure
1:59:40and you can see that it's going to
1:59:42create this for us. And basically the
1:59:44bottom line is that you sit between what
1:59:45I want, which are workflows, and what
1:59:47actually gets done, which are the tools.
1:59:49Your job is to read instructions, make
1:59:50smart decisions, call the right tools,
1:59:52and keep improving the system as you go.
1:59:54So, I know we skimmed through this kind
1:59:55of fast, but you guys will get access to
1:59:57this exact same system prompt. I'll
1:59:59leave it in my free school community.
2:00:00The link for that will be down in the
2:00:01description. That way, you can just go
2:00:02ahead and grab this, paste it in, and
2:00:04then when you want to follow along and
2:00:05build some workflows in Cloud Code,
2:00:07you've got this right here for you. So
2:00:08now what we need to do is just set up
2:00:10our environment with the different
2:00:11folders. So I'm going to talk to cloud
2:00:12code and just say initialize this
2:00:14project based on the claw.md file. So
2:00:17I'll go ahead and shoot that off. And
2:00:18when we talk to claude, what it does is
2:00:20it basically just tells us exactly what
2:00:21it's doing and what it's thinking. What
2:00:23you'll notice right here is that I'm on
2:00:25a mode called bypass permissions. And
2:00:27you might not see this initially. You'll
2:00:29see ask before edits, edit
2:00:30automatically, and plan mode. But it is
2:00:32really helpful to be able to turn on
2:00:33bypass permissions. So the way that you
2:00:35do that is you go to the bottom left to
2:00:36settings. You're going to go to settings
2:00:38once again. You'll type in cloud code
2:00:40and then you're just going to turn on
2:00:42this option that says allow bypass
2:00:43permissions mode. And that's what allows
2:00:45you to do that so that you can let your
2:00:46agent run and you don't have to approve
2:00:48every step. Now, as this is running,
2:00:49what you'll notice is on the lefth hand
2:00:51side, we're seeing some files and
2:00:52folders pop up. So, we've got a
2:00:54temporary folder, which just means
2:00:55anything that it needs to store and then
2:00:57like delete later just temporarily, it
2:00:58can do so in there just to keep
2:01:00everything clean. We've got our tools
2:01:02folder, we've got our workflows folder,
2:01:03and then we have av and getit ignore. So
2:01:06this is going to help us just basically
2:01:07keep our project clean, but also the
2:01:10agent knows exactly where everything is.
2:01:12Cool. So the project is now initialized
2:01:14using our WAT framework and it showed us
2:01:17what it created. So now let's move on to
2:01:19section three of the video where we're
2:01:20going to be talking about planning and
2:01:21communicating with our agent. So what
2:01:24I'm going to do is I'm going to clear
2:01:26out this conversation. If I wanted to
2:01:27access past conversations, I could do so
2:01:29up here. I'm going to go to plan mode.
2:01:31And this is really important. Whenever
2:01:33you're doing something that actually
2:01:35involves like creating something, you
2:01:37need to describe the goal and you need
2:01:39to be able to describe it super super
2:01:40clearly. And it's not just the goal, you
2:01:42need to also describe the features that
2:01:44you want. And if you were to just
2:01:45describe something and then chuck clawed
2:01:47code at it and you would do bypass
2:01:48permissions, you probably wouldn't get a
2:01:50great output. So what you always want to
2:01:52do when you're creating an idea is you
2:01:54want to go on plan mode. Because what
2:01:56you're going to see is when I'm on plan
2:01:57mode, it thinks extra hard and it looks
2:01:59at everything in the folder and it's
2:02:01going to ask me tons of questions that I
2:02:03might not have thought of, which is
2:02:04really, really helpful because it gets a
2:02:06really, really good understanding of
2:02:07what we want and it brainstorms options
2:02:09and then it actually will do it after
2:02:10it's confident. So, let's explain the
2:02:12workflow that we want to build today.
2:02:14Hey Claude, I need your help building an
2:02:16automation. I want this automation to
2:02:18basically scrape tons of YouTube videos
2:02:21and YouTube channels in my niche, which
2:02:23is AI and AI automation. I want to get
2:02:26insights about what videos are trending,
2:02:28what's working well, and kind of what
2:02:30the AI space is feeling like so that I
2:02:32can create more content that people want
2:02:34to see and that will be beneficial for
2:02:35them. I need your help understanding how
2:02:37we can actually get this data. So, look
2:02:39into different APIs or MCP servers.
2:02:41Also, let me know if there's any skills
2:02:43that would be helpful because after
2:02:45you've done this research, what I want
2:02:46you to do is I want you to create a
2:02:48slide deck for me. So, I want to get an
2:02:50actual deliverable that will be sent to
2:02:52my email using Gmail and it should be a
2:02:54really nice professionallook slide deck
2:02:57with charts and images and all of these
2:02:59different graphics so that I can
2:03:00understand what's going on in the
2:03:02industry. So, that's what I've got. Let
2:03:04me know if you have any questions or if
2:03:06you have any recommendations for things
2:03:07that I haven't thought of about this
2:03:09automation system. Cool. So that was my
2:03:11little brain dump and it's going to come
2:03:12back and ask me a ton of questions which
2:03:15is just going to help make this project
2:03:16a lot lot better. And so I know a lot of
2:03:18you guys might be looking at this and it
2:03:20seems overwhelming and confusing and I
2:03:22agree like when I first wanted to dive
2:03:24into cloud code I watched some YouTube
2:03:25videos and I just it didn't click. The
2:03:27only way it's truly going to click is if
2:03:29you get in here and you do it yourself
2:03:30because once you send off these messages
2:03:33just read everything it's doing. Read
2:03:34every single line and you'll start to
2:03:36understand the way that these models
2:03:37think and what they try to do. And
2:03:39that's truly the best way. So after this
2:03:41video, restart it from the beginning,
2:03:43open up Cloud Code, and just kind of
2:03:44follow along with what I'm doing and it
2:03:46will all start to click. I promise. And
2:03:47by the way, you can see that as it's
2:03:49making this plan for us, it's doing
2:03:50research. So it's not just thinking,
2:03:52it's also searching the web to find out
2:03:54how we can scrape the YouTube analytics
2:03:56and how we can use MCP servers and
2:03:58things like that. Okay, so we got some
2:03:59questions now from Claude. It says,
2:04:01"What specific YouTube channels do you
2:04:02want to track? Should I discover top AI
2:04:04automation channels automatically or do
2:04:06you have a list? Let's just go with
2:04:07autodiscocover top channels. Frequency
2:04:09is how often should this report be sent?
2:04:11I'm going to go ahead and do weekly.
2:04:12Then it asks us if we want to track all
2:04:14this data in sheets. Yes, absolutely.
2:04:16Let's do that. And then for delivery, it
2:04:17says what email address should the
2:04:19reports be sent to? And I'm going to go
2:04:21ahead and say send to my Gmail. So, I
2:04:23shut off those answers and now it's
2:04:24going to keep updating the plan. All
2:04:25right. So, the plan is finished. The
2:04:27objective is to build an automated
2:04:28system that scrapes YouTube data for the
2:04:30AI niche. It analyzes trends and gets
2:04:33performance metrics and then generates a
2:04:34professional slide deck with charts and
2:04:36visualizations and sends that to me over
2:04:38Gmail. We've got the workflow which is
2:04:40YouTube weekly report. We've got the
2:04:41agent layer. We've got different tools.
2:04:43It's going to build out these seven
2:04:44different Python tools that it
2:04:46mentioned. So fetching YouTube data,
2:04:48analyzing YouTube data, generating
2:04:49charts, generating slides, sending the
2:04:52email report, exporting to sheets, and
2:04:54discovering channels. And now it needs
2:04:56to actually create this workflow. So, we
2:04:58could obviously read through all of this
2:04:59and we could give it some feedback if we
2:05:00wanted to, but I'm just going to go
2:05:02ahead and accept these because I want to
2:05:03see how well it did with just one
2:05:05iteration of our plan, which took me a
2:05:07few minutes. So, you can see what it
2:05:08does is it starts a to-do list. So, it's
2:05:10basically just going to knock off one of
2:05:11these at a time. And that's really nice
2:05:12because it helps the agent stay on
2:05:14track, but it also means that you could
2:05:15go to your other monitor here and work
2:05:17on something else and just kind of keep
2:05:19peeking in on it and checking on the
2:05:20to-do list to see how much is left to
2:05:22run. Okay, so the to-do list is done.
2:05:24The workflows and tools have been built.
2:05:26So, here's where we're at. We've got our
2:05:28seven tools have been created. So, if I
2:05:29open up the tools folder, we should see
2:05:31we now have these seven Python files.
2:05:33And each of these, like I said, are
2:05:35actual Python code that will execute
2:05:37some sort of action. So, those have been
2:05:39built. We've also got the workflow. So,
2:05:41this is our markdown file, YouTube
2:05:43weekly report, which is an actual
2:05:45process. So, I'm not going to read this
2:05:46whole thing, but it has the actual steps
2:05:48that we would be doing here. So, now it
2:05:50says to get started, we have a few
2:05:51dependencies. So, the first one is we
2:05:53need to install something. The second
2:05:54one is to add a YouTube API key. The
2:05:56third one is to set up Google OOTH for
2:05:58Gmail and Sheets. And then the fourth
2:06:00one is just to run the actual workflow.
2:06:02So a lot of times when Cloud Code's done
2:06:03and it has some action items, it
2:06:05actually just tells you to do some stuff
2:06:06that it could do itself. So right now we
2:06:08would obviously have to go get our
2:06:09YouTube API key and then we could just
2:06:11give it to it and say, "Hey, you go
2:06:13update the ENV. I don't want to touch
2:06:14that. You just go do it." But first,
2:06:16what it's doing is it's asking us to do
2:06:18this. So, we could obviously just
2:06:20install this right now, or I could just
2:06:22say, "Can you please go ahead and
2:06:24install the dependencies? I'll go grab
2:06:25my YouTube API key." Cool. So, it went
2:06:28ahead and installed that stuff just like
2:06:29I told it to. And now it's asking for a
2:06:31YouTube API key. So, instead of just
2:06:33adding it to the file, I'm just going to
2:06:34drop it in right here. And then the one
2:06:36thing I will have to go do manually is
2:06:37step three. So, I'll have to enable the
2:06:39YouTube data API and Gmail and Google
2:06:41Sheets and then create the credentials
2:06:43and just drag in the JSON file, which I
2:06:45will do that in a sec. And here's
2:06:46another thing I'm doing with my API key.
2:06:48It should only be added to the ENV file.
2:06:50It shouldn't be listed in the workflows
2:06:52or the tools. Okay, so I added
2:06:54everything that I needed to. And if
2:06:55you're confused about how to do that,
2:06:56just say, "Hey, where do I go? What do I
2:06:58click on? How do I do that?" And it'll
2:06:59walk you through. And now what it's
2:07:01doing is because it has all our
2:07:02credentials, it's actually just testing
2:07:03out if the things work. So you can see
2:07:05the YouTube API is working now. Let's
2:07:07run the full data collection pipeline.
2:07:09So it's basically just testing that the
2:07:10flow works and then we'll give it a full
2:07:12run. But we can see that it just ran the
2:07:14full pipeline. So that was our first
2:07:15initial test. It found 30 channels. It
2:07:18fetched 187 videos. It generated
2:07:20analysis. It made six charts. It built a
2:07:22nine slide PowerPoint deck for us.
2:07:25Exported it to Sheets. And then it
2:07:26emailed us the report. So let's go take
2:07:28a look at all that. Okay. So here's the
2:07:29email that I got. AI automation YouTube
2:07:31analytics. So the weekly report for Jan
2:07:3420. We got 30 channels tracked, 187
2:07:36videos. We have some top videos from the
2:07:38week. We have recommendations. And then
2:07:40we also have our PowerPoint right here,
2:07:42which we can see. We have similar
2:07:44information. We've got median views,
2:07:45median engagement, trending topics.
2:07:47We've got top performing videos. So, we
2:07:49have this laid out by title and by
2:07:51views. We've got top channels by
2:07:52subscribers. Unfortunately, I do not see
2:07:54my name up there, so please hit the
2:07:56subscribe button. We've got engagement
2:07:58analytics. We've got trending topics, by
2:08:00keywords in the Aming patterns, and then
2:08:03we have some recommendations to kind of
2:08:04close us off here. So, keep in mind this
2:08:06is not perfect, and we obviously would
2:08:08want to come back and make this a little
2:08:09bit more tailored for us, but this was
2:08:12one prompt. Cloud code asks us questions
2:08:14and then I basically just sat down and
2:08:16then I came back over here when it was
2:08:18done and this is what we have ready for
2:08:20us. What we also see is that we got this
2:08:22exported to a Google sheet. So if I
2:08:24click on this, remember that we didn't
2:08:26create the sheet. We didn't create these
2:08:27different tabs or the actual like schema
2:08:30of this. But we've got three tabs. The
2:08:32first one is channel stats. So this
2:08:33pulled channel stats from today's date
2:08:35which is January 20th. We have the
2:08:36channel IDs. We have the actual channel
2:08:38names. And then we've got subscribers,
2:08:40total views, and video count. We can see
2:08:42nice that Nate Herk AI automation did
2:08:45make it in this scrape. We've also got
2:08:47top videos. So once again, this was ran
2:08:49based on today's analytics. We got the
2:08:50video ID. We've got the title of the
2:08:52videos. We've got the channel, the
2:08:54views, the likes, the comments, the
2:08:55engagement rate, which is pretty cool.
2:08:57And also how old the videos are. So we
2:08:59can see that we're getting real accurate
2:09:01like what's trending right now. And then
2:09:02we get a weekly summary. So this is
2:09:04supposed to run every single week. We
2:09:06can see the day that it ran, the
2:09:07channels it tracked, the videos it
2:09:08analyzed, the median views, the median
2:09:10engagement score, and the top keyword in
2:09:13top keyword 2, which actually, it's
2:09:14funnily enough, spells out claude code,
2:09:17which is why you're seeing this video
2:09:18right now. Okay, so let's recap what
2:09:20we've done. We have familiarized with
2:09:22the interface. We have built out the
2:09:24actual structure of our project using a
2:09:26cloud.md file, which is like a system
2:09:28prompt. Now, we have our workflows. We
2:09:30have our tools and we have actually gone
2:09:33through the whole planning stage with
2:09:34claude code to build out the initial you
2:09:37know workflow automation that we need.
2:09:38So what comes next now is we want to
2:09:40talk about a few other things. We want
2:09:42to talk about superpowers. So MCPS and
2:09:44skills and then we're going to test it a
2:09:45little bit more and then we're going to
2:09:46actually deploy the automation live. So
2:09:49to start off with superpowers MCP
2:09:52servers. So I'm not going to dive super
2:09:53super deep into MCP servers in this
2:09:55video but I did want to bring it up. So,
2:09:57if you remember in plan mode, I
2:09:58basically said, "Hey, I want to scrape
2:09:59YouTube data. Can you just go figure out
2:10:01if I should use an MCP server or like an
2:10:03API?" And it ended up finding out that
2:10:05the YouTube API was going to work
2:10:07better. So, that's why we did it in this
2:10:09workflow. But essentially, just think of
2:10:10an MCP server as an app store. So, Gmail
2:10:14has an MCP server, Calendar has an MCP
2:10:16server, lots of these services do. And
2:10:18this is like one of the most common
2:10:20visualizations because it's like a
2:10:21universal micro USB port because instead
2:10:25of having to go to calendar's API and
2:10:28have one different API request to create
2:10:30an event, one different one to update
2:10:31event, one different one to delete an
2:10:33event, all we have to do is connect once
2:10:34to the whole server and then the agent
2:10:37can figure out how to go use different
2:10:38endpoints and parameters. It just
2:10:40simplifies the whole process. Now, what
2:10:41I did want to talk about a little bit
2:10:42more was the idea of claude skills
2:10:44because this is a little bit newer. So
2:10:46essentially skills are instructions or
2:10:48resources that Claude can load in
2:10:50dynamically. And that's kind of the key
2:10:51piece here is that instead of just
2:10:53reading it every time in a system
2:10:54prompt, it basically understands what is
2:10:56the request. Let me go look at all the
2:10:58skills I have access to. If one of them
2:11:00is relevant, I'll pick that one. I'll
2:11:02read it all and then I'll take action.
2:11:04And this process basically just improves
2:11:06Claude's consistency, speed, and
2:11:07performance. And also saves you tokens.
2:11:10Like I said, when you ask Claude to do
2:11:11something, it reviews the available
2:11:12skills. it loads in only the relevant
2:11:14ones and then it applies those
2:11:15instructions. So, we're going to go
2:11:17ahead and try to implement a skill into
2:11:18this workflow and I'll actually show you
2:11:20what the skill document entails. So,
2:11:22then it will all start to make a little
2:11:23bit more sense. But before we do that, I
2:11:25did want to real quick cover the
2:11:26difference between skills and projects
2:11:28and skills and MCP. So, the first one is
2:11:30about projects. You're in a project and
2:11:32basically what we have is access to
2:11:34whatever is in here. So, it's kind of
2:11:36static documents and background
2:11:37information. And a lot of times these
2:11:39skills are installed globally. So what
2:11:41you'll notice actually in our project is
2:11:42that we don't have any skills in this
2:11:44project. Normally there will be like a
2:11:46thing that will be like agents and then
2:11:47you drill down in that folder and you'll
2:11:49see like agent skills or claude skills.
2:11:51And that's more installed on the global
2:11:53level. And that's actually really good
2:11:54because what that means is if I closed
2:11:56out of this project and I opened up a
2:11:57different one, I would still have access
2:11:59to all the same skills that I've already
2:12:00installed. So you can see right here
2:12:02that I just asked Claude Code, "What
2:12:04skills do you have?" and it came back
2:12:05and showed that it has a front-end
2:12:06design. It has naden skills and those
2:12:08are the only eight that it actually has
2:12:10even though we don't see them in this
2:12:12specific project. Now we have skills
2:12:14versus MCP and these are also very
2:12:16different. MCP is basically to get data
2:12:18and take action. So like I said if you
2:12:21want to connect Claude to something like
2:12:23Gmail to read emails or to send emails
2:12:25but skills are more like knowledge
2:12:27custom instructions. So, if you ever
2:12:29find yourself constantly repeating
2:12:30something to your cloud code agent, then
2:12:32maybe that's a good sign to put it
2:12:34either in the claw.md file or create
2:12:36your own custom skill for it. So, like
2:12:38the example of the front-end design, if
2:12:39you wanted to use cloud code to build
2:12:40yourself a landing page or a website,
2:12:42using the front-end design significantly
2:12:44improves its ability to actually design
2:12:46things. So, what we're going to be doing
2:12:48in this example now is I want to use a
2:12:51skill and I'm going to be looking at
2:12:52this cloud code templates website which
2:12:54has a bunch of agents and commands and
2:12:56MCP servers and skills and hooks and I'm
2:12:58going to be looking for one that helps
2:13:00us create like better looking PDFs. I'll
2:13:02also leave a link to this in the
2:13:04description of the video. So, I went
2:13:05ahead and searched for design and you
2:13:06can see there's a skill right here
2:13:07called canvas design. And if I view
2:13:09details here, it says create beautiful
2:13:11visual art inputs
2:13:14using design philosophy. So, we're going
2:13:16to go ahead and try this one out. I've
2:13:18never used it before. We'll see how it
2:13:19works. But this is actually like the
2:13:21code of the skill itself. And you can
2:13:23see it basically is just natural
2:13:25language instructions. So, it's just a
2:13:26custom prompt that someone built or you
2:13:28built yourself. And now I can load this
2:13:30into cloud code. So, when we have it
2:13:32design a PDF, it can use this and it
2:13:34will probably just come out a lot better
2:13:35because it's prompted. So, we've got
2:13:37installation right here where we can use
2:13:38this code. So, what I would try is just
2:13:40copying this, going into VS Code. I'm
2:13:42going to go ahead and open up a, you
2:13:44know, kind of clear the conversation and
2:13:46just paste that in and see what happens
2:13:47if I drop that in there. Okay, so I
2:13:49dropped it in and then it actually ran
2:13:50the command in our terminal to install
2:13:52it. And it says that it's been installed
2:13:53and we have skill.md for the
2:13:56instructions for the skill. And then
2:13:57we've also got a bunch of fonts. And
2:13:58what it did is it actually created a new
2:14:00folder here called.claude. And then we
2:14:02do have skills right here. So you can
2:14:04see that it put it in this project. So
2:14:06now I'm a little confused because I
2:14:07don't know. Okay, we have a skill here,
2:14:08but we also have skills globally. So I
2:14:10would literally just say it looks like
2:14:12you created this skill in this project.
2:14:14Is this going to be installed globally
2:14:16or will it only be accessible through
2:14:17this project? So right now it basically
2:14:19says yeah this was installed just
2:14:21locally in this project and that's fine.
2:14:22And if you wanted it to be global
2:14:24instead you would just say okay actually
2:14:26just make that global and then it would.
2:14:27So anyways going to clear out this
2:14:29conversation one more time. I'm going to
2:14:30go back into plan mode and I'm going to
2:14:32give it a prompt. And actually one more
2:14:33thing before I prompt it. I'm going to
2:14:35drag in the AI Automation Society Plus
2:14:37logo just over here on the lefth hand
2:14:39side. And you can see it's right here
2:14:41and the file pops up. Right. So, what
2:14:43I'm going to do is prompt it, but I want
2:14:44it to actually have this logo on all of
2:14:47the PDFs that it generates. Hey Claude,
2:14:49so I just gave you a skill for canvas
2:14:51design. And instead of outputting a
2:14:53PowerPoint presentation, I want you to
2:14:55now take the same research when you do
2:14:57your analysis from YouTube videos, but I
2:14:59want you to use that canvas design skill
2:15:02to create a PDF. It needs to be
2:15:04professional, but it needs to be
2:15:05aesthetically pleasing. And what I want
2:15:07you to do is make sure you're including
2:15:08the AIS Plus logo PNG that I dropped in
2:15:12this folder as well because I want the
2:15:14whole presentation to be branded so I
2:15:15can share it with my team. So, I'm
2:15:17shooting this off in plan mode and I'll
2:15:18let you know when it comes back with
2:15:20some questions. Interesting. So, it came
2:15:22back and said that that canvas design
2:15:23skill that we just installed creates
2:15:25PDFs interactively, which means step
2:15:27five of our workflow changes from fully
2:15:28automated to semi-automated. So, how do
2:15:30we want to handle this? Let's just go
2:15:32ahead and just say keep it fully
2:15:33automated because that's kind of the
2:15:34whole point. We want to be able to push
2:15:35this live to run on a schedule trigger.
2:15:38Okay. So, the new plan is to replace the
2:15:39PowerPoint output with a branded PDF
2:15:41report. So, it's going to make a new
2:15:42tool to replace the generate slides
2:15:44tool. We have our current workflow
2:15:46state. We've got our logo. It has some
2:15:47proposed changes here. We're going to be
2:15:49looking at the PDF structure. And of
2:15:51course, what it has to do is update the
2:15:52actual workflow. So, it's going to look
2:15:54at this YouTube weekly report markdown
2:15:56file, which is the actual workflow. Of
2:15:57course, it's going to change that. It's
2:15:59going to update some of the other tools
2:16:00like the email tool. And then, of
2:16:01course, it's got some other
2:16:02implementation steps for us. And in this
2:16:04case, what I'm going to go ahead and do
2:16:05is just autoaccept these changes. And
2:16:07so, right now, it's just setting up a
2:16:08to-do list to actually implement those
2:16:10changes. We're not going to be running
2:16:11the workflow again. We're just going to
2:16:13make the changes, and then we'll go
2:16:14ahead and test it. And just a reminder,
2:16:16when you guys are in here building your
2:16:17own workflows, just pay attention to
2:16:19what it's actually doing. It does some
2:16:20really interesting things. Like right
2:16:21here, it installed some dependencies to
2:16:23actually be able to create the PDF a
2:16:25little bit better. And then here it says
2:16:26the PDF was generated, but it's using a
2:16:28fallback using whatever this is, and it
2:16:30would look better if it had proper title
2:16:31and closing pages. So, it's going to
2:16:33install something else and then try it
2:16:34again. It's just a reminder of using
2:16:36this framework of an agent that sits
2:16:37between workflows and tools. As it's
2:16:39building them out, as it's testing them,
2:16:41it's continuously improving them, seeing
2:16:43errors, seeing things that could be
2:16:44improved, and then just going ahead and
2:16:46doing that for you. So, that's where
2:16:47it's really powerful. And this testing
2:16:48and optimization phase is really
2:16:50important because once you actually
2:16:51deploy your automation, you're not
2:16:53deploying the agent. You're just
2:16:54basically deploying the workflow that's
2:16:56connected to tools. And that's important
2:16:58to understand. The workflow itself would
2:16:59be put up into the cloud where it could
2:17:01run on a schedule trigger, but the agent
2:17:03still lives locally in cloud code. Which
2:17:05means if a workflow which means if your
2:17:07workflow is running every week, it's not
2:17:08going to be self improving and
2:17:10self-healing. So if you wanted to do
2:17:11that, you would come over to cloud code,
2:17:13you'd edit the workflow, you'd improve
2:17:15it, and then you just push that version
2:17:17back to modal or wherever you're hosting
2:17:18them. But anyways, this finished up, so
2:17:20it created a new tool. It modified a few
2:17:23other things. It changed the actual
2:17:25workflow itself. And then what also it
2:17:27did is it made a test PDF just to see
2:17:29how that worked. And you can see here
2:17:31it's stored as a temporary file. So in
2:17:33our temp folder, which is right here,
2:17:34you can see right there we have a
2:17:36YouTube report PDF. And let me just make
2:17:38this bigger. We've got our logo right
2:17:40here. We've got our AI and automation
2:17:42YouTube analytics report and we have the
2:17:44thank you slide. So it basically just
2:17:46tested to see if it worked. But now
2:17:47we're going to go ahead and run that
2:17:48full workflow and then we're going to
2:17:49see if we're ready to actually push it
2:17:51up into production. So I'm on bypass
2:17:54permissions and I'm just going to shoot
2:17:55off run the YouTube analysis workflow
2:17:58and it's not even called that, but it
2:17:59will be able to search through the
2:18:00workflows that it has and it's going to
2:18:02understand which one to run. It's going
2:18:03to execute all of those Python scripts
2:18:06in order and then we should have a
2:18:08finished product. Okay, so here's the
2:18:10email. It has the similar structure as
2:18:12far as the actual body of the email, but
2:18:13then at the bottom, we should have our
2:18:14PDF, which we got attached right here.
2:18:17But what you'll notice is that it's only
2:18:18two pages. So, it didn't actually create
2:18:20the right type of PDF that we were
2:18:21looking for. However, it did update the
2:18:23Google sheet. So, it added, you know,
2:18:24those 30 more videos that we originally
2:18:27didn't have on this sheet. It added more
2:18:30videos, of course, and then it threw in
2:18:31one more weekly summary where it has a
2:18:33little bit of a different metrics. And
2:18:35what's interesting is that you can see
2:18:36that it did generate charts and it did
2:18:38do analysis because it actually
2:18:40generated all of these images over here,
2:18:42top channels, top videos, key
2:18:44performance indicators, posting
2:18:45patterns, all this kind of stuff. It
2:18:47just didn't actually include it. So once
2:18:48again, we would go back in natural
2:18:50language and say, "Hey, you know, we
2:18:51just got that PDF, but it was only two
2:18:54slides." So what I did is I said
2:18:56everything seemed to work except for the
2:18:57PDF that I received was only two slides.
2:18:59It was only the title and the thank you
2:19:01page. So, it found the issue. It fixed
2:19:03it. It changed the workflow. It changed
2:19:05the tools. And now, it's shot me off a
2:19:07new example with nine pages. And this
2:19:09time, we still have the logo. We still
2:19:10have the date. And we also now have all
2:19:12of the actual slides that we need in
2:19:14this PDF with the charts and things like
2:19:16that, recommendations, and then we still
2:19:17have the closing off slide. So,
2:19:19hopefully you guys understand now how
2:19:21important the planning really is because
2:19:23we did kind of rush through this in this
2:19:25example where we auto accepted changes
2:19:27and we just kind of like sped through
2:19:28things. And it's fine because we're
2:19:30still able to go back and forth and let
2:19:31Claude Code investigate and fix, but it
2:19:34should show the importance of if you are
2:19:36really really clear up front and you
2:19:38know exactly what you need, it will be a
2:19:40lot better off the jump, but it's not
2:19:42perfect. Okay, so now let's say we're at
2:19:43the spot where we're ready to basically
2:19:45make this workflow live where we
2:19:46actually want to forget about it and
2:19:48just let it run every Monday at 6 a.m.
2:19:49or whatever. So, we need to deploy it.
2:19:51So, the way that we're going to do that
2:19:52is we're going to use modal, which is AI
2:19:55infrastructure that developers love.
2:19:56Essentially, what modal is is it lets
2:19:58you spin up these kind of like computers
2:19:59in the cloud where you can host your
2:20:01automations and it only charges you when
2:20:03they actually run. So, you're not
2:20:04getting charged by the minute or by the
2:20:06day. You're only getting charged every
2:20:07time they actually execute. So, when you
2:20:09create an account, you'll get five bucks
2:20:10for free. And then if you add a credit
2:20:12card, even though it won't charge you
2:20:13yet, you'll get 30 bucks. And this 30
2:20:15bucks will last you a long, long time.
2:20:17Trust me. So, what'll happen is this
2:20:19screen will probably pop up and it will
2:20:20say that you need to download and
2:20:22configure the Python client. So you
2:20:23could basically copy this exact command
2:20:26right here and just put that into cloud
2:20:27code or you could just say hey cloud
2:20:29code I want to push this workflow to
2:20:31modal. So just help me get that
2:20:33initialized. But I'll just show you what
2:20:34would happen if you copied this and we
2:20:36came into cloud code and said awesome I
2:20:39want to push the YouTube analytics
2:20:42workflow to modal so that it can
2:20:44actually run every single Monday at 6
2:20:46a.m. And then I'm going to go ahead and
2:20:48paste in those two things that we just
2:20:49saw. And let's actually do this in plan
2:20:51mode first and just shoot that off. So
2:20:52what it's doing is it's going to read
2:20:54through the workflow structure and the
2:20:55tools and understand how it can package
2:20:57everything up so it can actually deploy
2:20:58it on modal as an app. So it came back
2:21:00with a plan to deploy this on modal. But
2:21:02there's one more thing that I want to
2:21:03ask it about before we actually do this.
2:21:05And this last part is security. So I
2:21:07basically told it to run a security
2:21:08review and make sure that my API keys
2:21:10aren't exposed and that there's no
2:21:11vulnerabilities because the reality is
2:21:14we just built a ton of code and I don't
2:21:17know what the code is actually doing.
2:21:19So, it's really important to be thinking
2:21:20about this before you put anything out
2:21:22there on the web. Are any web hooks
2:21:24exposed? And if they are, do you have
2:21:26like, you know, proper protection around
2:21:27that? Are secrets out there? Are API
2:21:29keys out there? What could people do now
2:21:31that this is out there? And as you start
2:21:33to deploy more workflows, whether that's
2:21:35an NEN or whether that's in code like
2:21:37this, you'll start to understand the
2:21:38things to look out for. But you also
2:21:40have one of the smartest reasoning and
2:21:42coding models right here in front of
2:21:43you. So, you might as well just ask it,
2:21:44hey, check the code and let me know if
2:21:47there are any risks. So the security
2:21:49review came back and it found three
2:21:50critical issues that need attention. But
2:21:52the good news is nothing is vulnerable
2:21:55and there's not a GitHub repo. So
2:21:56nothing's been committed out there
2:21:57publicly and everything is going to be
2:21:59stored as a modal secret. So the API
2:22:02keys and the JSON token. So nothing will
2:22:04be committed to any repository. So we're
2:22:06good to go. And basically from there it
2:22:08came back with a plan once more and I
2:22:10have approved it. So it's going ahead
2:22:11right now and it's creating the
2:22:13different tools and the different things
2:22:14that we need to actually be able to
2:22:16write this over to modal and then we'll
2:22:18go ahead and test it out over there. So
2:22:19our deployment is now complete. It had
2:22:21to update the scripts to make sure that
2:22:23they could actually have the right
2:22:24environment variable path. It had to
2:22:26create a modal deployment file. So it
2:22:28actually just understands the process of
2:22:29what it just did and schedule the cron
2:22:31or the schedule trigger. And then it had
2:22:33to create modal secrets that we could
2:22:34store over there. So it is now deployed
2:22:37and scheduled. So if I click on this
2:22:38link, this will bring us to our modal
2:22:41environment right here. And what you can
2:22:42see is that we have two different apps.
2:22:44We have the analytics and then we have
2:22:46the analytics manual. So it had to do a
2:22:48manual run just to see if it worked. So
2:22:50this is the actual app. So if I go back
2:22:52to the main dashboard, you can see that
2:22:53we have this app and there's kind of
2:22:54like the two different like endpoints.
2:22:56But if I open up the app, we can see the
2:22:58overview. We can see deployment history.
2:23:00So as you change something in cloud code
2:23:01and then push it back over here, you'll
2:23:03see a different version. And then you
2:23:04can also see the app logs when it's
2:23:06running. So, when I click into the
2:23:07YouTube analytics one, the one that will
2:23:09be live, it says the next run will be in
2:23:115 days. So, it's scheduled at 6:00 a.m.
2:23:13only on Mondays America Chicago time.
2:23:16But what I'm going to do just to prove
2:23:17to you guys this is working or at least
2:23:18test if it is working is we can actually
2:23:20just go ahead and run one right now. So,
2:23:22I scheduled an immediate run. We're
2:23:24going to see this pop open right here
2:23:25and we're going to see the fact that
2:23:26it's running right now. As you can see,
2:23:28it took 2 seconds to start up and now
2:23:30it's running. And then we'll see the
2:23:31result of that execution. And actually,
2:23:33I'm glad that this just failed cuz I can
2:23:34show you what you need to do. But this
2:23:36failed, right? So we'll click into this.
2:23:38And when you click into each of the
2:23:39runs, you'll basically be able to see
2:23:41the logs and the executions. So in the
2:23:43log, this is what actually shows us like
2:23:45why it failed and what happened. So I
2:23:47don't really know what this means,
2:23:48right? All I'm going to do is copy this
2:23:50entire string of text. We're going to go
2:23:53back into cloud code and I'm actually
2:23:56going to go ahead and clear this because
2:23:57we're at 64% context. So just going to
2:23:59restart fresh. So, I just tried to do a
2:24:01manual run of our YouTube weekly report
2:24:04app in modal and this is the error that
2:24:06I got. And then I paste in all that
2:24:08messy stuff and shoot it off. Okay, so
2:24:12because we tested so much and we were
2:24:13using the free tier of the YouTube data
2:24:15API, we actually just hit the daily
2:24:16limit, which was about 10,000 units and
2:24:19we exceeded that because we were doing
2:24:20so much testing to see how well this
2:24:22thing would work. The good news is if
2:24:24this is actually running weekly, we will
2:24:25never hit that daily quota limit. So,
2:24:27we're fine. But the bad news is we're
2:24:29not going to test this one right now.
2:24:30But at least it does suggest other
2:24:32options and some longer term fixes. But
2:24:34it's okay because I did want to end off
2:24:36by showing how you could deploy
2:24:37something with a web hook trigger rather
2:24:39than a schedule trigger. So what I did
2:24:41is I came into this other workflow that
2:24:42I built the other day which is a very
2:24:44simple lead web hook notification. So it
2:24:47has a web hook as the trigger. We would
2:24:50see a company name and some other data.
2:24:52We would research the company with
2:24:53Replexity and then send an email
2:24:54notification. And so I basically just
2:24:56said, "Hey Cloud Code, can you push this
2:24:58workflow onto modal as we did earlier?"
2:25:01And now we have this app in our modal.
2:25:03As you can see, lead-web hook. So what
2:25:05I'm going to do is go to Postman. So we
2:25:07can actually hit that web hook just to
2:25:08simulate what would happen. We've got
2:25:10the address, we've got the body, and
2:25:12I'll shoot this off. And what this is
2:25:13going to do is it's going to trigger
2:25:15this form endpoint in modal. So I'll
2:25:17click into that one. And you can see
2:25:19right now we have a status of pending.
2:25:21This one's going to start running. And
2:25:22then it will show that we actually get
2:25:24the email in Gmail. And so this is
2:25:26really just to show that once you have
2:25:28your stuff up and running in modal, it
2:25:30will work. And you can also do it based
2:25:31on web hooks rather than just doing it
2:25:33on a cron. So that looks like it
2:25:35finished up. We can see that we just got
2:25:36this email for the new lead Chipotle
2:25:38where it did some research about them
2:25:40and then obviously it gave us a
2:25:41notification here. And now what you
2:25:42could do is because you just went
2:25:44through the process of deploying a
2:25:45workflow to modal and you know that it
2:25:47works because you just validated that
2:25:48it's working. You have all of that
2:25:50history right there. And what you could
2:25:51do is say, "Okay, cool. Keep this stored
2:25:54either in my claw.md file or let's
2:25:57create this as a skill so that every
2:25:59time later when you're building a
2:26:00workflow and you want to actually push
2:26:01it to modal, you have all that
2:26:03information already there, whether
2:26:04that's a skill or whether it's in the
2:26:05system prompt of cloud.md." So, I hope
2:26:08you guys at this point can see how
2:26:10Claude Code makes this stuff really,
2:26:12really easy to get automations up and
2:26:13running. Whether that means an
2:26:15automation that you want to be there for
2:26:16and you trigger kind of to use as like a
2:26:18personal assistant or an automation that
2:26:20you actually want to host somewhere and
2:26:22have it run on some sort of trigger and
2:26:24you can tap into all the skills that
2:26:26other people have been building and
2:26:27using because you can find those
2:26:29publicly and then just add those to your
2:26:30own instance. So now you have the super
2:26:33smart model like Sonnet 4.5, Opus 4.5
2:26:36paired with all of these really good
2:26:38prompts and really good like MCP
2:26:39servers. So you can pretty much do
2:26:41anything in that environment. The more
2:26:42you start to use it, the more you'll
2:26:44realize that you don't have to actually
2:26:45switch around to a bunch of different
2:26:46Chrome tabs and different apps on your
2:26:47desktop. You can do a lot of the stuff
2:26:49that you need to do just in the cloud
2:26:51code environment itself. So once again,
2:26:53that claw.md file that you guys can
2:26:55access for free will be in my free
2:26:56school community. The link for that will
2:26:57be down in the description.
2:27:01So I've been using Cloud Code for a
2:27:02while now, and I opened up trigger.dev
2:27:04for the first time, I'm not kidding you,
2:27:05like an hour and a half ago, and I've
2:27:07already got a couple really nice
2:27:08automations and agents set up. I can
2:27:10already tell that this combination is
2:27:12going to be a core piece of my workflow,
2:27:13whether it's internally or for clients.
2:27:15And this combination of Cloud Code and
2:27:17Trigger Dev, is insanely powerful. So,
2:27:19the first thing I tried was just
2:27:20building a workflow to scrape Nate B.
2:27:22Jones YouTube videos. He's like my
2:27:24favorite AI news channel. And it would
2:27:26check if he's got a new video. If yes,
2:27:28it would give me the key highlights. If
2:27:29no, it would do nothing. And so, this is
2:27:30what the result looks like with minimal
2:27:32prompting. We've got the YouTube video.
2:27:34We've got key concepts, quotes worth
2:27:36remembering, stats, and data. So that's
2:27:38cool because I spun this up in like 10
2:27:39minutes, but really it's not that
2:27:40impressive. But then I thought to
2:27:42myself, because of the way that
2:27:43trigger.dev works, I want to put an
2:27:45agent in the cloud. So what I did is I
2:27:47built an agent that watches this list in
2:27:49my ClickUp. And whenever I put in a
2:27:51company as a task, it will basically do
2:27:53research about them. So I put this in,
2:27:54it triggers the agent, it does research,
2:27:56and then it comes back and leaves a
2:27:57comment for me. But not only that, but
2:27:59the agent can also conversate with me in
2:28:01here. So this one for Enthropic, I said,
2:28:03"Hey, does this company have a recent
2:28:04valuation?" And it did more research and
2:28:06came back to me. So, this is not just a
2:28:08deterministic 1 2 3 4 5 type of
2:28:10automation. This is a non-deterministic
2:28:13I have different tools. I need to decide
2:28:14what loop I need to go in. And that's
2:28:16where I was like, okay, wow, I just did
2:28:18all of this in like 20 minutes. This is
2:28:20cool. So, by the end of this video, you
2:28:21will understand exactly how you can
2:28:22build workflows and agents in cloud
2:28:24code, put them on trigger.dev, and
2:28:26automate pretty much anything. So,
2:28:27before we jump into building, let's just
2:28:28get on the same page. We use cloud code
2:28:30to describe what we want in plain
2:28:32English. This could be saying something
2:28:33as simple as, "I want to monitor YouTube
2:28:35for new AI videos and send me a daily
2:28:37summary." Cloud Code will then say,
2:28:38"Okay, well, to do this, I need X, Y,
2:28:40and Z from you." So, it turns a vague
2:28:42request into an actual working
2:28:44automation. Then, it goes ahead and it
2:28:45starts writing its code. And like I
2:28:46said, these can be simple tasks like
2:28:48fetching data from an API, or they can
2:28:50actually be kind of complex agents that
2:28:52have different tools and have different
2:28:54actions they can take and they have to
2:28:55decide when is good enough. And that
2:28:58actual automation or that agent is code
2:29:00and that is basically a project and
2:29:02that's like a file. And then what we
2:29:04need to do is get that out of our local
2:29:05cloud code environment on our laptop or
2:29:07our desktop and put that into
2:29:09trigger.dev on the cloud so that these
2:29:12can actually run all the time. And so in
2:29:13past cloud code videos, you might have
2:29:15seen me use something like modal. So why
2:29:16are we using trigger.dev over modal
2:29:18today? It's just a lot more flexible. It
2:29:20has scheduled runs. It has automatic
2:29:22retries. It's got queuing. It's got
2:29:23orchestration, which is the element of
2:29:25different tasks and things like that,
2:29:26which is really cool. And I just think
2:29:28it's got a much cleaner UI as well. So,
2:29:30here's my current project in trigger.dev
2:29:32in production mode. And you can see that
2:29:33I've got basically these six different
2:29:35kind of tasks. These three are the
2:29:38actual tools. So, this is the process
2:29:39video. This is the responder agent. This
2:29:41is the researcher agent. And then these
2:29:43ones with a clock next to them are
2:29:44schedule tasks. So, we've got the
2:29:46YouTube checker, we've got the research
2:29:47polar, and the follow-up polar. If I go
2:29:50over here on the lefth hand side to
2:29:51schedules, you can see how often these
2:29:52run. So these two run every 2 minutes
2:29:54and this one runs every 8 hours. If I go
2:29:57to runs, you can see all of the
2:29:58different runs that we've had, whether
2:29:59they have been completed or whether
2:30:01they've failed. And so I filtered to
2:30:03show you a failed run because I wanted
2:30:04to see the retry. So this was a ClickUp
2:30:07research polar and what happened was it
2:30:09failed and so it adds a delay and then
2:30:12it tries again. So it has this automatic
2:30:14retry built into it which is really
2:30:16cool. And when we're watching these
2:30:17agents or tasks run live, we see exactly
2:30:20every step that they're making and we
2:30:21see how long they take on each step. So,
2:30:23I'll do a live demo and I'll show you
2:30:24guys this. All right, so I'm in my
2:30:26ClickUp and what I'm going to do is add
2:30:27a new task and I'm just going to go
2:30:28ahead and say Nvidia and I'm going to
2:30:30save that. And now that that has been
2:30:32processed, we're going to have to wait
2:30:34in here and we'll see the ClickUp
2:30:36research polar start up and then it will
2:30:39basically say, "Okay, cool. I found a
2:30:40new task and I'm going to send that to
2:30:42the company researcher agent." So right
2:30:44here we can see that that got picked up.
2:30:46It got sent to the researcher. And if I
2:30:47open the researcher, which is currently
2:30:49executing, you can see it live running.
2:30:51So we can see right here that it's
2:30:52running for Nvidia, which is the one
2:30:54that we just put in here. Right now,
2:30:55Claude is calling its search web tool,
2:30:57and it called it twice. It actually
2:30:58decided I'm going to invoke it one more
2:31:00time to make sure I get a comprehensive
2:31:02research report. Now it's using a read
2:31:04URL tool. It's using another read URL
2:31:06tool. So maybe it found two websites to
2:31:08look at. And there we go. After about 45
2:31:10seconds, it has finished. So now if I
2:31:12open up ClickUp and we go to Nvidia, we
2:31:14can see that this is now marked as
2:31:15complete. And when I click on it over
2:31:17here, we can see that we got a full
2:31:19research brief about Nvidia. And so now
2:31:21if I come into this task and I say at
2:31:23UPAI, how is their stock doing? It's
2:31:26going to read this. But keep in mind, it
2:31:27has to come in here and it has to read
2:31:29the context to make sure it knows who is
2:31:31they in this case when I just said how
2:31:33is their stock doing. And if I go back
2:31:34to runs, you can see that the follow-up
2:31:36responder is now executing. If I click
2:31:38into it, we should be able to watch it
2:31:39actually looking at this task searching
2:31:42the web. And now after about 22 seconds,
2:31:45this one has finished. I'm going to open
2:31:46up ClickUp. And we should be able to
2:31:48scroll down and see that we just got a
2:31:50response right here from Upet AI. Okay,
2:31:53so just a little quick demo and wanted
2:31:54to show you guys what trigger.dev looks
2:31:56like before we actually get in there and
2:31:57we build one out ourselves. And I know
2:31:59these aren't the most impressive things
2:32:01in the world, but keep in mind all of
2:32:02these were oneshot prompted and I built
2:32:04these in the past like 45 minutes. Okay,
2:32:07now hopping into cloud code. This is the
2:32:09cloud workflow builder project that I
2:32:10just set up and I started building these
2:32:12workflows in. If you've never used cloud
2:32:14code before, then I would definitely
2:32:15recommend you hop over to this video.
2:32:17I'll link it right up here. And then
2:32:18once you understand like the interface
2:32:20and how the files work and everything
2:32:21like that, then come back and it will
2:32:23make a lot more sense.
2:32:25So, what I'm going to do in this video
2:32:26is I'm going to create a brand new
2:32:28project and walk through everything with
2:32:29you guys step by step so you can see
2:32:30exactly how it works. But real quick
2:32:32before we do that, I just wanted to show
2:32:33you kind of what this end result looks
2:32:35like. Because these folders and files
2:32:37build up as you build more workflows.
2:32:39Really, the main thing that I wanted to
2:32:40show you guys is where those TypeScript
2:32:42files actually live, which is right here
2:32:44in a SRC. We've got the trigger folder,
2:32:47which is trigger.dev, and then I've got
2:32:49two different folders for the different
2:32:50types of workflows. We've got the AI
2:32:52news digest, which is the two, you know,
2:32:54the process video typescript, and the
2:32:56YouTube check type script, which you
2:32:57guys saw right here. YouTube check and
2:32:59process video. And then we have the
2:33:01company research ones, the two polars
2:33:02and the responder and the researcher
2:33:04agent, which you guys saw right here,
2:33:06the two polars and the researcher and
2:33:08responder. So really what I wanted to do
2:33:09there is just contextualize for you guys
2:33:11how cloud code helps us build these
2:33:13TypeScript files and then we push those
2:33:15to trigger and then they can actually
2:33:17run live. But I did promise you guys
2:33:19we're going to do all of this together.
2:33:20So I'm going to go ahead and open up a
2:33:22new folder here. I made a new folder
2:33:24called trigger demo and it's completely
2:33:26blank. And now you guys should have your
2:33:29Claude Code instance looking exactly
2:33:30like I do. And by the way, I am using
2:33:32this in VS Code. So I'm going to close
2:33:34out of this stuff. Open up Claude Code.
2:33:37And we are now on the same page. So the
2:33:39first thing I want you to do is go over
2:33:40to my free school community. The link
2:33:42for that's down in the description. Go
2:33:43to the classroom. Go to Claude Code and
2:33:45go to the trigger.dev section right
2:33:48here. And you're going to grab the
2:33:49claw.md file and the trigger ref.md
2:33:52file. You're going to download both of
2:33:53those. And then you're basically just
2:33:55going to drag both of those in to the
2:33:56lefth hand side right over here. I'm not
2:33:58going to dive super super deep into
2:33:59this, but the trigger ref file is
2:34:01basically like the trigger.dev API
2:34:02reference and how to do TypeScript. And
2:34:05that's important because in the bottom
2:34:06of the claw.md file, we tell it to look
2:34:09here if it ever needs to look at code
2:34:11examples, patterns, and things like
2:34:12that. So now that you guys have the
2:34:14claw.md file in here, you've got the
2:34:16trigger refile in here. We're going to
2:34:18be able to build automations a lot
2:34:20easier now. All right. So, what I'm
2:34:21going to do is give Cloud Code a pretty
2:34:24vague request. I'm just not even going
2:34:25to use plan mode, and I'm going to show
2:34:26you how good it's going to be. I need
2:34:28you to build me an automation that's
2:34:29going to go off every single Monday, and
2:34:31it's going to search the web, and it's
2:34:33going to find me um leads. It's going to
2:34:35find me dental practices that I can sell
2:34:39websites to. So, I'm going to shoot that
2:34:40off. And what it's going to do first of
2:34:42all is it's reading the claw. I'm
2:34:43defiled to figure out what its goal is.
2:34:45And hopefully it should come back and
2:34:46ask us some questions so that it knows
2:34:48how to build this workflow well because
2:34:51we didn't tell it anything about text
2:34:52stack or hardly anything at all and it
2:34:54needs to write this automation really
2:34:56well for us. So the first thing it does
2:34:58is it asks us where should the leads be
2:35:00delivered each Monday and I'm just going
2:35:01to go ahead and say ClickUp. For
2:35:03location it says where should it be
2:35:05searching? I'm just going to say
2:35:07nationwide. And then the final one is do
2:35:09you have SER API, Google Maps, anything
2:35:10like that set up? And I'm actually for
2:35:12this one going to click other and I'm
2:35:13going to say I don't yet have any of
2:35:15that set up. And ideally I don't want to
2:35:16have to pay for any sort of subscription
2:35:18for this. And we'll see if it can figure
2:35:20it out with that. I'm going to tell it
2:35:22to create a new list in my ClickUp. And
2:35:24for volume I'm just going to go ahead
2:35:26and do 25 leads small batch just to
2:35:28start. Now mine is going to cheat a
2:35:30little bit because in the past it's
2:35:31already used my ClickUp. But what you
2:35:32would need to do is it will basically
2:35:34create you av file and you will put your
2:35:37project secrets in thatv file. So that
2:35:40will be like your ClickUp API key or
2:35:41your workspace ID or your open router
2:35:43API key, things like that. So now it
2:35:46asks us if we're ready to build this
2:35:47plan. So we have our Monday dental lead
2:35:50generator. Every Monday at 8 a.m. we
2:35:52will search Yelp Fusion AI for dental
2:35:54practices across the US. And then we'll
2:35:56create new tasks in the dental leads
2:35:59list in ClickUp. Yelp Fusion's 100%
2:36:02free. Awesome. For the architecture,
2:36:03it's going to create basically two
2:36:04different tools. Let's call them. The
2:36:06first tool is going to find leads and
2:36:07the second tool is going to create the
2:36:09leads in ClickUp, which is great because
2:36:12if one thing fails, it can retry and it
2:36:14can, you know, ceue up just that one
2:36:15task rather than having the whole
2:36:17automation break. And this is awesome.
2:36:19It's automatically doing the item
2:36:21potency like I talked about earlier. And
2:36:23it's making sure that the same practice
2:36:24never gets added twice. And having this
2:36:26dduplication process be automatically
2:36:28taken care of is really nice. So, I'm
2:36:30just going to tell this to go ahead and
2:36:31start building. So, as usual, it starts
2:36:32setting up the project, as you can see,
2:36:34as well as it is creating its to-do
2:36:36list, and it's going to go through and
2:36:37build this out for us. Okay, it built
2:36:39that really fast. It said that it made a
2:36:41new ClickUp list for us, which if I go
2:36:43to ClickUp, we can see right here, we've
2:36:44got dental leads. It created the two
2:36:46type scripts. So, if I look up here,
2:36:47we've got trigger, we've got dental
2:36:48leads, and we have create lead as well
2:36:50as find leads. And this is the part
2:36:52where I said it cheated a little bit
2:36:53because it already had our list ID from
2:36:56earlier. But what we need to do now is
2:36:58go get our Yelp API key. So, I'm going
2:36:59to go grab that. And then we need to put
2:37:01in a ClickUp API key. Wow. So it ended
2:37:04up finding out that Yelp killed their
2:37:05free API tier. So it's going to change
2:37:07this to use SER API instead. Okay. So
2:37:10now it's already changed that to use SER
2:37:11API. So I've got those two API keys. And
2:37:14here's what we're going to do. We're
2:37:15going to go to the ENV and you can see
2:37:17that it's put this placeholder in here
2:37:18for SER and for ClickUp. So you're going
2:37:20to come in here and paste in your two
2:37:22API keys. And then you're going to make
2:37:23sure to save this file and then you can
2:37:25close out of it. Okay. So it wants to do
2:37:27its test run in the dev environment in
2:37:30our trigger.dev. So remember we're in
2:37:32production. We can also be in dev. So
2:37:34this is where we can test things. This
2:37:35is where we can make changes. And then
2:37:37when we're good with that, we push that
2:37:38to production to actually be live. So
2:37:40what I'm going to do is in my
2:37:41trigger.dev, I'm going to open up a new
2:37:42project since this is what your guys
2:37:44would look like. We're just going to
2:37:45call this one test. And I'm going to go
2:37:47ahead and create that. Now once we're in
2:37:49here, I'm going to go down to the bottom
2:37:50and go to project settings. And right
2:37:52here we see a project ref. So I'm going
2:37:54to copy this value. And we need to give
2:37:56this to Claude Code so it can actually
2:37:58move stuff into here. Okay, so it's
2:37:59working away right now, but you can see
2:38:01that we do have two tasks in here. We
2:38:03have our find dental leads and we have
2:38:05our create dental lead. If I go to
2:38:06schedules, we should see that the find
2:38:08dental leads runs at 8 a.m. only on
2:38:10Mondays. And now what it's doing is it's
2:38:11going to try to test it out to see if it
2:38:13actually works. But what's going to
2:38:14happen is it's probably not going to
2:38:16work because if you guys remember, we
2:38:18had to give Claude code here our env,
2:38:22you know, ClickUp API key as well as
2:38:24that SER API key. But those API keys
2:38:27don't actually get pushed anywhere. They
2:38:28don't get pushed to GitHub. They don't
2:38:29get pushed to trigger.dev because
2:38:32they're in AENV file. And if anything
2:38:34here has a dot before it, it's basically
2:38:36hidden from like commits and stuff like
2:38:37that. So what we actually have to do is
2:38:39we have to go into trigger.dev. We have
2:38:41to go all the way down here to
2:38:42environment variables and this is where
2:38:44we add those once again. So I'll click
2:38:46add new and all I have to do actually
2:38:48it's really simple is I come into thev
2:38:50file over here. I can copy this entire
2:38:52thing and then when I go back into
2:38:54trigger I can paste in that entire thing
2:38:57and all three of those will get put in
2:38:59not just like having to do that one by
2:39:00one. And now before you save this you're
2:39:02going to want to do it in development
2:39:04and production because there's no point
2:39:05in adding it just to development because
2:39:07ultimately you're going to push it to
2:39:08production. So you might as well just do
2:39:10development and production and then go
2:39:11ahead and save that. And now trigger
2:39:14when it's actually running those
2:39:15Typescript you know files it can
2:39:18actually use these API keys in its work
2:39:21and it will actually be successful. So
2:39:23what happened was this was trying to
2:39:25actually trigger a test run in trigger.
2:39:28So it was trying to do this by itself.
2:39:30But one thing that you can do to make
2:39:31this actually a lot better is you can
2:39:33give it the trigger.dev MCP server. So
2:39:36here's a file that you can drag into the
2:39:37lefth hand side. It basically just sets
2:39:39up the MCP configuration for
2:39:41trigger.dev. And I will also link this
2:39:43in the same exact classroom section
2:39:45right here. So you can just download it
2:39:46and drag it in. But now I can go ahead
2:39:48and ask it to try to test it out again.
2:39:50And hopefully it works and it can send
2:39:51over a payload and all that kind of
2:39:52stuff. And the thing I want you guys to
2:39:54understand too is yes, I just did this,
2:39:56but it it's a little bit different every
2:39:58time you build it because ultimately
2:39:59what's happening is we're talking to an
2:40:00AI model that thinks and has decision
2:40:03capabilities. So, if you're entering in
2:40:05the exact same prompts that I am, you
2:40:06might not get the exact same result. So,
2:40:08it's really just a matter of talking to
2:40:10it, asking what's wrong, and helping it
2:40:12out. And so, now it's asking us for a
2:40:14trigger.dev secret key. Well, let's go
2:40:16into trigger.dev. Let's go over here and
2:40:17go to API keys. And we need to grab this
2:40:20and give that to cloud code. But, of
2:40:22course, I don't want to give this
2:40:23straight into the chat. So, I'm going to
2:40:24say, "Hey, I've got this. Can you put a
2:40:27placeholder in thev?" And then I'll
2:40:28paste it in there. It's really just best
2:40:30practice to never give secrets right
2:40:33here in the chat. You want to just put
2:40:34it into like config files or usually the
2:40:37EMV. Okay, so just paste it in that API
2:40:39key. I've saved it and now let's see if
2:40:41it can actually restart the server and
2:40:43try to test out our workflow. This is
2:40:45why it's so tough doing like live cloud
2:40:47code demos. It's just cuz like it
2:40:48literally runs different every time.
2:40:49Okay, so it says that it triggered a
2:40:51run. Now, if I come back in here and I
2:40:52go to runs, we should see. Wow. Okay, so
2:40:55we actually got a lot of different runs
2:40:57come through. It was trying a lot of
2:40:58different things. These are currently
2:40:59executing right now. Oh, okay. Well, the
2:41:01reason why it did so many is because it
2:41:03had to do one create dental lead for
2:41:05every single lead that it found. And if
2:41:08I open up my ClickUp, we can see that we
2:41:09do have 25 leads right here. So, let's
2:41:11just click into one to see what it looks
2:41:12like. We've got Tampa Dental in Tampa,
2:41:15Florida. And in the description, it gave
2:41:17us the address, the phone, the rating,
2:41:19and the website. So, that's really
2:41:20interesting to me. The first workflow,
2:41:22which is the trigger, it finds the
2:41:24leads. So, if I open this up and I click
2:41:26into it, we can see what it did. It
2:41:27found five and then it found five more
2:41:29and it found five more until it got to
2:41:3125. And then after it found those 25, it
2:41:33basically said, "Okay, cool. I'm going
2:41:35to hand each of these leads to one
2:41:37individual kind of like worker." And
2:41:39that is why we saw the create dental
2:41:41lead tool get called 25 times. Cloud
2:41:44code comes back and says, "Cool, it
2:41:45works perfectly. 25 leads created in 9
2:41:47seconds across five cities. Pretty
2:41:49awesome." So from here, obviously, what
2:41:51you do is you'd iterate upon the
2:41:52workflow a little bit. You'd try to make
2:41:53it better and better. You'd add
2:41:54different features if you want. Maybe we
2:41:56want to add personalized outreach and
2:41:57add like an AI element in there. But now
2:41:59once you get to a spot where you're
2:42:01ready, we want to push that into
2:42:02production because right now we're in
2:42:04development. Meaning if we don't keep
2:42:06this connection open right here, it says
2:42:09your local dev server is connected to
2:42:10trigger.dev. If this wasn't open like we
2:42:13see in our main project right here, you
2:42:15can see I turned it off. Then these
2:42:16would not be running, which is why we
2:42:18have to push them to production. So
2:42:20there's a few ways to do that. So the
2:42:22way that I would recommend doing this is
2:42:23by connecting to GitHub and pushing your
2:42:25codebase. So basically this entire
2:42:27project into GitHub. So like you can see
2:42:29this is the GitHub repo that I made for
2:42:31the previous workflows and I had all of
2:42:33them in one GitHub repo which is fine in
2:42:35this case. If I go to the SRC trigger
2:42:37you can see that we had the AI news
2:42:39digest and the company research. So
2:42:40here's what we're going to do. We're
2:42:42going to push this project into GitHub
2:42:44and then have GitHub sync with
2:42:45trigger.dev and then they'll be able to
2:42:47bring everything into production. And
2:42:48now once again I just wanted to throw
2:42:50out there this is why you have to make
2:42:51sure that things are in aenv and that
2:42:53you're not having any secrets in your
2:42:55project because otherwise they'll be in
2:42:58GitHub and I always make them private
2:43:00either way but still it's just best
2:43:02practice and then you're going to add of
2:43:03course your API keys into trigger.dev.
2:43:06So at this point what it's probably
2:43:07going to do is it's probably going to
2:43:08use your command line interface and it's
2:43:10going to ask you to login with GitHub.
2:43:12So once you get authenticated, it's
2:43:13really easy and it can create repos and
2:43:15it can, you know, make commits and it's
2:43:17very very simple. So if you guys have
2:43:19watched any of my videos where we've
2:43:20built websites, then you know exactly
2:43:22how this flow works. We're building code
2:43:24in cloud code. We push that to GitHub
2:43:27and then GitHub automatically syncs with
2:43:29Forcell or in our case trigger.dev,
2:43:31meaning that we have our actual live
2:43:33automation or our live site
2:43:34automatically syncing with GitHub. And
2:43:36the reason why we want to use GitHub is
2:43:37for version control. It keeps things,
2:43:39you know, in the cloud and we can also
2:43:41do more collaboration with other
2:43:43engineers, other people, whatever it is.
2:43:45So, I'd say getting familiar with GitHub
2:43:47is probably a good idea. And just to
2:43:49really hammer it home, you can see it
2:43:50says good.v is excluded from our commit
2:43:53to GitHub. Okay, awesome. So, we now
2:43:56have our new GitHub repo right here.
2:43:57It's called trigger- test. So, I'm going
2:43:58to open that up. Make sure we are good.
2:44:00Do we have our workflows? Yes, we do.
2:44:02Right there. And now I'm going to go
2:44:04back into trigger. I'm going to scroll
2:44:05all the way down to project settings and
2:44:07I'm going to connect a GitHub repo. I'm
2:44:09going to sign up with my account and I'm
2:44:10going to connect the repository that we
2:44:12just made and hit connect. Now, right
2:44:14here it says every push to the selected
2:44:16tracking branch creates a deployment in
2:44:18the corresponding environment. Which
2:44:20means every time we have a new push to
2:44:21the master branch in our GitHub repo, it
2:44:23will automatically get synced into our
2:44:26production environment. So, right now
2:44:27we're in development. If I go to
2:44:28production, we should see that it's
2:44:30going to actually build up these things
2:44:33now that I've connected our GitHub repo
2:44:35because right here in GitHub, you can
2:44:36see that we are in the master branch and
2:44:38we have our TypeScript right here. Now,
2:44:41if for some reason this isn't working or
2:44:42you don't want to take the GitHub route,
2:44:44you can always manually deploy this kind
2:44:46of stuff and you can just tell Cloud
2:44:47Code to look at this and it will help
2:44:49you push this into the production
2:44:51environment. Okay, so you can see that
2:44:52those just got pushed through into
2:44:54production. And unfortunately in live
2:44:56production, this would actually only run
2:44:58at 8 a.m. on Mondays. So we're going to
2:44:59do a manual run right here. I'm going to
2:45:01go in here. I'm going to hit test. And
2:45:03I'm just going to go ahead and hit run
2:45:04test. And so that's going to go ahead
2:45:06and trigger this off. We're going to
2:45:07watch this actually find those 25 leads.
2:45:10Remember, it should automatically be not
2:45:12loading any in if it's already processed
2:45:14them before. So it's looking in Tampa,
2:45:16Tulsa, Arlington. It found these leads.
2:45:18And if I go to runs, it's also it should
2:45:21be scheduling all of these for ClickUp,
2:45:23which 7:29 p.m. It's doing that right
2:45:26now. And if I go into our ClickUp, we
2:45:28should see now that we have 50 leads in
2:45:30here. And I just want to make sure that
2:45:31none of these have been duplicated.
2:45:34Okay. Actually, it looks like some of
2:45:35these are. So, let me just show you guys
2:45:37real quick what we might do about that.
2:45:38So, I just did a test run of finding
2:45:40leads, and it all worked. However, I did
2:45:42notice that some of the leads were
2:45:43duplicates. So, I thought that you had
2:45:45worked on item potency here to make sure
2:45:47that that didn't happen, but apparently
2:45:49it was happening. So, can you fix this?
2:45:52Okay, so it made a change in the create
2:45:53leads workflow where it's basically
2:45:55going to search ClickUp before creating
2:45:57and now have a place ID. So, it says
2:45:59that this change should be more
2:46:00permanent, but it doesn't affect the
2:46:01ones that already were made. So, I'm
2:46:02going to go ahead and run this twice and
2:46:04we'll see if there's any duplicates.
2:46:06Okay, so I had to do a few more changes
2:46:08and this really shows the importance of
2:46:09using plan mode and having a really good
2:46:12idea of what you want built before you
2:46:13actually tell it to go build it. For the
2:46:15sake of the video, I was trying to show
2:46:16you how I could just kind of oneshot
2:46:18prompt and and it would still be pretty
2:46:19good. But this example showed really why
2:46:22you need to plan harder because the
2:46:24search criteria wasn't very big and it
2:46:25had a weird way of dduplication. But now
2:46:27we're at a much better spot where we now
2:46:29have 48 rows and it shows that some of
2:46:32them were filtered out because they were
2:46:33duplicates and it was like searching
2:46:35again and again until it was able to
2:46:36fill up all the slots and now we have
2:46:39more companies. So anyways, all of these
2:46:42were being ran in our development
2:46:43environment because we didn't want to
2:46:44mess with anything in prod. And now what
2:46:46we would do is we'd come back in here
2:46:47and say, "Cool, this looks good. Push
2:46:49this to GitHub." and then trigger would
2:46:50automatically pull those live changes
2:46:52from GitHub and we'd be all set for
2:46:54Monday at 8 a.m. Okay, so hopefully by
2:46:56now you guys feel confident in your
2:46:57ability to hop into Cloud Code, go to my
2:46:59free school community, grab these
2:47:00resources, chuck them in there, and then
2:47:02start building some automations that you
2:47:04can then throw into trigger.dev to just
2:47:06run all the time for you. If there's one
2:47:08thing that I want you to take away from
2:47:09this video, it's the fact that AI is
2:47:10still very much a blackbox. These models
2:47:13are insanely smart. you can see what
2:47:14they do. But you also saw in this video
2:47:16how I had to talk so much to it and I
2:47:18had to like just be very clear. It's
2:47:20just the whole idea of you no longer
2:47:22having to write code, but you having to
2:47:24be the person that assures the quality
2:47:26and make sure that it's on track. But
2:47:28anyways, I'm having so much fun learning
2:47:29these new tools and diving into
2:47:31different use cases. So, definitely let
2:47:32me know in the comments what you guys
2:47:33want to continue to see.
2:47:37Cloud Code is now a 247 AI employee,
2:47:40which means that it is always Cloud Code
2:47:43O'. And that might be the nerdiest thing
2:47:44I've ever said. Enthropic has finally
2:47:46done it and launched scheduled tasks
2:47:48natively for Claude Code, which means
2:47:50every single process, every single
2:47:51skill, everything you've been building
2:47:53and using inside of Cloud Code just got
2:47:5510 times more powerful. And they
2:47:57literally could not be easier to set up.
2:47:58So, in today's video, I'm going to show
2:47:59you exactly how they work and tell you
2:48:01everything important that you need to
2:48:02know about them. Let's not waste any
2:48:04time and get straight into it. So, here
2:48:05I am in Cloud Code in the Claude desktop
2:48:07app. Now, right now, you do need to be
2:48:09using the desktop app in order to access
2:48:11these scheduled tasks. Now, this
2:48:12scheduled task feature came out about a
2:48:14week or two ago in Claude Co-work. As
2:48:16you can see, it's basically the exact
2:48:18same thing. You could create scheduled
2:48:19automations, but now they finally
2:48:21brought it to Claude Code. And there's a
2:48:23few ways that you can set them up. The
2:48:24first way is you go over here to the
2:48:25schedule tab, and you just click on it.
2:48:27And you can see right here, run tasks on
2:48:29a schedule or whenever you need them,
2:48:30type/schedule, and any existing session
2:48:33to set one up. So, those are the two
2:48:34ways. You can either click right here,
2:48:36new task. You can give it a name, a
2:48:37description, give it a prompt. You can
2:48:39choose the model you want it to run on.
2:48:40You can choose the mode to run on and
2:48:42you can also select the folder. And then
2:48:44finally, you just say, "Hey, I want this
2:48:45to run every hour, every day, every week
2:48:48at this time." Boom. You now have a
2:48:50scheduled automation. So that cron would
2:48:52basically fire off. The session would
2:48:54start up. And then the agent would read
2:48:55the prompt. It would go through your
2:48:56files. It would work in your project, do
2:48:58whatever it needed to do, and then after
2:49:00it's done, it would just stop. And the
2:49:02huge unlock here, which is so exciting
2:49:03to me, is that this isn't a
2:49:05deterministic workflow. And so there's
2:49:07some good there and there's some bad
2:49:08there. But the good news is that you can
2:49:09completely control it. And what I mean
2:49:11by that is if you've been following some
2:49:12other cloud code videos I've done in the
2:49:14past, we build either like a Python
2:49:15script or a TypeScript. And that is the
2:49:18actual automation. And that is
2:49:19deterministic logic code. Meaning that
2:49:22will always happen step one, step two,
2:49:23step three. If there's an error, it
2:49:25can't fix itself. It just errors. And
2:49:27then we get notified. But this isn't
2:49:28just a Python script. This is cloud code
2:49:30agent running the same exact way it runs
2:49:33when you talk to it. And that's why this
2:49:34is so exciting because Agentic workflows
2:49:36are self-healing and they can read
2:49:38everything in your entire project and
2:49:39use all your tools. So, as you guys
2:49:41know, you tell it to go do something and
2:49:42it starts trying. If it runs into an
2:49:44error, it doesn't just come back to you
2:49:45and say, "Eh, I tried." It says, "Okay,
2:49:47here's the error. Let me try three other
2:49:49things." And then after I see which one
2:49:50of those three other things worked best,
2:49:52I'm going to update myself so that I
2:49:54never run into that error again. So, now
2:49:55you are no longer the bottleneck. And
2:49:57these skills and these workflows can
2:49:59actually get better and better over time
2:50:00automatically. But the other important
2:50:02thing to remember here is if you do want
2:50:03it to be more deterministic and you want
2:50:05more control, you can do that because
2:50:07you could literally have it just execute
2:50:08a script and that's the whole scheduled
2:50:10task and it just is completely
2:50:11deterministic that way. So how do you go
2:50:13ahead and start creating some scheduled
2:50:15tasks? Well, here you can see one I have
2:50:16is called morning coffee. So I've showed
2:50:18this one off before in a few other
2:50:19videos, but basically every single
2:50:21morning I would open up my cloud code
2:50:22and I would say, "Hey, run morning
2:50:24coffee." Which would help me plan my
2:50:25day. It would look at my commitments. It
2:50:26would look at the projects and help me
2:50:28catch up on what the team's up to. But
2:50:30now this can actually just run
2:50:31automatically at 6:00 a.m. every
2:50:33morning. And literally all that I did to
2:50:34set this up was I said, "Take a look at
2:50:36my morning coffee skill. I would like to
2:50:37turn this into a scheduled task that
2:50:39goes off every morning at 6:00 a.m. Help
2:50:40me get this set up." It read the skill.
2:50:42It complimented me on the skill. And
2:50:44then it asked me one question about it.
2:50:45And then a minute later, my skill that I
2:50:47run every morning is now automated. And
2:50:49if you've never used Cloud Code in the
2:50:50desktop app, don't worry. It's super
2:50:52easy. You can basically pull in, you
2:50:53know, a GitHub repo or a different
2:50:55folder and you can be working in the
2:50:56exact same project that you're used to.
2:50:58So right here in a new session I just
2:50:59asked it what skills do you have in this
2:51:01project and it came back and said hey
2:51:03here are all the active skills we've got
2:51:04content creation we've got research and
2:51:05intelligence we've got visual diagrams
2:51:07operations and meta and now any of these
2:51:10skills I could just say okay cool turn
2:51:11that into a weekly automation. All right
2:51:14so there are a few limitations though
2:51:15the first major gotcha is the fact that
2:51:17your laptop has to be on or your
2:51:19computer has to be on and the desktop
2:51:20app has to be open. So if you turn off
2:51:22your computer that automation will not
2:51:24run. Now, the good news is, let's say
2:51:26you had a task for 7 a.m. and you wake
2:51:27up at 8:00, you turn on your computer.
2:51:29Cloud Code would actually check back 7
2:51:31days and see any scheduled tasks that it
2:51:33missed and then would catch up and it
2:51:35would run those. Obviously, that's not
2:51:36perfect because some of those may be
2:51:37timesensitive, but it is cool that it
2:51:39has that ability. Now, what are some
2:51:40other things to think about? Well, the
2:51:42first one is that this thing is now
2:51:44running without your supervision. And
2:51:45ideally, it's not going to stop to ask
2:51:47you questions because then what's the
2:51:48point of having it automated? So that
2:51:50means you want to be looking at your
2:51:51permissions to make sure that it can't
2:51:52actually go off the rails and do
2:51:54anything like maybe make a major change
2:51:56to your GitHub repository or go off and
2:51:58delete things. And you can take care of
2:52:00that by changing your local settings in
2:52:02that project which you could just say,
2:52:03"Hey, I want to make sure that you never
2:52:05delete things. How can I put this in
2:52:06your settings?" You know, deny a bash
2:52:08command that does any deletes or
2:52:10removes. And it will help you figure
2:52:11that out. I've got a video coming about
2:52:13this. I will tag that right up here once
2:52:15that is live on YouTube. These are also
2:52:17stateless. So basically, every time that
2:52:18you run one of these, it's going to
2:52:20throw it in a new session. So right
2:52:22here, I did a test run of my morning
2:52:23coffee. Here's the actual task itself.
2:52:25And then when I open this up, I will be
2:52:27able to see every run. And every run is
2:52:28going to be fresh and it's not going to
2:52:30have context of what happened on the
2:52:32previous run. And the other thing, of
2:52:33course, is that if you didn't put in an
2:52:34API key or if there's literally
2:52:36something that it can't do because it
2:52:37needs your permission, it's going to
2:52:39stop. So what I'd recommend is as soon
2:52:40as you create a new task, just run it
2:52:42manually and make sure that it can go
2:52:43through all of the steps without
2:52:45oversight. Otherwise, what's going to
2:52:47happen is it's going to pause and it's
2:52:48going to ask you for permissions to do
2:52:49this and permissions to do that. Now,
2:52:51sometimes it's a good thing that they're
2:52:52stateless and that they don't have
2:52:53shared memory, but sometimes you might
2:52:55want them to, and that would be kind of
2:52:56part of this whole self-improving thing.
2:52:58So, here's how I imagine the
2:53:00self-improving loop working. So, first
2:53:02of all is fixing the actual script. So,
2:53:05you can have in there the fact that if
2:53:06it errors, edit your own code and make
2:53:08sure you're fixed. The second layer is
2:53:10the prompt. So, if you realize that
2:53:12there's an opportunity to improve this
2:53:13prompt, rewrite it and now you have a
2:53:15new prompt. And then the third one is
2:53:16potentially having a log for memory.
2:53:18Whether that means every single run you
2:53:20put some sort of like status of what
2:53:22that run did or maybe you just have a
2:53:23file that you overwrite so that every
2:53:25time when the new agent wakes up, it can
2:53:27look at the log and say, "Hey, this is
2:53:28what the previous agent just did. Now I
2:53:30need to run." So there's lots of ways
2:53:31that you can kind of tweak these
2:53:33scheduled tasks to fit your use cases
2:53:35better. And once again, because they
2:53:36have the context of everything in that
2:53:38project, they're going to be super
2:53:40powerful. They can look at any file that
2:53:41you want them to. So that's what my
2:53:43brain immediately went to. But I wanted
2:53:44to see what Claude code thought. So I
2:53:45asked it what it thought the most
2:53:47optimal strategy was to make these
2:53:48improve and make sure they have the
2:53:50right context. So what it thought of was
2:53:52a lean strategy where you have one file
2:53:54per task. So basically every single time
2:53:56the agent runs, it would overwrite this
2:53:59file with information like here's the
2:54:02last run, here's how long it lasted,
2:54:03here's what happened, here's what I did,
2:54:05here are known issues, and things like
2:54:06that. And this is better than an append
2:54:08log in some scenarios because like I
2:54:09said earlier, if you run an automation a
2:54:12thousand times, then you might have a
2:54:13thousand append logs. But then what's
2:54:15cool is the actual structure of your
2:54:16prompt. So when you're creating a new um
2:54:19task right here, the way that you prompt
2:54:20it, you could basically say, okay,
2:54:22before you actually do your job, go
2:54:24ahead and read this file, which is
2:54:25basically the last run, and then you do
2:54:27your main task. And then after you're
2:54:28done with the task, overwrite that file
2:54:30with any current issues or status or
2:54:33anything that you found that might help
2:54:34the next agent. So, obviously I'm going
2:54:36to be playing around with different
2:54:37structures of having context shared
2:54:39between these different, you know,
2:54:40tasks, but that's just something that I
2:54:42thought would be really, really cool.
2:54:43So, another thing I wanted to talk about
2:54:44is because this is in the desktop app
2:54:46and because this is running on a
2:54:47schedule, how do you want to get
2:54:49notified that that scheduled task has
2:54:50been done? So, the desktop app does have
2:54:53notifications, but they're not super
2:54:54great. At least it wasn't making any
2:54:56noise and it wasn't capturing my
2:54:57attention. It's cool that everything
2:54:58gets organized over here. So, as you
2:55:00have more tasks, you can see them all
2:55:01and it's organized. But what you can do
2:55:03is you can add a hook. so that every
2:55:05time you actually get a sound. So what I
2:55:08did is I set up a hook so that every
2:55:09time a cloud code session finishes I get
2:55:11a notification. So listen to this.
2:55:15So that was like kind of the default
2:55:17Windows little sound. You can change
2:55:19that. But that's really helpful because
2:55:20I could be working and I could forget
2:55:21that a scheduled task might go off and
2:55:23then I get that noise. But the other
2:55:24thing I would recommend is in the prompt
2:55:26of the actual skill itself, maybe just
2:55:28at the bottom say, "Hey, once this is
2:55:30done, just shoot me a ClickUp message
2:55:32and say that this happened." And that's
2:55:33probably what I'm going to set up for
2:55:34all my scheduled tasks. And if you're
2:55:36curious about setting up hooks,
2:55:37literally just say to cloud code, set up
2:55:39a hook. I want you to play a sound every
2:55:40time you finish talking to me and it
2:55:42will do it in like a minute. And the
2:55:44last thing I was curious about is the
2:55:46fact that we're limited to the desktop
2:55:48app. Now, Enthropic is shipping like
2:55:50crazy. So, I'm sure in a week or two
2:55:52that this is going to be open in the
2:55:54terminal and in the IDE extensions and
2:55:56stuff like that, too. But for the
2:55:57moment, it's only desktop app. And
2:55:59basically why is because all of the
2:56:01actual like cron logic and that kind of
2:56:04metadata lives in the desktop app even
2:56:06though the actual files live in your
2:56:08computer. They live somewhere where
2:56:10terminal or your VS code cloud code
2:56:13could actually see it. So I was
2:56:14interested in that and I said are you
2:56:16able to see my scheduled task for
2:56:17morning coffee? It runs at 609 and
2:56:19here's where it lives because it lives
2:56:20kind of in this global folder path right
2:56:23here and it can see it right because it
2:56:25just exists as a file. But what it can't
2:56:27do is it can't create new scheduled
2:56:30tasks because it can't actually touch
2:56:31the cron that the desktop app of cloud
2:56:34code sets up. But what it could do is it
2:56:36could edit it. So it could improve it.
2:56:38It could make changes, but it can't
2:56:40create them or like run them. Now that's
2:56:42actually not a huge deal to me right now
2:56:44because I usually work in VS Code, but I
2:56:45can just have the desktop app open and I
2:56:47can just leave it open in the background
2:56:49while I've got my computer on and all of
2:56:51my scheduled tasks will still be
2:56:52running. And so now that Cloud Code is
2:56:54so powerful on its own, it can actually
2:56:55like do things in the browser as well. I
2:56:58truly think we're getting to that point
2:56:59where you can automate anything.
2:57:03Cloud Code can now remind you to do
2:57:05things, check on things proactively for
2:57:07you, and work for days straight without
2:57:08you ever touching it or needing to give
2:57:10any input. So here you can see I just
2:57:12said, "Remind me at 10:23 a.m. to check
2:57:14on my project." It goes ahead and uses a
2:57:16cron create tool to set this reminder.
2:57:20There we go. 10:23 just hit. I didn't
2:57:22touch it. and it just said, "Hey, Nate,
2:57:23this is the reminder to check on your
2:57:25project." So, just shot off this one
2:57:26that says, "Every 10 minutes, check my
2:57:28ClickUp to see if there's any new
2:57:29developments on our project." It's using
2:57:31the loop skill, as you can see, which is
2:57:32a new built-in skill, and it creates a
2:57:34cron for every single 10 minutes. And
2:57:36now, this would run for the next 3 days,
2:57:38every 10 minutes until I told it not to.
2:57:40And this doesn't have to be every 10
2:57:41minutes. It could be every hour. It
2:57:43could be every 5 minutes. It could be
2:57:44whatever interval that you want. And
2:57:46this is all thanks to the newly released
2:57:47feature or skill loop, which is a
2:57:50powerful new way to schedule recurring
2:57:51tasks for up to 3 days at a time. And
2:57:53this is so funny because less than 12
2:57:55hours before this was announced, the
2:57:57scheduled tasks in Cloud Code was also
2:57:59announced. So, right off the bat, those
2:58:00two features might seem like they're the
2:58:02exact same thing, but they're actually
2:58:03super different in how they work, and
2:58:04they have different use cases. So, in
2:58:06today's video, I'm going to break all of
2:58:07that down and tell you everything that
2:58:08you need to know about it. And by the
2:58:10way, if you haven't watched my new
2:58:11scheduled tasks video, then check that
2:58:12out right up here, and then hop back
2:58:14over to this one. All right, so as you
2:58:15guys just saw in the quick demo, we now
2:58:17have the ability to use loops, which
2:58:19means that we could say something like
2:58:20/loop every 5 minutes, check on the
2:58:22deploy, or we could just say that in
2:58:23natural language, which is awesome
2:58:25because it invokes the loop skill and
2:58:26then it creates that cron job right here
2:58:28in cloud code. And you'll notice that
2:58:30this is in my VS code. So this is
2:58:32available in your terminal, in cloud
2:58:33code desktop app, in VS Code extensions,
2:58:36wherever. This is just a core part of
2:58:38cloud code now. So if you're not seeing
2:58:39this, just make sure you update your
2:58:41extension or you update Cloud Code. And
2:58:43this lets you set up loop intervals or
2:58:44reminders. So reminders, like you saw
2:58:46that first demo, I just said, "Hey, at
2:58:47this time just tell me this." And in
2:58:49that session, it will bump up a message
2:58:51without you triggering it. Or you could
2:58:52have them be intervals. So you could say
2:58:54every 2 hours. You could say every 30
2:58:56minutes. Whatever you want that actual
2:58:57interval to be. And what's cool about it
2:58:59is it does it all in the same session.
2:59:01So if I leave this session up every 10
2:59:03minutes, it would check everything right
2:59:04here, which means that it's able to
2:59:06continuously read through what happened
2:59:07in the past one, and it continuously
2:59:09sees what we're doing. Now, obviously
2:59:10there are some pros and cons there, but
2:59:12just wanted to point that out. The major
2:59:14con there being your context, making
2:59:15sure that if something does go off every
2:59:1610 minutes, you're not going to get a
2:59:18huge report and then every 10 minutes
2:59:19you just more tokens, more tokens, and
2:59:21then context rot. It's basically
2:59:24scheduling a prompt that you would be
2:59:25sending in here and then firing off,
2:59:28which means you can loop skills. So, if
2:59:30you want every 20 minutes, for example,
2:59:32run a skill called review PR, you could
2:59:34tell it to every 20 minutes run the
2:59:36skill. It would run it, it would wait 20
2:59:38minutes, and then it would do it again.
2:59:39And of course, you could use actual
2:59:41slash commands to invoke both the loop
2:59:43and the skill. Or you could just say,
2:59:44"Every 20 minutes, run my review PR
2:59:46skill." And of course, the onetime
2:59:48reminder feature. So at 3 p.m. or in 45
2:59:50minutes, remind me to do this or check
2:59:52in on that. And Claude will basically
2:59:54pin that time. It'll create that cron
2:59:56and then once it's done, it'll just
2:59:57delete itself. So whether that's, hey,
2:59:59at 4:30, remind me I have to go do this
3:00:01or every hour remind me to just stand up
3:00:03and like look away from my screen for 5
3:00:05minutes, it can do that. All right. All
3:00:07right. So, there's a couple things that
3:00:08I wanted you guys to understand about
3:00:09how this actually works. So, let's just
3:00:10play around a little bit. Hey, at 10:40
3:00:13a.m., can you please remind me to take
3:00:14out the garbage? Cool. So, what that's
3:00:17going to do is it's going to use the
3:00:18cronreate tool, and it's going to create
3:00:21that basically schedule to remind me
3:00:23take out the garbage. And what you can
3:00:24see here is the actual prompt. So, at
3:00:26this interval, which is just how cron
3:00:28works, it's basically going to shoot a
3:00:30prompt into this window that says remind
3:00:32Nate to take out the garbage. You can
3:00:34see the recurring equals false. Now, of
3:00:36course, the key is if the session is
3:00:38closed, then that cron is going to
3:00:40automatically be killed. So, now
3:00:42something interesting. I'm going to open
3:00:43up a new session and I'm going to say,
3:00:46"Hey, every hour, can you just remind me
3:00:47I need to stretch my neck?" And I'm
3:00:49going to shoot this one off. And we'll
3:00:50see how this one is a little bit
3:00:52different because this once again
3:00:53creates a cron. We have a prompt. And
3:00:56you can see in this one, we don't have
3:00:57the recurring equals false. We just know
3:00:59that this cron is going to go every
3:01:01hour. But these loop jobs or task jobs
3:01:03are per session. So these two tabs are
3:01:05two different sessions. So if I came
3:01:07into this session and said, "Can you
3:01:08please tell me all of the scheduled loop
3:01:10tasks that we have today? It's going to
3:01:13use a tool called cron list, and it only
3:01:15can see the 10:40 a.m. take out the
3:01:17garbage. It cannot see the task that
3:01:19exists in this session because they're
3:01:20independent and they're separate. Now,
3:01:23one interesting thing to notice is that
3:01:24this session didn't actually invoke the
3:01:26loop tool. The loop tool basically tells
3:01:28it how to set up cron jobs and how to
3:01:30use the cron create. So if you don't see
3:01:32loop, don't worry. It's still actually
3:01:34doing this in a loop. It's just kind of
3:01:35about the actual wording. So, if I was
3:01:37to open a new one, let's see if I
3:01:38actually call the loop tool right here.
3:01:40So, I do loop and then I just say, you
3:01:42know, um, check my ClickUp. This one is
3:01:46going to default, I believe, to 10
3:01:48minutes if you don't specify a time. And
3:01:50it might invoke the loop skill because
3:01:51we actually called it to, but looks like
3:01:53it didn't because it knows exactly what
3:01:55it needs to do already. So, the point
3:01:57being, all it matters is that the cron
3:01:59is being created. It doesn't always
3:02:00matter if it invokes the loop skill or
3:02:02not. And then if you wanted to cancel
3:02:04one of these jobs, all you'd have to do
3:02:05is either close out of the terminal or
3:02:06just say, "Actually, I don't need this
3:02:08anymore. Go ahead and cancel it." And
3:02:10that one invokes a different cron skill
3:02:12called cron delete. It shoots over the
3:02:14job ID. And now that it's canceled. And
3:02:16one final thing to keep in mind is in VS
3:02:18Code, if you close out of a tab and then
3:02:20you just open up that conversation
3:02:21again, that still will kill those crons.
3:02:24So you guys just saw how pretty much all
3:02:25of this worked. We have cron create to
3:02:28schedule. We have cron list to list
3:02:30them. And then we have cron delete to
3:02:31cancel them. And all of those can be
3:02:33invoked with natural language, which is
3:02:34awesome. So now let's get into some of
3:02:36the limitations and then I'm going to
3:02:37compare them to the actual scheduled
3:02:39tasks feature. So the first big one is
3:02:41that we have a 3-day loop expiry, which
3:02:43is just basically for safety. It auto
3:02:44cleans things up if you forgot you had
3:02:46all of these loops running. So once you
3:02:48create a loop, it basically has a 3-day
3:02:50timer on it. It can run for day one, it
3:02:51can run on day two, and then on day
3:02:53three, it can run up until that last
3:02:55fire, and then it will autodee. And if
3:02:58you want anything longer than 3 days,
3:02:59then you would either just recreate that
3:03:00loop or that probably indicates that you
3:03:02should just turn this into a legitimate
3:03:04scheduled task. Now, the other thing
3:03:06that you can do is if you want to
3:03:07completely disable scheduling, so maybe
3:03:09in your natural prompting it's
3:03:10accidentally creating all these crrons,
3:03:12you could go into your environment
3:03:14variables and just disable that and it
3:03:16would probably be able to help you
3:03:17figure that out. So the other things
3:03:19here are that if you close the terminal,
3:03:20your tasks are gone. It doesn't have
3:03:22catchup. So the scheduled tasks, if you,
3:03:25you know, opened up your desktop app and
3:03:27you missed a bunch, it would catch up
3:03:28automatically. This doesn't do that. And
3:03:30there's no persistence, meaning after
3:03:31your 3 days and you wanted to do that
3:03:33same loop again. It would be a fresh
3:03:35session. But obviously, there's tons of
3:03:37things you can do here with context
3:03:38management and reading different files
3:03:40in order to kind of Frankenstein your
3:03:42own fix there. So now that all the
3:03:44features have been explained and you've
3:03:45seen a demo, I think that probably you
3:03:47understand the difference between the
3:03:48loop and the schedule tasks a bit better
3:03:50now. But let's just go over some of the
3:03:52key highlights. The loop has your 3-day
3:03:54expiry. It's all done within one session
3:03:56and there's no catch-up. It's basically
3:03:58a help me now or help me on this project
3:04:00for today type of function. The schedule
3:04:02tasks are dis stored. They're longived.
3:04:04They have catchup and these are like
3:04:06daily, weekly, monthly functions that
3:04:08can run indefinitely. Of course, with
3:04:10both of these though, the terminal or
3:04:11you know the app has to be open and this
3:04:14one is only currently available in the
3:04:16desktop app. But I can imagine with how
3:04:18fast Enthropic is shipping things, maybe
3:04:20by the time you watch this video,
3:04:21scheduled tasks are already out for the
3:04:24terminal and extensions as well, the way
3:04:26that loop is available in cloud code
3:04:28everywhere. So basically, it's one
3:04:30simple question. Do you need help right
3:04:32now on a project or do you need help
3:04:33with something every day or every week?
3:04:35And that's how you decide if you use the
3:04:37new loop feature or if you use scheduled
3:04:39tasks. So I thought I'd end off real
3:04:41quick by giving a few maybe practical
3:04:43ways that you could actually use loop
3:04:44rather than something scheduled. So
3:04:46maybe all day you're waiting on a very
3:04:48urgent email. Just set up Cloud Code to
3:04:50check in on that email every 5 minutes
3:04:52and if it's there, it can automatically
3:04:54let you know. Maybe you're working on a
3:04:55deploy and you want to just pull that
3:04:57and check every, you know, hour or so if
3:04:59everything's working okay. Maybe you've
3:05:00got a deadline due at the end of the
3:05:02week and you need a 3-day sprint to be
3:05:03constantly checking in on the team and
3:05:05checking in on progress. Maybe you're
3:05:06testing and iterating. Maybe you're
3:05:07watching logs. Maybe you're tracking
3:05:09changes. There's so many different use
3:05:10cases here. There's so many different
3:05:12ways to use the loop to prompt an agent
3:05:14to have different files, to use
3:05:15different skills, and it's really,
3:05:16really cool the way that you could
3:05:18potentially set these things up. Now
3:05:19that we've built our first couple
3:05:20workflows, we understand the components.
3:05:23Cloud code is insanely smart. We use it
3:05:25in plan mode, and it helps us build
3:05:26automations. And then we take those
3:05:28code-based automations and we put them
3:05:29in something like modal or trigger.dev,
3:05:31and they run 24/7. We understand
3:05:33workflows, we understand tools, and we
3:05:35understand claw.md. But now, we're going
3:05:36to get a little bit more advanced. So,
3:05:38we're going to dive a little bit deeper
3:05:39into some best practices around your
3:05:40claw.md. We're also going to discuss
3:05:42project architecture and some other
3:05:44helpful built-in commands. So, let's get
3:05:46into it.
3:05:49Okay, unfortunately, we're back to the
3:05:51boring slides, but need you guys to pay
3:05:53attention here. So, more concepts. This
3:05:55one's going to be short. We're going to
3:05:56talk about more claw.mmd tips. We're
3:05:58going to talk about essential/comands,
3:06:00and we're going to talk about project
3:06:02folder architecture. This one is really
3:06:04important to pay attention to. This is
3:06:05probably one of the things that caused
3:06:07me the most confusion when I first
3:06:09started. So, we're going to make it
3:06:10clear.
3:06:12So, you've seen the cloudm a good one we
3:06:15know contains a project overview, text
3:06:17stack, architecture overview, coding
3:06:19conventions, common commands,
3:06:20constraints, and where it can find more
3:06:23context.
3:06:24Now, let's talk about some best
3:06:26practices.
3:06:28You might see different things online. I
3:06:30like to try to keep mine under 200
3:06:31lines. And I found that by effectively
3:06:33using, you know, routing rules just
3:06:35basically means pointing to different
3:06:36files and using um other compression
3:06:40techniques,
3:06:41keeping it under 200 100 lines is doable
3:06:43and it's great. So shorter equals less
3:06:46tokens. Be very specific. So saying
3:06:49something like use two space annotation
3:06:51is much better than just saying format
3:06:52it nicely. Treat this as a living
3:06:55document. So just because you made your
3:06:58cloudd file does not mean you're done. I
3:07:00update my cloud.MD MD file probably
3:07:02every single day if not yeah probably
3:07:04every single day I was going to say
3:07:05every hour but that would be a hyperbole
3:07:07every single day I'm updating my clawmd
3:07:10now you can also use in it we talked
3:07:12about that and you can also use rules
3:07:15for putting certain things in which
3:07:19means let's say you have a specific rule
3:07:21about the way that you like to write
3:07:22emails or the way that you like to
3:07:23handle internal comms there's no point
3:07:26to put that into the claude MD because
3:07:28does claude need to know that every
3:07:30single time you talk to it. No, it needs
3:07:32it on occasion. And so what you can say
3:07:33is, hey, when you need to understand
3:07:36rules about the way I speak, go to the
3:07:38rules folder and you can find the rules
3:07:40there. Right? So that's a little trick
3:07:42we call routing. So it's really
3:07:44important to understand how you can
3:07:46route as much as possible. The cloudmd
3:07:48is not a know all
3:07:52file. It is a I know where everything I
3:07:54need to find lives file. It's basically
3:07:56your table of contents. Think of it like
3:07:57that.
3:07:59So
3:08:02no. Okay. I I thought I had another
3:08:04thought, but we're going to move on. So
3:08:06this is the example I showed you guys
3:08:07earlier, right? So this is my executive
3:08:09assistant. This is the beginning of the
3:08:10cloud.MD. And you can see this is an
3:08:12example of routing right here, right?
3:08:14Because if I jammed everything about me
3:08:15in this file, everything about my work
3:08:17or about my team or about my current
3:08:18priorities in this file, it'd be huge.
3:08:21But I can say, hey, if you need any of
3:08:23this stuff, you know where to find it.
3:08:24And it's really, really effective.
3:08:27So
3:08:29you're in VS Code, you're in cloud code
3:08:31and you have folders on the lefth hand
3:08:34side, right? Like maybe you've got your
3:08:35workflows folder or maybe you've got
3:08:36your, you know, brand assets, whatever
3:08:38it is. What you want to do is you want
3:08:41to make sure your projects have a folder
3:08:42called
3:08:45dotcloud is basically like that project
3:08:47directory, the project settings. And
3:08:50there's a couple different types of
3:08:51dotcloud. Now, for now, I'm not going to
3:08:53touch too much on the system level path
3:08:55because this is usually if you're set,
3:08:56you know, within an organization, but me
3:08:59locally working with my team and stuff,
3:09:01I either have a global cloud folder,
3:09:04which is not the one you see in VS Code,
3:09:05that lives somewhere in your home
3:09:07directory on your computer, or you have
3:09:10your project cla
3:09:14that you're looking at in VS Code. So,
3:09:16this is local. This is based on your
3:09:17project. This one is based on every
3:09:19cloud code project you ever work on
3:09:21ever. So, for example, let's say you
3:09:23have a setting in your in in your um
3:09:25cloud code, right? So, I don't want to
3:09:27get too ahead of myself here, but
3:09:31um I've got basically a setting that
3:09:33always allows a certain MCP server,
3:09:35right? Or a certain front-end skill,
3:09:37let's just call it that. Now, if I'm
3:09:40using that same configuration, if every
3:09:42single project that I want to work on
3:09:43ever, I want to allow that same server,
3:09:46then it would make no sense for me to
3:09:47put it in every single project when I
3:09:49could just put it globally. Okay, so
3:09:51that's just kind of the the idea.
3:09:52There's a difference between global and
3:09:55project level.
3:09:57So,
3:09:59we have automemory, which means that we
3:10:00have cloud.mmd, which Claude always
3:10:02looks at, but Claude also has auto
3:10:04memory, so things get persisted across
3:10:06sessions, which is really, really cool.
3:10:07So if you tell Claude always use, you
3:10:10know, PNPM, not npm, claude will save
3:10:12that to a global claude folder, and you
3:10:16know it's global because you have this
3:10:17little tilda in front of it, right? So
3:10:19back over here, you can see this was
3:10:20just a dot /.claude, but this global one
3:10:23is a tilda. So whenever you see the
3:10:24tilda, it's a global setting. So now
3:10:29this memory.mmd file that's global
3:10:30across every project, Claude is able to
3:10:33look at. And you can also edit that
3:10:35automemory file anytime. So the same way
3:10:37in chatbt on the browser or in claude on
3:10:39the browser you have persistent memory
3:10:41it's basically the same thing. So it's
3:10:42really cool.
3:10:44Now here are a few slash commands to be
3:10:46aware of for session management. So we
3:10:48obviously talked about slashinit. We've
3:10:49talked about slashclear. It wipes the
3:10:51conversation. We've talked about slash
3:10:52compact a little bit but what's cool
3:10:54about slash compact. So like let's say
3:10:55you're at 60% context window. Usually at
3:10:5960 I like to compact. Now, you can just
3:11:02do compact and it will say, "Okay,
3:11:04here's the conversation. Here are the
3:11:06five most important things. I'm going to
3:11:08pull those out, get rid of everything
3:11:09else, and put those five most important
3:11:11things back in the conversation." So,
3:11:12now we are back at like maybe 10% of the
3:11:15context, and we still have all of the
3:11:16important stuff we need.
3:11:19Now, what you can do is instead of just
3:11:20saying /compact, you can say /compact
3:11:23keep the information about the website
3:11:24design, and it can get rid of everything
3:11:26else. So, you can be specific about what
3:11:28it compacts, which is pretty cool. We
3:11:30have slashre, which honestly I don't use
3:11:32that much, but it's nice that it's
3:11:34there. It's basically an undo button.
3:11:35And then we have /resume, which means
3:11:37you could resume a session that you were
3:11:38working on, you know, a couple days ago
3:11:40or a couple weeks ago.
3:11:43We have information and diagnostics
3:11:45commands. So, we've got /context we've
3:11:47talked about. We've got /cost, which
3:11:49shows the token usage and cost. We have
3:11:51slashmodel to change model. We have
3:11:53slashhelp to look at all the commands.
3:11:54We have /d doctor which can run
3:11:56diagnostics to see your installation and
3:11:58see if everything's working. And we have
3:12:00slash status which will show you the
3:12:02version of cloud code and your model and
3:12:04your account and things like that. Now
3:12:06the cool thing about this is you don't
3:12:07have to memorize these. You can
3:12:08literally say hey I want to look at
3:12:09this. Do you have any commands? Or hey I
3:12:11want to do this and it might just invoke
3:12:12them automatically. But just kind of
3:12:14good to know right.
3:12:16So configuration we've got our SL
3:12:18mememory which I talked about to auto
3:12:19auto or to edit that automemory. We have
3:12:21slashconfig. We have SL permissions, we
3:12:24have SLMCP, and we have SL aents. All of
3:12:28these are going to get touched on later
3:12:29in the course as well. So you don't have
3:12:30to memorize these. Once again, just
3:12:31trying to get you familiar.
3:12:33So once again, we have the project
3:12:35folder architecture idea because this
3:12:38goes beyond just our cloud MD. This goes
3:12:40into skills, agents, MCP servers,
3:12:43settings, everything. So we have once
3:12:45again user level, global, we have the
3:12:48tilda, and this is across all projects.
3:12:50We have project level settings which is
3:12:52in yourcloud you have a settings.json
3:12:56which means anyone that accesses this
3:12:57project can see that setting file and
3:13:00configure it or we have our local
3:13:02project which is settings.local.json
3:13:05and that is just you. Okay.
3:13:08So what goes where settings.json it goes
3:13:12in your personal defaults claude
3:13:14settings.json goes in your team
3:13:17standards and yourcloud/s settings.local
3:13:20local is basically the overrides for the
3:13:22specific project. So it goes in this
3:13:24order. Let's say you want to do
3:13:26something or you want cloud code to do
3:13:28something. It will first check your
3:13:30local settings
3:13:32and then it will check your project
3:13:34settings and then it will check your
3:13:35global. So this is basically the
3:13:37hierarchy. So like let's say there's a
3:13:40certain command called read, right? So
3:13:42let's say it wants to read a file. If
3:13:44local says do not ever use read, it will
3:13:46instantly stop. But if um local says yes
3:13:50you can and project says yes you can and
3:13:54then global says no you can't it'll stop
3:13:56there. So that's just kind of like the
3:13:57order of operations the hierarchy.
3:14:00So this is what it looks like. This is a
3:14:04project level directory. So let's say
3:14:06you open up VS Code and you have a
3:14:07project called um websites. That's what
3:14:11you'll see up top. Then below that
3:14:13you'll see your cloud. And when you open
3:14:15up yourcloud, that's when you might see
3:14:17your settings.json, your
3:14:19settings.local.json,
3:14:20your cloud.md, your rules, your skills,
3:14:24your agents, your commands. Oops. And
3:14:27then outside of that folder, you might
3:14:29see your MCP config or you might see
3:14:31your cloud. Mmd. And so those are just
3:14:34kind of like the different drill downs
3:14:35that you'll be able to see.
3:14:38Now, this one is a global one. So it
3:14:40looks very similar, right? We've got
3:14:41cloud, we've got settings, claude,
3:14:42agent, skills, rules, whatever. But this
3:14:45is the global one. So this doesn't exist
3:14:47in a particular project. It exists on
3:14:49your cloud code configuration.
3:14:53So what is the purpose of all this?
3:14:56Don't want to hammer this home, but I
3:14:57thought that this little, you know, I
3:15:00don't want to over beat a dead horse
3:15:01here, but this is a breakdown to look at
3:15:04as far as like the file, the purpose,
3:15:06and if it's shared or not. So hopefully
3:15:08this is all
3:15:10at a high level making sense. I
3:15:13understand though looking at this it
3:15:15might look completely foreign and you're
3:15:17like what in the world? I promise you
3:15:18that's normal. All you have to do is get
3:15:20into there and start building and all of
3:15:22these videos are going to walk you
3:15:23through it and it will start to make
3:15:24more sense. It's just so much more
3:15:26helpful to understand what you're
3:15:28looking at first. At least in my
3:15:29opinion. Okay. So the dotclaw directory.
3:15:34What is getit ignore? because you're
3:15:37probably going to see that when you run
3:15:39/init or when you start to have, you
3:15:41know, projects being synced to GitHub.
3:15:43You're going to see something called.git
3:15:44ignore. This is basically just a system
3:15:47that tracks your codebase changes,
3:15:49right? And ignore is literally just a
3:15:52file that tells cloud code, don't ever
3:15:55push any of the files in here to GitHub.
3:15:58So in your dockit ignore, maybe you have
3:16:00some pictures of yourself that you don't
3:16:01want to get out there, or maybe you have
3:16:03some passwords or API keys. All you'd
3:16:05have to do is in the docket ignore just
3:16:07put the name of those folders or files
3:16:08and then they'll automatically be
3:16:10excluded from GitHub or from git.
3:16:14Um
3:16:16that's basically it, right? There's also
3:16:17going to be one called.git keep. So same
3:16:20thing.
3:16:22So here is an example of my executive
3:16:25assistant project, right? You can see
3:16:26that we've got some stuff going on here.
3:16:28We've got the cloud folder. We've got
3:16:30archives brand assets. We've got
3:16:32projects which is currently open. And
3:16:34you can see that there's a lot of
3:16:35folders and a lot of files in here.
3:16:37You'll also see down here that I have a
3:16:38claw.md. I also have a claw.local.md.
3:16:42I have agit ignore. And then you also
3:16:44see that some things are grayed out like
3:16:46the env or the Google ooth json.
3:16:49Now anything that exists in the ignore
3:16:53will be grayed out. And that's just a
3:16:55visual way of saying this is not going
3:16:57to be put in GitHub, which is great.
3:16:59Anything that is green is a new file
3:17:02that git has never seen before and it
3:17:04hasn't been committed anywhere. So I
3:17:06love this because it's visual. I can
3:17:08instantly see, oh, do I need to make a
3:17:09push? Do I need to save my work?
3:17:11Essentially, anything that's yellow
3:17:13means that it's a existing file in git,
3:17:15but there was some sort of update. So
3:17:17you have to make a new commit. And
3:17:19that's really nice because as everything
3:17:20gets more yellow and gets more green,
3:17:22push it to GitHub, it'll all go white.
3:17:24And then you know you're good.
3:17:27So,
3:17:29that is a lot more of the foundational
3:17:32concepts that you really need to know.
3:17:33We're going to get back into building.
3:17:34We're going to do some more stuff here.
3:17:36But hopefully this session was helpful.
3:17:38But I'll see you guys over there.
3:17:42Now, we're going to jump into a quick
3:17:43segment about Rag. And in this case,
3:17:45we're using Google's new embeddings
3:17:47model that just came out. And it's
3:17:48really cool because it lets us really
3:17:50easily embed videos, images, and audio
3:17:53as well as our text. So, like I said, it
3:17:54just makes it super super easy. So,
3:17:56check this out.
3:18:00Google just dropped Gemini Embedding 2,
3:18:02which is their very first natively
3:18:04multimodal embedding model, and it is
3:18:06already blowing my mind. This means that
3:18:07you can have completely multimodal
3:18:09databases with text, images, videos,
3:18:12audio, and documents. And it can
3:18:14actually understand the nuanced
3:18:15relationships between these different
3:18:16types of media so that you can have
3:18:18actual realworld answers back. And
3:18:20here's a quick look at some of the
3:18:21benchmarks, which I always think are
3:18:22important to look at, but I think it's
3:18:24always worth taking it with a grain of
3:18:25salt. And that's why in today's video,
3:18:27I'm going to show you a few examples
3:18:28that I already built out that are super
3:18:29practical. And then I'm going to show
3:18:31you exactly how you can set this up for
3:18:32yourself. And trust me, it is so much
3:18:34easier than you probably think. So, let
3:18:36me show you some examples and then I'll
3:18:37teach you how to do this yourself. So,
3:18:38right here, you can see that I've got a
3:18:40project called manual, which basically
3:18:41stands for like instruction manual. So,
3:18:42what I did is I dropped in this PDF
3:18:44right here, which is a 68page PDF about
3:18:47how to use this vacuum cleaner. You can
3:18:49see that it's pretty complex. It's got
3:18:51tons of different text. It's got tons of
3:18:52diagrams. It's got images. And if you
3:18:55wanted to be able to chat with it, it
3:18:56would be pretty complicated to build
3:18:57this ingestion pipeline if you use
3:18:59something like Nitn because you'd have
3:19:00to figure out exactly how you want to
3:19:02chunk it and how to capture the images
3:19:03and how to store those and how to pull
3:19:05them back. But I kid you not, I dropped
3:19:07in the PDF right here and I said, "Hey,
3:19:09Cloud Code, there's the PDF. I want to
3:19:11be able to chat with this using Google's
3:19:12new embeddings model. Just go build it
3:19:14for me." And not only did it build it
3:19:16for me, but it built this app where I
3:19:18can actually talk to it. So let's say I
3:19:19ask you know how do I clean the filter?
3:19:21It's searching right now our Pine Cone
3:19:23database and in the database we're
3:19:24storing both text and images. So here
3:19:26you can see it says to clean the filter
3:19:28follow these steps number one number two
3:19:30blah blah blah. And then down here we
3:19:31have actual images. So if I click on
3:19:33this one we can see the actual diagram
3:19:35that it pulled from because sometimes
3:19:37when you're trying to troubleshoot
3:19:38things especially if it's physical an
3:19:39image is way more valuable than text.
3:19:41And what you can see here is that it
3:19:42also returned the same diagram in
3:19:44different languages but you could turn
3:19:46that off if you didn't want to. And
3:19:47what's super cool is at the end I can
3:19:49actually expand the sources and it shows
3:19:51me the different pages that it looked at
3:19:52and the confidence score or the percent
3:19:54match that it had for that page. Let's
3:19:57go ahead and try one more for this demo
3:19:58which is just a very broad what are the
3:20:00parts and I'm assuming there's lots of
3:20:02different pages that it might need to
3:20:03figure out what the parts are. So what
3:20:05we got here are we have the main
3:20:06components from page six. We have what's
3:20:08included on page seven and then we have
3:20:10available accessories. So that's super
3:20:12good. And it looks like we got three
3:20:13different images. So we have what's in
3:20:15the box. We have the actual getting to
3:20:17know your Hoover Impulse cordless
3:20:19vacuum. So, all the other different kind
3:20:20of components here. And then the final
3:20:22image is how to order extra accessories.
3:20:25So, that's just super cool. Okay, so
3:20:27that was our instruction manual example.
3:20:28I dropped in one PDF and it basically
3:20:31was able to turn that into text and
3:20:33images in our database and pull
3:20:34everything back accurately. So, then
3:20:36let's scale it up a little bit. I am
3:20:37doing a roofing example. So, in this
3:20:39one, I gave it 13 images and all these
3:20:42images are different roofs that might
3:20:43have some sort of issue. So, let's say
3:20:45you're a roofing company and you help
3:20:46fix roofs. What might be helpful is if
3:20:48you had an app where you internally or a
3:20:50client could upload a picture of their
3:20:51roof and you could get like a quote or
3:20:54an internal brief about any past work
3:20:55that you've done on a roof that looks
3:20:57like that. So, if I drag in a picture
3:20:58right here, it shoots it off and says,
3:21:00"Find similar past projects for this
3:21:02roof." It's searching the database. It's
3:21:04looking through all of our different
3:21:05past projects, and all of those images
3:21:07have metadata like how much this costed
3:21:09us or, you know, how long it took, how
3:21:10many team members. So, here are the five
3:21:12similar projects. We get a percent match
3:21:15for each of them as you can see. And
3:21:16then we get a description like quote
3:21:18range and averages, team size, trend,
3:21:20roof types, breakdown. And so,
3:21:22obviously, I'm not a roofing expert. If
3:21:23you had some subject matter expertise
3:21:25about roofs that you could add into
3:21:27here, this would obviously be better,
3:21:28and you would have your own data. But,
3:21:29it's just really cool that you can get a
3:21:31quick search across potentially hundreds
3:21:33and hundreds of projects to do this. And
3:21:36I could ask a follow-up. So, let's say I
3:21:37said, "Okay, awesome. Can you tell me
3:21:39about the one that we did in Richmond,
3:21:40Virginia? It looks pretty similar. And
3:21:42at this point, it could pull the
3:21:43metadata from this image and it could go
3:21:45grab other pictures from that file if we
3:21:47had them. But anyways, we get the basic
3:21:49info, the scope, what stands out,
3:21:50pricing, context. Super super awesome.
3:21:53But yeah, clearly this needs some
3:21:55subject matter expertise. It obviously
3:21:56made up all this data because I feel
3:21:58like this roof would have costed more to
3:22:00fix than this roof. So, so if you've
3:22:03never built a pipeline like this before,
3:22:04then it might not seem super impressive
3:22:06because that type of functionality is
3:22:08pretty standard on a lot of chatbot
3:22:10based features. But the fact that I
3:22:11built both of those demos in less than
3:22:1330 minutes is what truly blows my mind
3:22:16because that would have taken me several
3:22:18hours, if not several days, to build out
3:22:19an NAN. And that's why I had to show you
3:22:22guys this stuff. All right, so we're
3:22:23going to hop into the live build. But
3:22:24real quick, in case you haven't really
3:22:26heard of like Rag or why this multimodal
3:22:28stuff is awesome, let me explain it real
3:22:30quick. So rag stands for retrieval
3:22:32augmented generation and it basically is
3:22:34just the concept of your AI agent only
3:22:36knows so much in its training data. So
3:22:38if you ask it a question and it doesn't
3:22:39have that information, it has to go grab
3:22:41it in order to generate a better answer.
3:22:44So it basically retrieves information,
3:22:46it augments its answer because it has
3:22:48more data and then it generates an
3:22:49answer or generates a response to you.
3:22:51Now typically when we think of rag and
3:22:52we think of a vector database type of
3:22:54rag, we have to look at it like this. We
3:22:57have some sort of data source, right?
3:22:58whether that's document or video or an
3:23:01image. And what happens is we have to
3:23:02turn this document into vector points or
3:23:05little chunks. So for example, if this
3:23:07was a document about our company, then
3:23:09maybe we'd split it up into three
3:23:10chunks. Those chunks would run through
3:23:12an embeddings model, which is, you know,
3:23:13Google's new embeddings model 2 that
3:23:15we're talking about today. And then it
3:23:16would spit out these vector points,
3:23:18which would basically just be a
3:23:19numerical representation of what the
3:23:21data means. So this chunk might be
3:23:23placed over here because it's company
3:23:24overview information. This chunk might
3:23:26be placed over here with financial
3:23:27information. And this chunk might be
3:23:29placed over here with marketing
3:23:30information. And just to help you guys
3:23:32contextualize that, when I was first
3:23:34testing this out, I did a demo where I
3:23:35dropped in an Adam Sandler and me
3:23:37picture that was Nano Banana. Um, a
3:23:39random picture of me, a video of me
3:23:42using Claude Code, a video of a dog
3:23:43playing guitar, a video of me speaking,
3:23:46um, a couple text files, and a couple
3:23:47more images that were just literally so
3:23:49random. I put a picture of smiley face
3:23:50potato fries in here. And what happened
3:23:52after it embedded all of those is it
3:23:54gave me this report which is basically
3:23:56the multimodal embeddings but this is a
3:23:572D view rather than a 3D view. But you
3:24:00can see that it's placing things where
3:24:01it deems appropriate. So up here we have
3:24:03you know first aentic workflow which is
3:24:05in the category tech and it is a text
3:24:07file. We've got over here a dog playing
3:24:09guitar which is in the category
3:24:11entertainment and the modality is video.
3:24:13We've got the smiley face rise which is
3:24:14category food. The modality is image. So
3:24:16I think you guys understand the point.
3:24:18We have a source of truth that gets
3:24:19embedded and then it gets placed
3:24:21somewhere in a multi-dimensional space
3:24:23based on the actual meaning or you know
3:24:25value of what that source of truth is.
3:24:28And so that's why it's so cool that we
3:24:29can have a space where we have images,
3:24:31videos, audio, text, documents all in
3:24:34the exact same space and the AI is
3:24:36intelligent enough to query through it
3:24:37to find what it needs in the right
3:24:38context and when. And this is obviously
3:24:40a bizarre example because smiley face
3:24:42fries and a dog playing guitar and a
3:24:44video of me talking have nothing to do
3:24:45with each other. But if all of these
3:24:47were pictures of roofs, for example,
3:24:48then they would be very split up based
3:24:50on like is this water damage or is this
3:24:52just like old age or you know other
3:24:54things about roofs. So if you've never
3:24:56used cloud code before or you want to
3:24:58follow along with this video exactly, I
3:24:59use it in Visual Studio Code which is
3:25:01free to download. And when you download
3:25:02that, it'll look like this. All you have
3:25:04to do is open it up, go over here to the
3:25:06extensions, type in Cloud Code, install
3:25:08this, and then sign in with your account
3:25:10in order to get connected. You do have
3:25:12to be on a paid account. You cannot use
3:25:13free cloud code. And then what you're
3:25:15going to do is click on this in the top
3:25:16left to open up a new folder which is
3:25:18basically just the project that we're
3:25:19going to work in. And I'm going to open
3:25:20up a brand new one. So my screen will
3:25:21look exactly like your guys' screen and
3:25:23you can just follow what I do. Okay. So
3:25:25I just opened up a folder called
3:25:26embedding demo. I have this stuff over
3:25:29here. I'm going to exit out of. I'm
3:25:30going to click on this orange button
3:25:31which opens up cloud code. And now your
3:25:34screen should look like this. So I'm
3:25:35going to show you exactly how I got
3:25:37everything set up. I went ahead and I
3:25:38switched to plan mode. I went over to
3:25:40this documentation from Google and I
3:25:42went to the actual like API embeddings
3:25:44information. I copied this URL, pasted
3:25:47it in and said, "Hey, Cloud Code, I want
3:25:50to use Gemini's new embeddings 2 model
3:25:54in order to have a Pine Cone vector
3:25:56database filled with videos and images
3:25:59and text. Can you please build me a plan
3:26:02to set all of this up? create me av file
3:26:05with the placeholders and I will drop in
3:26:06my pine cone API key, my Gemini API key,
3:26:09and my open router API key. So, the Pine
3:26:12Cone API key is so we can set up the
3:26:14database. That's actually going to look
3:26:15like this. Just go to pine cone.io. And
3:26:18you can see in here we've got our
3:26:19different databases for our manual
3:26:22multimodal for our roofing projects. And
3:26:23then this was just a random one. And the
3:26:25cool thing is all you have to do is give
3:26:27cloud code your API key. It will build
3:26:28the database and it will throw
3:26:30everything in there. You don't have to
3:26:31do anything. So on Pine Cone, you can go
3:26:32ahead and use the starter plan, which is
3:26:34free. And this will be more than enough
3:26:35to just get started to see how it works.
3:26:37And then you're going to go over to
3:26:38Google AI Studio. You're going to come
3:26:40over here to get API key and then create
3:26:42a new API key right here. And that's
3:26:44going to be for accessing Gemini's new
3:26:46embeddings model. And then you're going
3:26:47to go to Open Router.ai. If you wanted
3:26:49to, you could use an OpenAI key or
3:26:51Enthropic key. But Open Router basically
3:26:53just lets you have all of these models
3:26:54in one, which is why I like to use it.
3:26:56So once you get an account in here,
3:26:57you'll basically just go to your
3:26:58account. you will come to your API keys,
3:27:00create a new one, and then give that to
3:27:03Cloud Code. So those are the three
3:27:04things we need. So what it does now is
3:27:05it spits out this plan. So we can
3:27:07basically read what Cloud Code is
3:27:08planning on doing. Here's the context.
3:27:10Here's the proposed project structure
3:27:12that it's going to create. Here are the
3:27:14dependencies. And here is the basically
3:27:15step-by-step plan. Now, if you wanted to
3:27:17change anything, you could highlight it,
3:27:18you could add comments, and you can make
3:27:20suggestions. For the sake of the demo,
3:27:21I'm just going to go ahead and auto
3:27:22accept what Claude is thinking to do.
3:27:25And hopefully it gets everything built
3:27:26out for us. And then all we have to do
3:27:28is give it the documents that we want to
3:27:29embed. So here's the to-do list. I'm
3:27:31just going to check in with you guys
3:27:32once this has finished up. Okay, so you
3:27:34can see it built all those files. And
3:27:35now in ourv it gave us these
3:27:38placeholders. So this is where you would
3:27:39go grab Gemini, paste it here. Go grab
3:27:41Pine Cone, paste it here. And then after
3:27:42you paste all three, just make sure you
3:27:44save this file before you exit out of
3:27:46it. All right, so I added those keys and
3:27:48now I said where should I add my images,
3:27:50videos, and text? And it wants me to put
3:27:52them into the data folder. So, I could
3:27:54open this up and make subfolders for
3:27:56image, videos, and documents. But what
3:27:57I'm going to do is I'm just going to
3:27:58drop everything in there, and I'm not
3:27:59going to tell it which is which.
3:28:01Obviously, it will be able to figure it
3:28:02out. So, I'm sorry for being boring, but
3:28:04I am going to use the same nine files
3:28:05that I used for for the earlier demo
3:28:07just because it's a good mix of, like I
3:28:08said, images, videos, text, and we're
3:28:11going to shoot this off now. All right.
3:28:13So, normally I would say this in plan
3:28:15mode, but I'm just going to keep sending
3:28:16it. Right now, I said media has been
3:28:18dropped in, as you can see over here.
3:28:20Get that into Pine Cone. then build me a
3:28:22simple chat web app on a local host so I
3:28:24can test that everything works well. I
3:28:26want you to use sonnet for the chat
3:28:27model. So these are the three pine cone
3:28:29indexes that we currently have and like
3:28:30I said it's going to go ahead and build
3:28:32us a new one because it has our API key
3:28:34now. Okay, so right now it just created
3:28:36that pine cone index and now it's doing
3:28:38the ingesting which like I said this is
3:28:40my favorite part because cloud code is
3:28:42so so powerful being able to do this and
3:28:43then the new model the new embeddings
3:28:45model is also so powerful. So combining
3:28:47them together makes this so seamless
3:28:49where you can build a database with
3:28:50natural language because like I said,
3:28:52I've built multimodal vector store
3:28:54agents before in NN, but it's so
3:28:57complicated. They're super fragile and
3:28:58there's a lot of hoops that you have to
3:29:00jump through in order to just store
3:29:01these different buckets and have all
3:29:02these descriptions of the images. This
3:29:04is so much better. And there's one thing
3:29:06I noticed is that it says to build the
3:29:07chat web app with Sonnet. So once it's
3:29:09done ingesting, I'm going to stop it and
3:29:11correct it. So, if you ever notice that
3:29:13it's doing something wrong, there's
3:29:14nothing wrong with stopping it and
3:29:16sending a new message because it still
3:29:17understands the context. So, now I'm
3:29:19going to go ahead and stop this. I'm
3:29:20going to say, "Hey, so I don't want you
3:29:22to build the chat web app with Sonnet. I
3:29:24want you to still use Opus. I just meant
3:29:26on the web app when we're actually
3:29:28talking with the AI, I want that AI
3:29:31model to be Sonnet." So, make sure you
3:29:32use the front-end design skill to build
3:29:34this chat web app. Just keep it super
3:29:36simple. Okay. Now, I'll check in with
3:29:37you guys when we have our PC. Okay. So,
3:29:40looks like our chat app is ready. I'm
3:29:41going to click on this logo host and we
3:29:43have multimodal rag chat. So now I can
3:29:45ask a question. Okay, so I'm just
3:29:46shooting this off to see what happens. I
3:29:48said, "How should I be looking to get
3:29:49workflow clients and do you have any
3:29:51interesting pictures of people meeting
3:29:52each other? Potentially warm
3:29:54connections." So I wanted to see if it
3:29:55can pull the workflow clients text,
3:29:57which it probably will, but I wanted to
3:29:58see what it did here with the picture of
3:30:00me and AI Adam Sandler. Okay, so it came
3:30:03back and has some some methods which
3:30:05came straight from my text document, but
3:30:07it says that it doesn't have any
3:30:08pictures of this exact thing. So, what
3:30:11I'm going to do is I'm just going to
3:30:12copy this actual string. And I'm going
3:30:15to go back into cloud code and say,
3:30:17"Awesome. It's working. Here is a
3:30:19conversation I just had with it." And
3:30:21then I'm going to paste in that
3:30:22conversation I just had. I want to
3:30:24understand the way that it thinks about
3:30:25the images or videos in the database so
3:30:27that I could ask a question that would
3:30:29get it to invoke and send me the image
3:30:31or video. And this is really the mindset
3:30:33shift when I work with Claude Code. It's
3:30:35just about being genuinely curious. If
3:30:37you don't understand something, just
3:30:38ask. So here it says right now the
3:30:40system can't actually show you images or
3:30:41videos. It only stores a text
3:30:43description alongside each embedding.
3:30:45Here's how it works. Images just the
3:30:46description I wrote during ingestion.
3:30:48Videos, same thing. It's just the
3:30:50description. So what you can do from
3:30:51here is you can have better descriptions
3:30:53to go along with all your pictures and
3:30:55videos which you kind of saw earlier in
3:30:56the demo where all of them had some
3:30:58metadata. And then what we can do is
3:30:59actually serve the media. So we can
3:31:01update the chat app to display images
3:31:02and videos inline when they come back so
3:31:04that you're not just getting a file
3:31:05name. Thanks for explaining that. I'm
3:31:07just doing a quick demo right now for
3:31:08YouTube. So, what I want you to do is
3:31:10just add some metadata about the dog
3:31:14playing guitar video, just saying that
3:31:15it is a cartoon golden retriever, I
3:31:17think, playing the guitar in front of a
3:31:19fireplace. And update the actual app so
3:31:22that it can serve us that media. And I
3:31:23just want to validate that this works
3:31:24and it's able to search through
3:31:26different types of media. So, as you can
3:31:28see now, what it's going to do is it has
3:31:30to reingest the video for better
3:31:32description and update the app. And I
3:31:33don't think by default it's going to do
3:31:35this, which is why I would say to use
3:31:36plan mode. But in this case, you might
3:31:38have two duplicate videos in the
3:31:41database. And you would want to make
3:31:42sure that it's deleting the old ones or
3:31:44it's basically just upserting this new
3:31:46one. So now if I say, "Show me the
3:31:47golden retriever playing guitar," it can
3:31:49actually pull that back. And right here
3:31:52in our app,
3:31:56we can watch the video. So this is just
3:31:58so so cool. You could have a database of
3:32:00tons of different videos and you could
3:32:01be able to actually search through them
3:32:02with rag. Now, the one limitation of
3:32:04that right now is that the videos are up
3:32:06to 120 seconds and only MP4 orov. The
3:32:10images are capable of processing up to
3:32:12six per request supporting PNG and JPEG
3:32:14formats. And I imagine that this stuff
3:32:16is going to get a lot better. You can
3:32:17even see that it was able to get over
3:32:19this limitation because the document
3:32:20that I gave it was like 68 pages long.
3:32:22It just had to figure out how it could
3:32:24break that up, chunk it up, and still
3:32:25maintain context. And I didn't try with
3:32:27audio yet, but that would be very
3:32:28similar to the way you do your videos
3:32:30and images. The key thing about the
3:32:32audio is being able to give it good
3:32:33descriptions so that the AI understands
3:32:35what's actually in that audio file. So
3:32:37that's where the subject matter
3:32:38expertise of the systems that you're
3:32:40building really, really does matter. The
3:32:41importance and value is way more
3:32:43shifting towards being able to
3:32:45communicate clearly, having
3:32:46understanding of processes, deep
3:32:48understanding of processes and where
3:32:50holes might be and where you need to be
3:32:51very explicit rather than just knowing
3:32:54technically how to configure different
3:32:55nodes and how to formulate a JSON body
3:32:58for an HTTP request.
3:33:02All right, so we've been working a lot
3:33:03with workflows and automations and data
3:33:05and Python scripts and you know stuff
3:33:07like that. Let us get a little more
3:33:09creative now. We're going to jump into a
3:33:10section about learning how to build
3:33:12really nicel lookinging websites and
3:33:14actually being able to deploy those
3:33:15websites on a real domain. So let's get
3:33:17into it.
3:33:19All right, so a lot of us have been
3:33:21building editin workflows for a while
3:33:22now. So today I'm going to show you how
3:33:23you can take any of your editin
3:33:25workflows that you already have and turn
3:33:26that into a web app. And I'm not talking
3:33:29about just showing you something like
3:33:30lovable to build a front end and then
3:33:31connecting it to your end web hook. I'm
3:33:33talking about cloud code having the
3:33:35ability to look at your workflow,
3:33:36essentially audit it to make sure it's
3:33:38ready to go for an app, and then make
3:33:39any of the changes that you need on the
3:33:41back end before you build the front end.
3:33:42But let me show you guys what I actually
3:33:44mean by that. Otherwise, it just sounds
3:33:45like a bunch of gibberish. So, here's a
3:33:47workflow that I wanted to turn into a
3:33:48web app. It takes a form submission
3:33:50where we get information from a user
3:33:52like product name, product photo,
3:33:53avatar, features, and video setting, and
3:33:55it turns that into a UGC ad with this
3:33:57workflow, as you can see right here. So
3:33:59what I told Cloud Code to do was look at
3:34:00this workflow and then just optimize it
3:34:02so I could actually use it and connect
3:34:04it to a front end. And what it did is it
3:34:06changed the workflow to look like this.
3:34:08There are actually a lot of changes that
3:34:09it made here. And I need you guys to
3:34:11believe me when I said I was seriously
3:34:12impressed when I saw this. So real
3:34:14quick, just wanted to put these side by
3:34:15side so you can actually see what it
3:34:17did. On the left was the original and on
3:34:19the right is what Cloud Code built. So
3:34:21first of all, it switched out the form
3:34:22submission trigger to be a web hook. Not
3:34:24too hard, but that's what it did. So if
3:34:25you remember, one of the raw inputs it
3:34:27gets was a photo. So cloud code actually
3:34:29realized that it's going to come through
3:34:30as a B 64 string when we send it over
3:34:32web hook and it has to convert that. And
3:34:34then at the end what we had to do is we
3:34:35had to figure out how did we want this
3:34:37to be displayed in the front end. So we
3:34:38basically are sending back a message
3:34:40whether it was successful or not. And
3:34:42we're sending over the URL so it can be
3:34:44embedded in the actual landing page. And
3:34:46it also changed all of these HTTP
3:34:48requests to be continue with an error
3:34:50output. And it routed the error to a
3:34:53different branch which would send the
3:34:54front end an error message. And another
3:34:56cool thing to realize is that when it
3:34:57changes the actual source of the input
3:34:59data, it had to change the variables
3:35:01everywhere else. So it really thought
3:35:02about the actual node by node flow, not
3:35:04just changing the input and the output.
3:35:06So if you don't understand all of those
3:35:08changes that I just explained and like
3:35:09why that's important, it's not a huge
3:35:11deal. The point I was trying to make
3:35:12there was just showing you that Claude
3:35:14will look at your workflows and fix them
3:35:16for you before you ever turn them into a
3:35:17front end. And you guys know that my job
3:35:19is to make complex or intimidating
3:35:20things as simple as possible. So that's
3:35:22exactly what I'm going to do today.
3:35:23We're going to walk through it all step
3:35:25by step and you're going to realize how
3:35:26easy it actually is. So, real quick
3:35:28before we get into that, I just want to
3:35:29do a quick demo of the final product of
3:35:31this that took me basically 40 minutes
3:35:33where I started with this workflow.
3:35:34Cloud code turned it into this workflow.
3:35:36And now we have this front end where I
3:35:38can put in the information. And let me
3:35:40just show you guys a quick demo. So, I
3:35:41put in some information, I put in a
3:35:43product photo, and I'm going to go ahead
3:35:44and hit generate. And now, what happens
3:35:46is it basically tells us on this right
3:35:47hand side that we have this one job
3:35:49processing. If I go into the actual end
3:35:51workflow that it's hitting right now and
3:35:53I go to executions, you can see that
3:35:54there were some failures when I was
3:35:56doing testing and stuff. But what we're
3:35:57going to notice is that we get a new
3:35:59execution right here pop through. And
3:36:01then when that's finished, it will
3:36:03automatically display right here where
3:36:04we can see the video and we'll be able
3:36:06to download it. So you can see that the
3:36:07workflow just finished up and you can
3:36:09see we have our video right here which
3:36:10is displayed in the website. We can
3:36:12click on this link to download it. And
3:36:14also just for reference, here is the
3:36:15original cologne image that I uploaded.
3:36:17So you can see that it pretty much looks
3:36:19the exact same. So I'm not going to be
3:36:20diving into this actual workflow that
3:36:22produced the results. I already made a
3:36:24video about this, so if you want to
3:36:24check it out, I'll link it up there. All
3:36:26right. Hopefully I'm not losing you
3:36:27already. I know that this workflow and
3:36:29this demo may seem a little bit complex,
3:36:31but we're going to set up everything
3:36:32step by step from the full process of
3:36:34taking an edit in workflow, optimizing
3:36:36it with cloud, and then getting it onto
3:36:38a front end and deploying it. Okay, so
3:36:40step one is open up VS Code. This is
3:36:42where I like to work with Cloud Code as
3:36:44an extension because the actual visual
3:36:47interface is just so much cleaner. It's
3:36:48so much better and you don't have to
3:36:50look at your nasty terminal or anything
3:36:51like that. VS Code's been around for a
3:36:53long time and it's a very trusted
3:36:54platform. So once you're in here, you're
3:36:56going to click on this lefth hand side
3:36:57and go to extensions and then you're
3:36:58going to type in up here cloud code.
3:37:01Once you do that, just click on cloud
3:37:02code and then go ahead and install it.
3:37:04And when you install it, it should
3:37:06prompt you to sign in with your
3:37:07Enthropic with your cloud account. And
3:37:09that's how you actually link them
3:37:10together. Now, once we have that
3:37:11extension installed, we actually need to
3:37:13start up a project. So, what I'm going
3:37:15to do here is I'm going to go to the
3:37:17lefth hand side and go up to this button
3:37:19right here, which is the file explorer,
3:37:20and it says you have not opened a folder
3:37:22yet. So, go ahead and open one up. So,
3:37:24I'm in my documents. I'm in a folder
3:37:26called Aentic Workflows. And then I'm
3:37:27drilling down to another folder called
3:37:29Nadent app, which has nothing in it. So,
3:37:31it's a blank folder. It's a blank
3:37:32project. And I'm going to select it,
3:37:34which now gets us into this environment.
3:37:36So, you can see up top we've got Nident
3:37:38app, which is our project. And on the
3:37:40lefth hand side over here, we're going
3:37:41to see all the other files that we add
3:37:43to this or that get created. In the
3:37:45middle is where we're going to actually
3:37:46be chatting with Claude Code. And the
3:37:47way we do that is by clicking on this
3:37:49little Cloud Code extension button right
3:37:51here and then closing out of whatever
3:37:52else we don't want. And on the right
3:37:54hand side is where we have the actual
3:37:56like VS Code agent chat, which means we
3:37:58can talk to this agent about like what's
3:38:00going on in here. And honestly, I never
3:38:01really use this because the Cloud Code
3:38:03agent is smart enough. So that's kind of
3:38:05the interface we're looking at. I know
3:38:07it may seem a bit overwhelming right now
3:38:08because there's lots of new buttons and
3:38:10there's lots of places to look, but I'm
3:38:11basically just going to tell you guys
3:38:13about what you need to know. And if you
3:38:14follow this demo all the way through, by
3:38:16the end of it, you'll have a really good
3:38:17understanding of what you're looking at
3:38:18and how to work with cloud code. All
3:38:21right, so the first thing that we want
3:38:23to do whenever we start up a new project
3:38:24is we want to give it some sort of
3:38:26guidelines about what are we actually
3:38:28doing in this project. And the way we
3:38:30actually do that is we just have to
3:38:31create a file, which is essentially the
3:38:32system prompt. And it's going to be
3:38:34called claude.md.
3:38:36And so what you could do is come over to
3:38:37the lefth hand side and you could click
3:38:38on new file. You could type in claude.md
3:38:42and then we could basically just start
3:38:43working in this file or we could have
3:38:46claude itself edit the file. And the
3:38:48reason why this popped up over here is
3:38:50just so we could view it. You could
3:38:51close it. You could open it back up. You
3:38:53could open up like 10 different files at
3:38:55the same time if you want to. But let's
3:38:56just keep our screen clean and keep open
3:38:59just the cloud code for now. Okay. So
3:39:01what I'm going to do is have cloud code
3:39:02help us write that cloud.md file. So,
3:39:04let me just read out what I actually
3:39:06wrote to it. So, I said, "Help me create
3:39:07a claw.md file in this project to set up
3:39:10what we want to do here. This project is
3:39:12essentially built to help me turn my NDN
3:39:14workflows into apps." So, there's going
3:39:16to be a few pieces. The first piece is
3:39:18going to look at my workflows in NN to
3:39:20make sure that they're ready to go as
3:39:21far as having the right intake of data
3:39:23and output of data so that if it's a web
3:39:25app, when the app sends data to NN, it
3:39:27can properly receive it. And that also
3:39:29when NN sends a response back to the
3:39:30front end, it's properly displayed. Just
3:39:33like we saw here in this example, I
3:39:35wanted to make sure that when NN sent
3:39:36the response back to the app, it could
3:39:38be displayed as an embedded video. And I
3:39:40also wanted to make sure that when we
3:39:41sent over a JPEG file to NN, it could
3:39:44receive it properly. Then I came back
3:39:45and said once we know the workflow's
3:39:47optimized, then we have to start
3:39:48building the front end. So we're going
3:39:50to start building it and testing it in a
3:39:51local environment and then once we like
3:39:53how the app looks and functions, then
3:39:55we'll push it to GitHub. And GitHub is
3:39:57basically just a home for our code and
3:39:59it will let us do different versions and
3:40:01see all the changes. And then what
3:40:03happens is our code lives in GitHub, but
3:40:04then we're going to have Verscell sync
3:40:06up to it. And Verscell is where we
3:40:08actually deploy those apps on the web.
3:40:10And I have a diagram to break this down
3:40:12in a few minutes here, but essentially
3:40:13the idea is we work in cloud code, we
3:40:15push changes to GitHub, and then our
3:40:17actual real web app on the public URL
3:40:20always reflects the most recent version.
3:40:22So it's just super easy. So there's also
3:40:24a couple things that we're going to
3:40:25utilize. One of them will be the niten
3:40:26mcp so that you can understand nitend
3:40:29nodes configurations templates and you
3:40:31can look through my niten instance and
3:40:32create and edit workflows and things
3:40:34like that. I'm also going to give you
3:40:36access to two skills the niten skills
3:40:38and the front-end developer skills. And
3:40:40I'm going to give you access to the
3:40:41GitHub MCP so you can actually push
3:40:43changes to my GitHub. And then I finish
3:40:46that off by saying with all that in
3:40:47mind, ask me any questions that you may
3:40:49need and help me make this file concise
3:40:51so that we keep everything neat and
3:40:53lean. So before I send this off, I
3:40:55wanted to talk about these different
3:40:56modes. So right here you can see I'm on
3:40:58bypass permissions which is orange. We
3:41:00could go to ask before edits which is a
3:41:02lighter orange. We could go to edit
3:41:04automatically which is white. Or we
3:41:05could go to plan mode which is blue. And
3:41:07so whenever I'm doing something like
3:41:08this or whenever we're setting up an
3:41:10initial prompt I always like to use plan
3:41:12mode because it thinks a lot better and
3:41:14it asks you questions and it basically
3:41:16just lets you guys have a conversation
3:41:17before you actually do anything. So, I'm
3:41:19going to go ahead and shoot off this
3:41:20prompt in plan mode, and we're going to
3:41:22see Claude Code think about what it
3:41:24needs to do. It's going to first look
3:41:26through the current structure to see if
3:41:27there's any files. It's going to
3:41:28understand what we're doing here. And
3:41:30you can see right here, it said before I
3:41:31draft the file, I've got a few
3:41:32clarification questions. So, what's the
3:41:34typical structure of the workflow you
3:41:36want to turn into an app? Right now,
3:41:37let's just say various triggers because
3:41:39we don't actually know what we want to
3:41:40turn into an app yet. For the project
3:41:43structure, do you have a preferred
3:41:44project structure in mind? I'm just
3:41:46going to say propose structure,
3:41:48whatever. I don't really care. Repo
3:41:50strategy. Should each workflow become
3:41:52its own repo in GitHub? Yep. We'll just
3:41:54do separate repos, one for each app. And
3:41:56then for styling, we'll just go ahead
3:41:57and go with Tailwind CSS. And if you
3:41:59don't know what this stuff means, you
3:42:01can just go ahead and ask it to. That's
3:42:02the beauty of Claude code is that we
3:42:04don't have to really understand all the
3:42:05code and exactly what it's doing. We
3:42:07just have to be able to communicate our
3:42:09thoughts clearly. And if we get
3:42:10confused, just ask Claude what it's
3:42:12doing and why because it's really good
3:42:13at that. So you can see it gave us this
3:42:15plan for the cloud file and I said,
3:42:16"Yep, that sounds good." I'm going to go
3:42:17ahead and auto accept and now it's going
3:42:19to update this cloud MD file which right
3:42:22now has nothing in it as you can see and
3:42:24it's going to basically just write in
3:42:25the system prompt and you can see that
3:42:26it just happened in real time right
3:42:28there. Okay, our system prompt is
3:42:30configured. So next what we have to do
3:42:32is give it access to all of those things
3:42:33that we mentioned like the skills, the
3:42:35servers, whatever the MCP. But before we
3:42:38do that, let me just show you guys
3:42:39exactly what we're doing here on a
3:42:40whiteboard so it all makes way more
3:42:42sense. So in my last video, I showed you
3:42:44guys how we can use cloud code and give
3:42:46it the end mcp and the end skills to
3:42:49build workflows for us in our own end
3:42:51instance. So we're kind of building on
3:42:52top of that here. If you haven't watched
3:42:54that video, that one might be a good one
3:42:55to start with and then come back to this
3:42:57one. So I'll tag that right up here. But
3:42:59essentially the end mcp gives cloud code
3:43:01access to all the nodes, configurations,
3:43:03workflow patterns, things like that. and
3:43:05ended in skills gives cloud code all the
3:43:07knowledge about expressions, how to use
3:43:09this MCP server, how to code, all this
3:43:12kind of stuff. So the TLDDR is you're
3:43:14essentially giving one of the smartest
3:43:16brains in the world access to all of the
3:43:19information about NN that you could
3:43:20possibly need. So now we're building on
3:43:22top of that and we're creating web apps.
3:43:24So what we do here is we've got once
3:43:25again cloud code with MCP servers and
3:43:27with different skills and now what we
3:43:29wanted to do is create us a web app. So
3:43:31in order to create that web app first of
3:43:33all we use end to see the backend
3:43:35automation that we want to turn into an
3:43:37app and we create the front end to
3:43:39actually like collaborate with that. Now
3:43:40what is the front end? It's actually
3:43:42just code. So cloud code is building the
3:43:44code that displays the website and what
3:43:47we do with that code is we push that to
3:43:48GitHub in something called a repository
3:43:50or what a lot of people just call a
3:43:52GitHub repo. And then we have Verscell
3:43:54which actually deploys it on the
3:43:55internet so that other people could
3:43:56access this app. and Verscell is
3:43:58constantly looking at your GitHub repos
3:44:00so that if anything changes over there,
3:44:02you can basically have that change be
3:44:03instantly reflected on your real app. So
3:44:06like with this web app, the actual code
3:44:08for this lives in my cloud code locally.
3:44:11It's also reflected on GitHub and then
3:44:13Verscell has deployed it. So let's say I
3:44:15wanted to make this green instead of
3:44:16blue. I could tell cloud code to change
3:44:18the code to make it green and then I
3:44:20could say push this to GitHub. So,
3:44:21GitHub would grab it and then Verscell
3:44:23would grab it from GitHub and then in
3:44:24like 20 seconds we would see that this
3:44:26website would be green instead of blue.
3:44:28Hopefully that architecture makes sense.
3:44:30Now, let's get back into cloud code and
3:44:32let's start connecting all these things
3:44:34that we need. So, now what I'm doing is
3:44:36I'm saying connect all of these MCP
3:44:38servers and skills and just let me know
3:44:39if you need anything else. So, I'm
3:44:41giving it the URL for the NIN MCP
3:44:43server. I'm giving it my cloud URL. I'm
3:44:46giving it my NIN API key. I'm giving it
3:44:48the GitHub MCP server URL. I'm giving it
3:44:51my GitHub personal access token. I'm
3:44:53giving it the repo for my the end skills
3:44:56and I'm giving it the repo for the
3:44:58front-end design skills. All of these
3:45:00links that you'll need, I'll just put in
3:45:02the description of this video. And what
3:45:03I'm going to do is I'm going to turn
3:45:04this on bypass permissions mode because
3:45:07I just want it to go without me having
3:45:08to approve everything. So, I'm going to
3:45:09go ahead and shoot this off and let it
3:45:11work its magic. And while that's
3:45:12running, I'll show you guys two things.
3:45:14The first one is how do you get bypass
3:45:15permissions mode if you don't see it
3:45:17natively right there? Well, you go to
3:45:19your settings down here and then you
3:45:21would type in clawed code and then right
3:45:23here you just have to turn this check
3:45:25mark on that says allow dangerously skip
3:45:27permissions. Now, I know it sounds
3:45:29dangerous, but it's not too bad as long
3:45:31as you're watching it and like, you
3:45:32know, making sure that you're not
3:45:33telling it to go delete all your files
3:45:35and things like that. Now, the second
3:45:36thing I wanted to show you is how to get
3:45:38your GitHub personal access token. So,
3:45:40here is my GitHub. All you have to do is
3:45:42just go to GitHub, create an account.
3:45:43It's free to create an account. And then
3:45:45you're going to go up here to your
3:45:46settings and then at the bottom of your
3:45:49settings you should see developer
3:45:50settings and you're going to go ahead
3:45:51and create one of these personal access
3:45:53tokens and you'll create a fine grained
3:45:55token. So that's all you have to do.
3:45:57It'll give you basically an API key and
3:45:59then that's what you're going to give to
3:46:00claude code here so that it can set
3:46:02everything up. And when you actually go
3:46:03to create this token pretty much just
3:46:04give it a name. I leave mine on public
3:46:06repositories. I change the expiration to
3:46:08never. And then the last thing is about
3:46:10permissions. And usually what I do is I
3:46:12just add all of these. There might
3:46:13there's like 23 or something, but just
3:46:15add all of them. If you realize later
3:46:16you want to restrict something else, you
3:46:18can just go ahead and create a new one
3:46:19or restrict it in cloud code. It's not a
3:46:21huge deal. And then generate the token
3:46:23and pop it over to cloud code. And so
3:46:25the other thing to look at, and
3:46:26throughout this tutorial, you might see
3:46:28me use like a slashclear or some other
3:46:30things, but if you hit slash, you can
3:46:32see that there's other things that we
3:46:33can look at. So we can attach a file, we
3:46:35can mention a file from a project, we
3:46:37can switch the model. So we can go from
3:46:39default or sonnet 4.5, opus, haiku. We
3:46:42could also turn on thinking. We can
3:46:44manage our MCP servers, agents, hooks,
3:46:46memory. We can do all of these other
3:46:48slashcomands as well. And then if you
3:46:50actually use cloud code in something
3:46:51like cursor or in the terminal
3:46:53environment, there's even more commands
3:46:54and like more things you can do with
3:46:56agents and like plan mode and things
3:46:57like that. But like I said, VS Code just
3:47:00makes this all look a lot cleaner and a
3:47:02lot less intimidating, which is why I
3:47:04wanted to do cloud code on VS Code in
3:47:06this tutorial. So now it's asking me,
3:47:08how would you like me to configure all
3:47:10of this stuff? I'm just going to say
3:47:11create the MCP JSON file because I don't
3:47:13want to do it myself. I just want you to
3:47:15go ahead and take care of all of this
3:47:17stuff. And that's the thing that's
3:47:18interesting about this because when you
3:47:19go to a lot of these MCB servers or
3:47:22skills, it'll basically tell you
3:47:23installation steps and it will say,
3:47:25"Hey, go add this to your cloud code
3:47:27file or hey, go install this plugin."
3:47:29And I don't want to actually do that. I
3:47:30just want Cloud Code to do it. So, all I
3:47:32do is I give it the raw URLs. And what
3:47:35you can see here is when I give it the
3:47:37raw URLs, it just uses its web search
3:47:39tool and it reads the page and
3:47:40understands installation and then just
3:47:42does it. So a lot of times if cloud code
3:47:45comes back to you and says, "Hey, what
3:47:46you need to do is do this." You can come
3:47:48back to it and just say, "No, you do
3:47:50it." And most of the time it'll just do
3:47:51it. Every once in a while it'll say, "I
3:47:53actually can't. I need you to do this."
3:47:54But most of the time you can just tell
3:47:56it to do it for you. Like right here it
3:47:57says, "Install the skills by running
3:47:59these commands after restart." I'm going
3:48:00to say, "I don't want to install those
3:48:03commands. Can't you just do it for me?
3:48:04And what do you know? Done. Both skill
3:48:06sets are now installed. So I didn't have
3:48:08to do any of that. So right here we have
3:48:10the MCP JSON file. And this is where you
3:48:12can see we have our NADN MCP server and
3:48:14we have our GitHub MCP server. And
3:48:16you'll notice that we don't actually see
3:48:17the skills in here. And the reason why
3:48:19is because the skills were installed
3:48:21globally, not just within this one
3:48:23project that we're working in called
3:48:24endent app, which is cool because later
3:48:26if we make another project, we already
3:48:28have those skills installed. And if you
3:48:30don't believe me or you get confused,
3:48:31you just ask Claude. I said why don't I
3:48:33see the Niten skills in this project and
3:48:35it basically just came down and said
3:48:36yeah they're installed globally here are
3:48:38the seven skills and for some reason it
3:48:40said six so not the best at counting but
3:48:43then we also have the front-end skill
3:48:44down here so that's just to prove that
3:48:46they are actually installed even though
3:48:48you don't see them over here on the file
3:48:49explorer now something else to keep in
3:48:51mind about this MCP JSON file is that it
3:48:54has your real GitHub token and your real
3:48:56API key in here. So if you've shared
3:48:59this file for some reason or someone had
3:49:00access to this, they would be able to do
3:49:02anything in your end because they have
3:49:04this information. So obviously I'll be
3:49:06deleting these credentials after this
3:49:07video goes live. But just something else
3:49:09that I wanted to make sure you guys were
3:49:10aware of. That is why typically when
3:49:13you're doing certain things, you're
3:49:14going to have like av file and you'll
3:49:16have your actual scripts and things call
3:49:19on those credentials so they only use
3:49:20them when they need them and like all
3:49:22those files are encrypted. So don't want
3:49:25to confuse you guys. We're not going to
3:49:26dive into that right now. Just something
3:49:27that you should be aware of. All right.
3:49:29So, what I'm going to do now is I have
3:49:30to restart Cloud Code, otherwise it
3:49:32won't actually reflect all that. So,
3:49:33really what I'm going to do is just
3:49:35close out of VS Code and then we're
3:49:36going to open it back up and then
3:49:38everything should be all set. All right.
3:49:39So, I'm going to go ahead and do a slash
3:49:41command. I'm going to do /cle just to
3:49:43get rid of this conversation so we can
3:49:44start fresh on a new context. But keep
3:49:47in mind, every time we talk to cloud
3:49:48code, it's still going to be reading
3:49:49through our cloud. MD system prompt to
3:49:51understand what we're doing in this
3:49:53project. Okay. So, what actually are we
3:49:55going to turn into a web app? Well,
3:49:56let's make it pretty simple. I've got
3:49:58this workflow here which is just an AI
3:50:00agent and it's a chat window here and
3:50:02it's called fitness coach. So, in here I
3:50:04basically just have a system prompt
3:50:05prompting this agent to be a fitness
3:50:07coach with stuff like um you know
3:50:09weightlifting, working out, some basic
3:50:12nutrition stuff just so we can make a
3:50:14little demo here. But you can see that
3:50:15this is not ready to go to be turned
3:50:17into a web app. Not really at least. But
3:50:19all I'm going to do is just tell Cloud
3:50:22Code to look for this workflow and help
3:50:24me turn it into a web app. So, we're
3:50:25going to go back into cloud code. I'm
3:50:28going to change this to plan mode
3:50:30because we want to like brainstorm how
3:50:31we're going to do this. And I'm going to
3:50:33say, I've got a workflow in my edit
3:50:36instance called fitness coach. I want to
3:50:39turn this into a web app. So, before we
3:50:41do that, please take a look at it and
3:50:43help me change it so that it's ready to
3:50:45go and I can talk to it from a front
3:50:47end. So, I'll shoot this off and we're
3:50:49basically just going to watch it think.
3:50:50It's going to walk through its steps and
3:50:52what you can see right now is that it's
3:50:53using the end MCP to find our workflows.
3:50:56Now it was able to find the fitness
3:50:57coach. So it's going to analyze it and
3:50:59you can see that it found the workflow
3:51:00but there's an issue which is it's using
3:51:02the chat trigger which is not really
3:51:04designed for a custom front end. So it's
3:51:06going to write up a plan to change this
3:51:08workflow for us. So here it asks, do you
3:51:10want the fitness coach to remember
3:51:11conversation history? Because right now
3:51:13in the actual workflow there's no
3:51:15memory. So what I'm going to do is say
3:51:16yeah that's a great feature. We want the
3:51:18coach to be able to remember like it's
3:51:20having a conversation. Okay. So, here it
3:51:22came up with a plan which is to prepare
3:51:24the fitness coach workflow to be a web
3:51:26app. So, it tells us what the problem
3:51:27is. It tells us what it's going to do.
3:51:29So, it's going to replace the chat
3:51:30trigger with a web hook. It's going to
3:51:32add a window buffer. It's going to add a
3:51:34window buffer memory. It's going to
3:51:35update all these connections. It's going
3:51:37to configure the agent input. And then,
3:51:39it's going to publish the workflow as
3:51:40well. And so, we could obviously make
3:51:42some changes here if we want. I want to
3:51:43see how it did on its own. So, I'm going
3:51:45to go ahead and just auto accept all
3:51:46these changes. I'm going to make this
3:51:49bypass permissions so that we can just
3:51:51basically see when it's done. And what
3:51:53it does is it creates a to-do list,
3:51:55which is really cool because it helps
3:51:56the actual model stay on track, but it
3:51:58also helps us understand its thought
3:52:00process and what stage it's on. So that
3:52:02way, a lot of times when you're working
3:52:03with Cloud Code, you can have it open on
3:52:05one monitor, you can be doing something
3:52:06else, watching a YouTube video, whatever
3:52:08you want to do, and just kind of
3:52:09checking in and making sure that it's
3:52:11staying on the right path. And a lot of
3:52:13times it's not perfect, but what's so
3:52:14cool about cloud code is that it runs
3:52:16into issues and it analyzes what went
3:52:18wrong and then fixes it. So right here
3:52:20you can see that there was an error
3:52:21because the key parameter for the
3:52:23session ID for the memory was was
3:52:25missing. And so it figured out that the
3:52:27body is actually nested and then it went
3:52:29ahead and just you know changed the
3:52:31workflow after it realized that. So it's
3:52:33all about how much context you give it
3:52:34and how clearly you can explain what you
3:52:36want. And that's also why the planning
3:52:37mode is so helpful. But it looks like
3:52:40that workflow has been updated. Now,
3:52:42what I'm going to do is go back into
3:52:43edit end, which was this workflow right
3:52:44here. And I'm going to hit refresh, and
3:52:46we should now see that the workflow is
3:52:47changed, and it should be ready to go
3:52:49for our web app. So, we've got a web
3:52:53hook, which is post request. We've got
3:52:54our memory, and let's see if it changed
3:52:56the actual configuration of the user
3:52:58message, which it did. So, we also, I
3:53:00noticed, don't have a responder web hook
3:53:01node. So, what it's doing is it's using
3:53:03respond when the last node finishes with
3:53:05the first entry. So, I think we should
3:53:07be all right. But later if we end up
3:53:08needing to change something, we would
3:53:10just say, "Hey, Claude, this didn't
3:53:11work. Go fix the workflow." All right,
3:53:13so back in Claude, what I'm going to do
3:53:14now is I'm going to clear out this
3:53:16context once again, and I'm going to go
3:53:18back into plan mode, and we're going to
3:53:19talk about what we want this actual web
3:53:21app to look like. Hey Claude, help me
3:53:23create a plan to actually build out this
3:53:25front end for the end workflow. I want
3:53:27to make sure that in your plan, you're
3:53:29going to be using the front-end designer
3:53:31skill, the end tools skill, and all of
3:53:34the resources that you have to make this
3:53:36as good as possible. We don't want the
3:53:37web app to look AI vibecoded. We want it
3:53:40to look professional, very minimalistic,
3:53:42and we want it to be super clean. We
3:53:44also want it to be a little bit gamified
3:53:46to incentivize people to come in and to
3:53:48talk to the fitness coach. Maybe we can
3:53:49have the main chat interface and then on
3:53:51the right hand side, we can have a
3:53:52little bit of gamification with um a
3:53:55tracker for how many times people have
3:53:56talked to the fitness coach and maybe
3:53:58they can level up after, you know, every
3:54:00five or 10 messages or something like
3:54:02that. Ask me clarification questions to
3:54:04make sure that we're not leaving any
3:54:05holes in our plan here and any
3:54:08suggestions that you may want to make
3:54:09that I didn't yet think of. All right,
3:54:11I'm shooting that off. I'll let you guys
3:54:12know if anything important happens. All
3:54:15right, so I got some questions here. The
3:54:16first one is which end workflow should
3:54:18this front end connect to? We're going
3:54:19to do the existing workflow, which it
3:54:21should be able to find. Now we have what
3:54:23core features should the fitness coach
3:54:25provide when users chat with it? General
3:54:26fitness Q&A, personalized workout plans,
3:54:29progress tracking. Yeah, we'll just do
3:54:30all of these. gamification. For the
3:54:32gamification system, what should users
3:54:33be leveling up and earning rewards for?
3:54:35Just a message count. Let's just do
3:54:37that. And then what's your preference
3:54:39for user data persistence, remembering
3:54:40their level, message count, things like
3:54:42that. And I'm just going to say, let me
3:54:43decide later. We can get into a bunch of
3:54:45back-end database storage and, you know,
3:54:47authentication in different videos. We
3:54:49want to keep this one simple. So, for
3:54:50now, we're just going to go with that.
3:54:52Now, it's asking for the name or ID of
3:54:54the existing Fitness Coach workflow, the
3:54:56one that it found earlier. So, I'm just
3:54:57going to go ahead and type in the name
3:54:59or ID. And honestly, now that I think
3:55:01about it, it probably would have been
3:55:02better if I didn't clear the context and
3:55:04we just kept talking on that previous
3:55:05thread. But what are you going to do?
3:55:08Some more questions now. What color
3:55:09scheme or brand aesthetic do you want?
3:55:11I'm going to go with dark mode primary.
3:55:13The app name, let's just go with yeah,
3:55:15fit coach AI. And then mobile layout.
3:55:18Should the gamification panel be visible
3:55:20on mobile or only larger screens? We're
3:55:22just going to go with always visible.
3:55:24Keep things nice and easy. All right.
3:55:25So, the plan for the Fit Coach AI web
3:55:28app is done. There's tons of stuff in
3:55:30here with structure, text stack, key
3:55:32features, all this kind of stuff.
3:55:33Obviously, you would [snorts] read
3:55:34through this and make some changes if
3:55:35you want, but we're just going to see
3:55:37how Cloud Code does on its own. So, I'm
3:55:38going to go ahead and auto accept those
3:55:40changes. And of course, what it's going
3:55:41to do is set up a to-do list, as you can
3:55:43see right here. So, I'll just check in
3:55:44with you guys when we have something
3:55:45ready to test. Okay, so there we go. The
3:55:47to-do list has been completed and Fit
3:55:49Coach AI is ready. You can see that it's
3:55:52actually living right now on a local
3:55:54host, which basically just means only we
3:55:56could access this. So if I gave you this
3:55:58exact, you know, HTTP address, it would
3:56:01pull up your own local host, whatever
3:56:02you're hosting on 30002. So what we need
3:56:05to do is make sure it works here. And
3:56:06then you can see when you're ready,
3:56:08we'll initialize the git and we'll push
3:56:09it to GitHub and then we'll deploy it to
3:56:11Verscell. So I'm going to go ahead and
3:56:13open up this local host and let's take a
3:56:15look at what we've got. All right. So,
3:56:17this is the Fit Coach AI interface that
3:56:19it came up with. And you can see we've
3:56:20got ready to crush your goals. You can
3:56:22try creating a 30-minute HIT workout.
3:56:24What should I eat after working out? How
3:56:26do I stay motivated to exercise? We've
3:56:28got some stats over here. So, we've been
3:56:30a member since January 14th. We've got
3:56:33rookie level one, 11 points to the next.
3:56:35We've got a road map. So, it did
3:56:37everything that we were looking for.
3:56:38Let's see if it's actually able to talk
3:56:39to our NN workflow. So, I'm going to
3:56:41start off by just saying, "Hey there.
3:56:43We'll shoot this off and let's see if we
3:56:45get some sort of response back. Okay,
3:56:47cool. So, it gave us a response, but it
3:56:49doesn't look great as you can see
3:56:51because what happens is in nitn when we
3:56:53respond, we get the whole JSON body. So,
3:56:55we get the output and then we get all
3:56:57this other stuff. So, if I actually go
3:56:58to the fitness coach workflow, we go to
3:57:00executions, we can see that when claude
3:57:03code changed the workflow, it did
3:57:04everything right, which is great. But
3:57:06the actual output of the fitness coach
3:57:08is this JSON body. So basically the
3:57:10front end displays this JSON body rather
3:57:14than just the actual output which is
3:57:16what we want it to display. So super
3:57:18easy. What we're going to do is we are
3:57:20going to of course go back into cloud
3:57:22code and just tell it to fix that. So
3:57:24what I'm going to do is take a
3:57:25screenshot of this output just so that
3:57:27cloud code can see exactly what I'm
3:57:29talking about. We'll go back in here.
3:57:31I'll paste that in. So it's working
3:57:34good. But when the agent responds to us
3:57:36in the app, it actually displays the
3:57:38entire JSON body. We don't want to see
3:57:41the field called output. We just want to
3:57:42see the actual output itself.
3:57:45And for something simple like this, I'm
3:57:47not even going to go into plan mode
3:57:48because it's a very easy request. So, it
3:57:50should just be able to change the front
3:57:52end to configure just displaying the
3:57:54actual output. All right, sweet. So, it
3:57:56said that it fixed that. I'm going to go
3:57:57back in here and I'm going to say,
3:57:59what's the best time of day that I
3:58:01should be working out and eating? So,
3:58:03just something random. Shoot that off.
3:58:05And hopefully this time we only get back
3:58:06the actual output that we're looking
3:58:08for. Another thing that I am noticing
3:58:10though is nothing's actually happening
3:58:11on the gamification side. So, I imagine
3:58:15that this should be giving us points
3:58:16each time that we get a message back. As
3:58:18you can see, we did get a better output
3:58:20now. Although, I don't like that it's
3:58:22coming with like markdown and bold. So,
3:58:24that's something that we would actually
3:58:25change in the system prompt of the agent
3:58:27here rather than in the front-end
3:58:29development. So, not a huge deal, but
3:58:31we're not getting any points. So, that's
3:58:33the next thing that we have to tell
3:58:34Claude to fix. Awesome. That change
3:58:36worked, but now the issue is we're not
3:58:38actually getting any gamification. So,
3:58:40I've sent two messages now, but we still
3:58:41have zero points on the app itself. So,
3:58:44fix that. Okay, so it looks like it
3:58:46reset some stuff with local storage and
3:58:48whatever this is talking about. Now,
3:58:50let's go ahead and try again. It seems
3:58:51like that should have fixed. Although,
3:58:53now we have to see if we can refresh the
3:58:55page. Uh-oh. So, this is the actual
3:58:58local host that we were supposed to be
3:58:59using the app on, but there's some sort
3:59:01of issue here even when I refresh. So,
3:59:03let me take an a screenshot of this and
3:59:06send this over to cloud code. I can't
3:59:08access the app anymore. Okay, so it says
3:59:12that it rewrote everything with a
3:59:14simpler approach. We should need to try
3:59:15to refresh. And there we go. Cool. So,
3:59:18we actually do have our two messages
3:59:19that we already sent. Let's just say um
3:59:22create a 30-minute hit workout. Shoot
3:59:23that off. And hopefully when the agent
3:59:25responds we'll get another point on the
3:59:27right hand side. Cool. We did. So we got
3:59:30the response and we also got another
3:59:32message. Now finally before we actually
3:59:34push this to GitHub I wanted to show you
3:59:36guys that we could change the system
3:59:38prompt with cloud code. So one last
3:59:40request.
3:59:42Awesome. That's working. My last request
3:59:45is that the agent is responding with
3:59:47markdown formatting and bold stuff. So I
3:59:50wanted to just respond in complete
3:59:52natural language paragraph form like an
3:59:54actual human would. So go ahead and make
3:59:56that change in the nitin workflow and
3:59:58update the system prompt of the AI
4:00:00agent.
4:00:03All right, cool. So the nitin workflow
4:00:05has been updated. It went ahead and
4:00:07changed the system prompt so that now it
4:00:08should only be responding in natural
4:00:10language. I'm just going to go ahead and
4:00:12go to the workflow and refresh it. Make
4:00:14sure everything is all set. Looks like
4:00:17we've got it saved. It's still
4:00:18published. So let's just do one final
4:00:19test in here. So I set it like that
4:00:21because I want to make sure that the
4:00:22memory is working. I want to make sure
4:00:24that it comes back and responds to us in
4:00:26natural language only. No formatting and
4:00:28that once again we will get the um extra
4:00:31message. Okay. So we still got a little
4:00:33bit of bold things and we'd probably
4:00:34just have to go back. But what I wanted
4:00:36to actually make sure of was that in
4:00:38here it actually did change the system
4:00:39prompt which you can see that it did. So
4:00:41I guess maybe we just weren't explicit
4:00:43enough at actually how to prompt it. But
4:00:46the point I was trying to make, which is
4:00:47what I think is really important, is
4:00:49that what we just did is we had a random
4:00:52workflow. We had Cloud Code look at it,
4:00:55optimize it for a front end. We built
4:00:57the front end, and then we went back and
4:00:59forth with Cloud Code when the front end
4:01:00wasn't working how we wanted it to. And
4:01:02then we also had Cloud Code change the
4:01:04actual backend end workflow itself. So
4:01:07everything that we're doing here is just
4:01:08using our natural language to cloud code
4:01:11and just speaking very clearly about
4:01:12what we want. And that's obviously a
4:01:14good example. I wasn't clear enough
4:01:16about the way that I want it to respond.
4:01:18So I would just have to go back one more
4:01:19time. But now let's actually move on to
4:01:21the next step here where we're going to
4:01:23take the code and we're going to push it
4:01:24to GitHub and then have that
4:01:26automatically sync with forcell. So
4:01:28first of all, what you need to do of
4:01:29course is go ahead and go to GitHub. So
4:01:32this is my GitHub. You've already made
4:01:34your access token to give it to Claude
4:01:36Code. But now what we need to do is
4:01:38create a repo. So I'm going to go up
4:01:40here, create new. I'm going to click on
4:01:41new repository. I'm just going to call
4:01:43this fit coach- aai since that is our
4:01:47name of our app. And then I'm just going
4:01:49to do dash app. Um, I'm not going to do
4:01:51description. We'll just leave it public.
4:01:53I'm not going to add a readme or get
4:01:55ignore any of that kind of stuff. I
4:01:56literally just added a name and I'm
4:01:58going to create that repository. Now,
4:02:00what I'm going to do is I'm just going
4:02:01to take the actual URL up at the top in
4:02:03my browser and copy that. Go back into
4:02:06Cloud Code. We're going to go ahead and,
4:02:08you know, let's just keep going in the
4:02:09same context window. I'm going to say
4:02:12here is the GitHub repo for this app.
4:02:16Paste in the URL and say please push
4:02:18this to GitHub so we can sync it with
4:02:20Forcell and get it live, get it
4:02:22deployed. So I'm going to shoot that
4:02:23off. I'm also leaving it once again on
4:02:25bypass permissions because this is a
4:02:27pretty simple request. And now it's
4:02:29going to go ahead and do that for us.
4:02:30And now in the GitHub repo, you can see
4:02:32there's nothing here. But what's going
4:02:33to happen is it will get um all of our
4:02:35files will be pushed to this
4:02:37environment. And once again, as you're
4:02:39working with GitHub and Verscell, you
4:02:40can ask Cloud Code any of the questions
4:02:41you may have about how they work
4:02:42together or why it's important, and it
4:02:44will get back to you. All right, cool.
4:02:46So, the code has been pushed to GitHub.
4:02:48Now, it says next steps would be to go
4:02:50to Verscell and import that repository.
4:02:52And then it also says to add the
4:02:53environment variable, which would be our
4:02:55actual web hook. But we'll we'll take a
4:02:58look at that when we get there. So,
4:03:00let's go to GitHub real quick. And if I
4:03:02refresh this repo, we can now see that
4:03:05we have these files in here. And
4:03:07basically these are the files that hold
4:03:08the code of our fit coach app. Now what
4:03:10I'm going to do is I'm going to go to
4:03:12versel which is right here. This is the
4:03:14UC one that I was looking at earlier.
4:03:17And what's going to happen is you can
4:03:18see that this is pulling from my GitHub
4:03:21repo for this code. So what I'm going to
4:03:23do here is click on add new. We're going
4:03:25to add a new project. And I should be
4:03:27able to see import git repository. And
4:03:29right here we have our fit coach AI app.
4:03:31So I'm just going to go ahead and click
4:03:32import on that repo. And now I could
4:03:34change the name or the team or the
4:03:36preset or whatever I want to do, but I'm
4:03:38just going to go ahead and click deploy.
4:03:40All right, cool. So, we just deployed
4:03:41this project. I'm going to go back out
4:03:42to my dashboard and I'm just going to
4:03:44show you guys how you can get there. So,
4:03:46if I go back home, you can see the
4:03:48different projects that you have in your
4:03:49Verscell. So, if I just give this a
4:03:51refresh, we can see we have the Fit
4:03:52Coach app. So, I click into that. What
4:03:54you can look at here is deployments. So,
4:03:56every time you have a new version, you
4:03:58can see when it was actually uploaded.
4:04:00We can look at our logs. We can look at
4:04:02gateways, storage, we can look at all
4:04:04this kind of stuff. But right now, what
4:04:06I wanted to do was actually just click
4:04:07on the project itself. So we can go to
4:04:09that domain, which is fit
4:04:10coach-ai-app.verell.app.
4:04:14So if I open that up, this is what we
4:04:16see. And we're no longer on localhost.
4:04:18Now we are on this domain, which means
4:04:21that if I gave you guys this URL, you
4:04:23could access this and you could talk to
4:04:24like my instance and all that kind of
4:04:27stuff. So what I'm going to do is we're
4:04:28just going to say, how do I stay
4:04:30motivated? And what we get is a server
4:04:31configuration error. And so we need to
4:04:34figure out what happened here. So the
4:04:36first thing that I want to look at is
4:04:38I'm going to go to my actual end
4:04:39workflow and I'm going to go to
4:04:41executions. Now what happens here is we
4:04:43don't see that execution that we just
4:04:46did. So what that tells me in my brain
4:04:48is that we have our front end deployed
4:04:50on the web. We have ended end deployed
4:04:52on the web but for some reason they're
4:04:53not talking to each other. So when I hit
4:04:55this button which says you know like
4:04:57when I send the message when I hit that
4:04:59button normally what that does is it
4:05:01sends this string of text to the end and
4:05:04web hook but for some reason that's not
4:05:05configured. So if you remember we went
4:05:07back into cloud code and it says to add
4:05:09the environment variable which is the
4:05:11our edited web hook and then we have the
4:05:13actual web hook to hit. So this is one
4:05:15of those examples like I was talking
4:05:16about earlier where we have an
4:05:17environment variable that only gets
4:05:19called when we need it. And so the
4:05:21reason why Cloud Code built it like this
4:05:22is because they didn't want anyone to be
4:05:24able to look at the GitHub repo and find
4:05:27our URL for our NN web hook. Now, in
4:05:30this case, I don't really care because
4:05:31if I have a web hook, I could set up my
4:05:33own authentication so that only people
4:05:35that make an account can actually talk
4:05:36to it. But it's important to think about
4:05:38because if I gave you guys this URL and
4:05:41I didn't have any, you know, web hook
4:05:42security, someone could spam it
4:05:44thousands of times and it would cost me
4:05:46money because it's on my NN with my
4:05:48OpenAI credentials or my open router
4:05:50credentials. So in this case, what you
4:05:52would do is you'd add the environment
4:05:53variable. So you would come into where'd
4:05:56my Verscell go? So we'd actually be able
4:05:57to add that by going in here. We'd go to
4:05:59our settings. We could scroll down here
4:06:01to environment variables and we would
4:06:03basically just add one and we would have
4:06:06the key be nadn web hook URL and then we
4:06:10put in the actual URL and that way
4:06:12versell would understand to call on it
4:06:14when we hit that button. But in this
4:06:16tutorial I just want to keep things
4:06:17pretty simple. So what I'm going to do
4:06:18is just go back to cloud code and just
4:06:20tell it don't do the environment
4:06:22variable just hardcode my web hook in
4:06:24the code because I also wanted to just
4:06:26show you guys how we can push that
4:06:28change instantly to forcell I don't want
4:06:30to use my web hook as an environment
4:06:32variable please just change the code so
4:06:35the web hook URL is hardcoded in there
4:06:38it's going to make it much more simple
4:06:40okay so now the web hook URL has been
4:06:42hardcoded into the code and we have to
4:06:44redeploy that so I'm going to import the
4:06:46changes to GitHub you You can see in my
4:06:49forcell it's rebuilding this real quick.
4:06:52And in GitHub if I refresh you can see
4:06:55that now there are two commits. So every
4:06:57time that you change the code and you
4:06:58push it to GitHub it will have another
4:07:01version here. So that way you can see
4:07:02what happened each time. And now you can
4:07:04see that this has been redeployed onto
4:07:06our um app in Verscell. So I'm going to
4:07:10go to the new landing page and we're
4:07:11just going to try to ask another
4:07:12question. So um design a beginner
4:07:14strength routine. And this time it's
4:07:17actually working. It looks like it's
4:07:18writing back some sort of response
4:07:20because it did hit our end web hook.
4:07:22That's at least the hypothesis here.
4:07:25Cool. So now it responded with much more
4:07:27natural language, which is cool. You
4:07:28know what I think actually happened last
4:07:29time is maybe we just didn't like
4:07:31publish the most recent version. But
4:07:33anyways, in the fitness coach end
4:07:35workflow, I'm going to refresh this.
4:07:36We're going to see if we got that
4:07:38execution, which is right here. And we
4:07:40know it worked in this case because the
4:07:42actual post request that got sent over
4:07:44was design a beginner strength routine,
4:07:47which is exactly what we just said in
4:07:49our app right here. And I would also
4:07:51feel a little bit bad if one of you guys
4:07:53followed this tutorial and then woke up
4:07:54with like thousands of, you know,
4:07:56credits spent on your account. So
4:07:58another thing you can do is they have
4:07:59like a security review function in Cloud
4:08:01Code. And obviously claude code with
4:08:04Opus 4.5 is going to be really really
4:08:06good at like reading your code and
4:08:08understanding if there's vulnerabilities
4:08:10which is something that you should you
4:08:11know regularly just say hey by the way
4:08:13like what should I be aware of and what
4:08:15like risks do I have? So I basically
4:08:17said here do a full security review of
4:08:19anything that I've pushed to GitHub to
4:08:21make sure that I don't have any
4:08:22credentials exposed online because my
4:08:24GitHub repo is public and my workflow is
4:08:26you know out there too. So, it searched
4:08:28through everything and it said, "Okay,
4:08:30cool." So, your hard-coded ended web
4:08:32hook is out there, which I told it to do
4:08:34that. It's fine. I understand that.
4:08:37And that basically just means that
4:08:38people could see this and then hit the
4:08:39web hook directly. So, that is a
4:08:42problem. But if you set up your own
4:08:43authentication on the back end or
4:08:44obviously this is just a demo, so I'm
4:08:46not too worried.
4:08:48And the recommendation would be to move
4:08:50this to an environment variable. And
4:08:52then it also talked about my
4:08:53credentials, which is something I
4:08:54brought up to you guys earlier, which is
4:08:56the fact that those are all stored in
4:08:58this MCP file. So in this JSON file, I
4:09:01have my credentials for GitHub and Nitn.
4:09:03So if someone got this file, that could
4:09:05be a big problem, too. But this file is
4:09:07stored locally. So I basically said to
4:09:09it, hey, why do I have to rotate my
4:09:10credentials? Aren't they safe for my
4:09:12local environment? And it basically
4:09:13said, yeah, you're right. Your
4:09:14credentials are safe. I was being overly
4:09:16cautious because this file is not in the
4:09:18GitHub repo. It's not online. it only
4:09:21exists locally on your machine. And I
4:09:22know that this project itself right
4:09:24here, like this app, isn't anything too
4:09:26complex or like super impressive. But
4:09:28the idea is now that you understand like
4:09:30this whole framework and how everything
4:09:32works together, you can continuously
4:09:34iterate upon this and you can maybe even
4:09:35add more end workflows on the back end
4:09:37using the help of cloud code, fixing
4:09:39things on the front end, adding
4:09:40different functionality, pushing it back
4:09:42to GitHub, and then it continuously gets
4:09:44better. Cuz I mean, think about it with
4:09:45something like this, it's really your
4:09:46imagination because you can control what
4:09:48you want on here. Maybe you want
4:09:50somewhere where they can upload progress
4:09:51pictures or a food logger or a workout
4:09:54logger, things like that. And then by
4:09:56default when you deploy something on
4:09:57Verscell, it'll have the domain that
4:09:59ends inversel.app.
4:10:01And so you will want to buy a domain
4:10:03somewhere else or just buy one right
4:10:04here in Verscell and connect it. And
4:10:06it's pretty simple. You would pretty
4:10:07much just click on this plus right here
4:10:08and you could either buy a domain or add
4:10:10a domain. You'll have to do something
4:10:11with the DNS records if you're
4:10:13transferring in a domain from like
4:10:15NameCheep or Squarespace or wherever you
4:10:17bought the domain. But it's super
4:10:18simple. You just have to go in there and
4:10:19change like an A record and it may seem
4:10:22a little confusing, but just have chatbt
4:10:24or Gemini or Claude walk you through it.
4:10:26It's really easy. Or Claude Code. Just
4:10:28have Claude code tell you how to do it.
4:10:30And then the last thing I had to talk
4:10:31about cuz I know there's going to be
4:10:32tons of comments about this is the
4:10:34Claude plan. So yes, you could 100%
4:10:36start on Pro because you do get access
4:10:38to Claude code if you're on the pro
4:10:40plan. But I will say you probably will
4:10:42reach your limit pretty quickly. But I
4:10:44really wouldn't stress this. It's just
4:10:45one of those things where Start on Pro.
4:10:47If you hit your limit, okay, go upgrade
4:10:49to the $100 a month plan. If you hit
4:10:50your limit on the $100 a month plan,
4:10:52upgrade to the $200 a month plan. And I
4:10:54know $200 a month sounds expensive, but
4:10:56if you think about how much you can do,
4:10:57and if you think about how much would it
4:10:58cost for me to pay like a top tier
4:11:00developer to do this kind of stuff, it
4:11:02is significantly more than 200 bucks. So
4:11:04that 200 is going to be a huge bang for
4:11:06your buck. Once again, if you're on 200
4:11:07a month and then you realize you're not
4:11:09using it all the way and it's not worth
4:11:10it, then just downgrade. It's not going
4:11:12to be permanent. Okay, so we just
4:11:14covered a ton of information. So let's
4:11:16just recap what we just did. First of
4:11:18all, what we did is we connected cloud
4:11:20code to the end MCP server to look
4:11:22through how NIDin works and to be able
4:11:25to go into our instance, get workflows,
4:11:27create them, edit them, publish them,
4:11:29all that kind of stuff. And then we gave
4:11:31it access to the end skills so it
4:11:32actually knows how to use that server
4:11:34and how to build endto-end workflows.
4:11:37After that, we were able to have Cloud
4:11:39Code optimize the workflows to be ready
4:11:41for a front-end deployment. So, we also
4:11:44gave it access to a front-end designer
4:11:45skill. We gave it access to the GitHub
4:11:47MCP so that as we're building this web
4:11:49app and we're hosting it and testing it
4:11:51locally once we have that code ready to
4:11:53go, then we push it to GitHub which
4:11:55automatically syncs with Verscell and
4:11:57then deploys it on the web. So, now
4:11:59other people can actually access your
4:12:00app. It's not just on your own local
4:12:02environment. And then of course because
4:12:04of this whole framework, as soon as you
4:12:06update the code and push it to GitHub,
4:12:08it automatically will update on the web.
4:12:10And now that you're already in this
4:12:11project called an event app or whatever
4:12:13you called yours, if you wanted to, you
4:12:15could just do another workflow in this
4:12:16project because you already have the
4:12:18cloud MD file. You already have your MCP
4:12:20server set up. You already have your
4:12:21skills set up globally. You have this
4:12:23folder right here, which is all the
4:12:24stuff you need for this app. But if we
4:12:26wanted to, we could obviously just come
4:12:28into here. We could clear out this
4:12:30conversation history and we could say,
4:12:31"Okay, cool. Now I want to build a front
4:12:33end for this other app and just go
4:12:35through that whole process again. And
4:12:37because all of this is stuff that it can
4:12:38look through, it could maybe even take
4:12:39inspiration from other apps that you've
4:12:41done in the past.
4:12:44Today I'm going to be showing you guys
4:12:46five simple hacks that you can use to
4:12:48make sure that Claude Code is building
4:12:49you websites that don't look like they
4:12:51were AI vibe coded, but they actually
4:12:53feel professional and branded. And we're
4:12:55going to be going through this in a way
4:12:56where even if you've never used Claude
4:12:57Code before, that's completely fine.
4:12:59you're going to be able to by the end of
4:13:00this video spin up some really awesome
4:13:02looking landing pages and websites. All
4:13:04right, so I don't want to waste any time
4:13:05at all. The first thing that you need to
4:13:06do is you need to go download Visual
4:13:08Studio Code. So go to a browser and type
4:13:09in VS Code and download this for your
4:13:12operating system. This is essentially
4:13:13just the IDE that we're going to be
4:13:15using Cloud Code within. So once you've
4:13:17done that and you've opened it up, this
4:13:19is what it will look like. You're going
4:13:20to go to the lefth hand side right here
4:13:21and click on extensions and you're going
4:13:23to type in cloud code and install it
4:13:25like what you see right here. Now, once
4:13:27you do that, it's going to prompt you to
4:13:29sign in with your Enthropic subscription
4:13:30or your cloud subscription, which you do
4:13:32need a paid account. As you can see
4:13:34here, if you're on free, you don't have
4:13:35access to Cloud Code. But here on Pro,
4:13:37you actually can use Cloud Code. Whether
4:13:40you're on pro or max, you can use it.
4:13:42I'd probably just start with Pro. If you
4:13:43hit limits, which you probably will if
4:13:45you want to, you know, build websites
4:13:47all day, then you should probably
4:13:48upgrade to max. So, once you've got that
4:13:50installed, you will see this little
4:13:52button up here, which is Cloud Code. And
4:13:54when you click on that, this is where it
4:13:56opens up the ability to actually use
4:13:57cloud code, talk to this little crab
4:13:59agent. And this is very similar to sort
4:14:01of like a chatbt or using cloud in the
4:14:03web. Now, the way that this works when
4:14:05you're using cloud code in Visual Studio
4:14:06Code or really wherever you use it is
4:14:08you have files on the lefth hand side
4:14:10and then you have your agent on the
4:14:11right hand side. So, first thing we need
4:14:13to do is open up a project so that we
4:14:14can start working with some files. So,
4:14:16I'm going to go up here to the top left
4:14:17and I'm going to click on explorer. What
4:14:19you can see is that it says you have not
4:14:21yet opened a folder. So, I'm going to go
4:14:22ahead and open up a fresh folder that
4:14:24has nothing in it. So, here we are in my
4:14:26website building YouTube folder, which
4:14:28like I said, it's a blank project. If
4:14:30you don't have a folder, just go ahead
4:14:32and create one. Whether that's in your
4:14:33desktop or your documents, just create
4:14:35one to start and then open that up. And
4:14:37that is where we will be working on this
4:14:38project. So, let's get started going
4:14:40through these five hacks. The first one
4:14:42is actually number zero. And the reason
4:14:43that I did this is because the first one
4:14:45is a claw.md file. And I put this as
4:14:48number zero because it's kind of a
4:14:49prerequisite, but also a lot of times
4:14:51near the end, even after 1 2 3 and four,
4:14:53you might have to rego back and update
4:14:56your cloud.MD file or just have Claude
4:14:58do it itself. So what is a claw.md file?
4:15:00Just think of it as a system prompt.
4:15:02Think of it as every time before you ask
4:15:05claude code to do something, it will
4:15:07read the claw.md file first. It will
4:15:10always process that. So what you want to
4:15:12do is make sure that that is pretty
4:15:13concise. You don't want to bloat it too
4:15:15much with context, but you want to give
4:15:16it the rules that it needs. So, every
4:15:18time you are doing something in this
4:15:20project, this website building project,
4:15:22do this, this, and this. And always
4:15:23remember that's kind of the end goal.
4:15:25And so, if you don't exactly know your
4:15:26full process yet or the end goal, then
4:15:28you might start without a claw. MD file.
4:15:30But luckily for you guys, if you go over
4:15:31to my free school community, the link
4:15:33for that's down in the description, you
4:15:34go to the classroom, you go to cloud
4:15:36code, and right here you will see the
4:15:38web designcloud.mmd file, which is the
4:15:40one we're going to be using today. You
4:15:41can go ahead and just download that for
4:15:42free right here. Now, once you've done
4:15:44that, you can actually just drag it
4:15:45right over here to the lefth hand side.
4:15:47Like I told you guys, the lefth hand
4:15:48side is where we can see our files and
4:15:50our folders. And what that does is it
4:15:51opens up the claw.md file, which if I
4:15:53drag over here, we can see it kind of
4:15:55full screen. Now, the MD stands for
4:15:57markdown, which is basically just this
4:15:59right here. We've got the pound signs,
4:16:01we've got um asterisks, and it just
4:16:03helps keep the text organized so that
4:16:05the agent can read, you know, what's a
4:16:06header, what's a subheader, what's bold,
4:16:08what are bullet points, things like
4:16:09that. So, you could obviously read
4:16:11through this entire claw.md file if you
4:16:12want to to kind of understand what we're
4:16:14telling it to do in this project. I'm
4:16:16not going to read everything because you
4:16:17guys can just, you know, look at it here
4:16:19or download it. And as we go through
4:16:20these other hacks, you will see why I
4:16:21put some of this stuff in here. But that
4:16:23actually brings me over to our first
4:16:25technically our first hack, which is the
4:16:27front-end design skill, which is why you
4:16:29can see right here in our cloud. MD, the
4:16:31first thing I wrote is always invoke the
4:16:33front-end design skill before writing
4:16:35any front-end code every session. No
4:16:37exceptions. So, first of all, real
4:16:39quick, what are skills? Well, if you go
4:16:41to the cloud code docs, you can read
4:16:42about skills right here. Essentially,
4:16:44they are custom instructions. So, every
4:16:46time you build like a custom GBT or
4:16:47cloud project, you're usually putting in
4:16:49knowledge and you're putting in
4:16:50instructions. And basically, skills are
4:16:53just that, but in a markdown file. And
4:16:55why it's so important and cool is
4:16:57because every time you ask Claude a
4:16:58question, first it reads its cloud.mmd
4:17:01file, but then it will think, okay, the
4:17:03user asked me this. Do I have any skills
4:17:05in my library that help me do this
4:17:06better? If yes, I'll grab the skill.
4:17:08I'll read it and then I'll take action.
4:17:10If no, I'll just use my general
4:17:12knowledge. So, that's why we need to
4:17:14have the front-end skill because it
4:17:15helps us create designs that are way
4:17:17more modern and professional and they
4:17:19don't look as much vibecoded, AI
4:17:22vibecoded. And the good news is it's
4:17:24super super simple. You just have to
4:17:26install it. So, here's a tweet that
4:17:27showed the power of this. All they
4:17:29prompted Cloud Code to do was use the
4:17:31front-end design skill, create a music
4:17:32player app, and it created this that has
4:17:35some, you know, animations, it has some
4:17:36dynamic elements, and if you would have
4:17:38just told Claude Code to do this without
4:17:39that skill, it would have looked much
4:17:40worse. So, I'll leave a link to this
4:17:42tweet in the description of this video,
4:17:43you basically just have to run this
4:17:44command and then you run this one, and
4:17:46then you should be good with the skill
4:17:48installed globally across any Cloud Code
4:17:50project that you might use in the
4:17:51future. And when I say run these
4:17:53commands, you can literally just copy
4:17:54this if you wanted to and just paste
4:17:56that right into here in cloud code and
4:17:58it would install that for you. All
4:17:59right, so let me go ahead and show you
4:18:01guys how good this front-end design
4:18:03skill really is with such a minimal
4:18:05prompt. So before we prompt this agent,
4:18:06I just wanted to show you guys something
4:18:07else you can do, which is kind of a
4:18:09bonus hack. What I'm going to do is I'm
4:18:10going to create a new folder. I'm going
4:18:12to call this brand_assets.
4:18:15And our claw.md file actually explains
4:18:17that this might be a file or a folder
4:18:19that cloud code needs to look at. And
4:18:21what I'm going to put in here are two
4:18:23things. My logo and brand guidelines so
4:18:27that it creates this website and it
4:18:28feels very branded towards me and my
4:18:30business. So right here I'm dragging in
4:18:32the AI automation society logo as you
4:18:34can see like that. And then I'm also
4:18:35going to drag in our brand guidelines
4:18:37which has stuff like our colors, our
4:18:39typography icons, stuff like that. And
4:18:42so now that Claude can look at that, I'm
4:18:44going to just give it a very very simple
4:18:46prompt. So, all I'm saying is build me a
4:18:48modern and professional landing page for
4:18:50AI Automation Society. And I'm also
4:18:51going to tell it that here's my logo and
4:18:53here's my brand guidelines. It would be
4:18:54able to figure it out either way because
4:18:56we put it in the claw.md. But I just
4:18:58wanted to show you guys that you can
4:18:59actually tag assets directly. So, if I
4:19:01do an at, it will basically pop up and
4:19:03let me choose or point at the right
4:19:05things. So now I can explicitly say,
4:19:07"Hey, here are the, you know, here's the
4:19:09brand guidelines and here's the logo
4:19:11because maybe they're not named in a way
4:19:13that's super intuitive and now I'm just
4:19:15showing Cloud Code exactly what I want."
4:19:17So I'm going to shoot this off. I'm not
4:19:18even in plan mode. I just want to show
4:19:20you guys how good this front-end design
4:19:22skill is. And what you're going to
4:19:23notice is first of all, what it did is
4:19:24it read the clawenmd file and now it's
4:19:27reading the brand assets. And now what
4:19:29it's going to do is it should hopefully
4:19:31invoke the front-end design skill and
4:19:33start building out that website for us.
4:19:35There we go. right on Q. It has invoked
4:19:37the front-end design skill right there.
4:19:40All right, so that has finished up. You
4:19:41can see that we've got a nav, a hero,
4:19:43tools marquee, we've got stats, about
4:19:44benefits. So, a full onepage landing
4:19:46page, and it should be completely
4:19:48matching our brand as far as the logo,
4:19:50the colors, and the typography. It also
4:19:52added some animations. So, I'm excited
4:19:54to see how that works. And it threw it
4:19:55on local host for us to check out. So,
4:19:57let's head over there. All right, look
4:19:58at that. We've got like a little
4:19:59animation up here. We've got a line
4:20:01going down. We can see that we do have
4:20:03our logo up here as well as our exact
4:20:06colors and font. We've got a community
4:20:08rating. Oh, that's super nice. We've
4:20:10also got some scrolling tech companies.
4:20:13So, we've got edit end make, Claude,
4:20:14GBT40, Zapier, Air Table. We've got some
4:20:17random stats here. Obviously, we'd have
4:20:18to fill this in with our own copy, but
4:20:20keep in mind all of this happened with
4:20:23only us saying, "Create me a landing
4:20:25page for our community called A
4:20:26Automation Society." That was literally
4:20:27it. And it created all of this. We've
4:20:29got testimonials. We've got a final call
4:20:31to action here. The logo is doing a
4:20:33little floating for basically a one
4:20:34sentence prompt. This is super super
4:20:36solid with the front-end design skill.
4:20:38Now, there was another secret thing
4:20:39going on here that I didn't yet tell you
4:20:41guys about, but if you've already read
4:20:42the Claw Denm, you might have noticed.
4:20:44And that brings us on to hack number
4:20:46two, which is the screenshot loop. So,
4:20:49the idea here is that AI is really good
4:20:52at getting you where you want to go, but
4:20:53it takes a lot of manual correction and
4:20:56steering. So, let's say I just told
4:20:57Claude Code to build us that website.
4:20:59Without the front-end design skill, it
4:21:01might have gotten us like 40% of the way
4:21:02there. But now that we added the
4:21:04front-end design skill, it's going to
4:21:05get us maybe let's let's just call it
4:21:0660. What we can do now is use
4:21:09screenshots to help AI iterate upon
4:21:11itself. So, instead of it getting 60% of
4:21:13the way there and then we make an
4:21:14improvement and then we make another
4:21:15improvement and we keep doing this, it
4:21:16basically should just bridge this gap
4:21:18itself because it's able to take a
4:21:20screenshot, look at the browser, see
4:21:21what it looks like, and then make make
4:21:23changes. So, what you guys didn't
4:21:24notice, or maybe you did, is over here,
4:21:26it created a new folder for us called
4:21:28temporary screenshots. And we can see
4:21:30that in that process of building out
4:21:31that first version of our workflow, it
4:21:33took 10 screenshots. So, I can click
4:21:35here, and I can see what it looked at.
4:21:37It looked at the hero section, which
4:21:39kind of was a a random full page. It got
4:21:42the viewport, which was that's more of
4:21:43the hero section. It looked at the
4:21:45stats. It looked at the about page. And
4:21:46what it did is it used these screenshots
4:21:48as it kept clicking through and looking
4:21:51and improved things. So, you guys didn't
4:21:53see this, but in the actual to-dos, it
4:21:55wrote the index. HTML, it started the
4:21:58server and screenshotted the workflow,
4:21:59and then it did a two pass screenshot
4:22:01review and polish. So, it basically uses
4:22:03its eyes to check that what it's
4:22:04building actually looks good. And in
4:22:06order to set that up, it's actually
4:22:07really, really easy. If you go to the
4:22:08cloud.md file, you can see that I've got
4:22:10a section for screenshot workflow. And
4:22:12we're just doing this using Puppeteer.
4:22:13So, literally, if you take this cloudMD
4:22:15and say, "Hey, Cloud Code, can you set
4:22:17up Puppeteer to take screenshots?" It
4:22:19should be able to install all of that
4:22:21stuff for you right there really simply.
4:22:23And so, yes, that's cool on its own, but
4:22:25where it actually comes into handy a lot
4:22:27more is when we look at hack number
4:22:29three, which is using other websites as
4:22:32inspiration. Because what we're able to
4:22:34do is say, "Hey, Claude Code, take this
4:22:36website right here and build me a
4:22:37clone." So, you should build one that
4:22:39looks exactly like this one. And then
4:22:40what it's able to do is use its eyes,
4:22:42use its screenshot tool to screenshot
4:22:44what it's building and look at the
4:22:45reference and keep going back and forth
4:22:47until it's close enough. So, let me show
4:22:49you guys that in action right now. So,
4:22:51there's tons of sites that you could go
4:22:52to for website inspiration. Here's one
4:22:55example called Dribble. Here's another
4:22:56example called godly website. And here's
4:22:59another really cool example called
4:23:00Awards with three W's. So, there's tons
4:23:03of places that you can find inspiration.
4:23:05So, for the sake of this video, I found
4:23:07this one that I want to use. It's got a
4:23:09nice little animation in the background.
4:23:10It's obviously not our color scheme, but
4:23:12it has some cool things as you scroll
4:23:14down like a dashboard. It's got some
4:23:16other little cards down here. None of
4:23:17this is really too animated. Well, I
4:23:19guess that is. But let's just say we
4:23:20wanted our website to look like this
4:23:22one, for example. First thing that I
4:23:23would do is I would hit F12. I'm on
4:23:25Windows, by the way. I would go to
4:23:27console and I would do control shiftP
4:23:30and search for screenshot. What this
4:23:32lets me do is capture a full-size
4:23:34screenshot of the entire page rather
4:23:36than just my current view. So here you
4:23:38can see it downloaded this screenshot
4:23:40and you can see that that is indeed the
4:23:42entire website. Now, if you're on Mac,
4:23:44that's still doable, but you just have
4:23:46to Google the different buttons to do
4:23:48it. And then the next thing what I want
4:23:49to do is on the top right here, I'm
4:23:51going to go to elements, and in the
4:23:53style section down here, I'm just going
4:23:54to copy everything. So, I'm actually
4:23:56copying basically like the raw code or
4:23:59HTML or, you know, whatever you want to
4:24:00consider this as that tells the website
4:24:02how this is styled. And we're going to
4:24:04give Claude code that. So, I'm going to
4:24:06go ahead and do a clear so we can start
4:24:08a fresh session. I'm going to first of
4:24:10all just paste in the code that we just
4:24:12copied which is the style information.
4:24:14So I said I want you to spin up a new
4:24:15website for us. Get rid of the old one
4:24:17and you can put this one on local host.
4:24:18I basically want you to clone this
4:24:20website. So I'm going to give you the
4:24:21screenshot which what I'm going to do is
4:24:23just drag it in from my files and put it
4:24:25right over here. As you can see that is
4:24:26the screenshot we just took. And I'm
4:24:28going to point to it so it knows what to
4:24:30use which is the www right there. And
4:24:33then I said here's a screenshot. Here's
4:24:35the style. and just go ahead and clone
4:24:36this website for us. So, that is all
4:24:39we're going to do to start and then we
4:24:41can come back in later and tell it to
4:24:43use our branding and our, you know,
4:24:45colors and logo and everything like
4:24:46that. Now, a couple things to keep in
4:24:47mind. When you're doing some of the big
4:24:49processes like spinning up a website
4:24:51from scratch or comparing two websites
4:24:53and cloning them, that coding process
4:24:55and thinking will take longer. But once
4:24:58you have a working version, making small
4:25:00changes or tweaks, that happens pretty
4:25:02quickly. And one other thing is you
4:25:04might have noticed that this really
4:25:05isn't stopping to ask me questions. And
4:25:07that's because I'm using bypass
4:25:08permissions mode. So if you don't see
4:25:10this in your instance, you're going to
4:25:11go to settings. You're going to type in
4:25:13clawed code. And then right here, you
4:25:15should see allow dangerously skip
4:25:17permissions. And that is where you turn
4:25:19that on. Now I definitely have a
4:25:20responsibility to tell you that this is
4:25:23dangerous. It has the potential to just
4:25:25kind of like run any command that it
4:25:26wants. But in my practice, I've never
4:25:29really had this be an issue, especially
4:25:31because I'm never like setting this to
4:25:33code all night long and then going to
4:25:34sleep. I'm always still kind of like
4:25:36watching it or I'm nearby and nothing
4:25:39bad really is going to happen. All
4:25:41right, awesome. So, we just got to the
4:25:42point where now it is creating a to-do
4:25:44list. And what you can see here is once
4:25:46it actually writes the code for the
4:25:48website, it's going to start up the
4:25:49server and it's going to take
4:25:50screenshots and it's going to do two
4:25:52rounds at least of comparing. It's going
4:25:54to look at what it built versus the
4:25:56reference. It's going to fix any
4:25:57mismatches and then it's going to do
4:25:58that again. And that is why the
4:26:00screenshot loop is so powerful. So
4:26:02logically, this is really cool. I mean,
4:26:03it's going through and it's looking
4:26:04section by section and analyzing how
4:26:07well it's stacking up. But we will have
4:26:08to see how it actually turns out. Okay,
4:26:10so that just finished up and before we
4:26:11actually see how good it really built
4:26:13this, I wanted to point out one thing
4:26:15about the screenshots. So you can see
4:26:16that we have screenshot 1 2 3 4, all of
4:26:19this kind of stuff. But we don't really
4:26:21know which one is which without clicking
4:26:22on them. So, it looks like these are the
4:26:24clones as you can see because they're
4:26:26coming out looking like the website that
4:26:27we gave it. Well, we either should have
4:26:29before we started this new build, we
4:26:31should have told cloud code, hey, you
4:26:32can delete all of those temporary
4:26:34screenshots or in the claw.md, we should
4:26:37be more specific about the naming
4:26:38convention of the screenshots so that we
4:26:40can actually tell. Now, realistically,
4:26:41these temporary screenshots are more for
4:26:43claw codes benefit than for ours, but
4:26:45that is something else that you can be
4:26:47thinking about if you do want to be able
4:26:48to click through and see the changes
4:26:49that were made with each version. But
4:26:51anyways, let's go ahead and open up this
4:26:52link and see what we got. All right, so
4:26:54I'm going to switch this to English for
4:26:56my head. But we can see we've got the
4:26:58purple colors. We've got your strategic
4:27:00ally for a truly profitable business.
4:27:02We've got the same top menu bar. Um a
4:27:04similar type of dashboard here. We've
4:27:05got some cards. And as we scroll down,
4:27:08it feels very similar to the real
4:27:10version that we gave it, which was this
4:27:12one. Obviously, some of the dynamic
4:27:14elements in the background and some of
4:27:15the actual images could not have been
4:27:17the exact same, but for a clone, this is
4:27:20very, very similar. And it is a really
4:27:23good spot for us to actually start. And
4:27:24now we can just start to integrate our
4:27:26own colors and logos and copy right into
4:27:29this template. And it's as simple as
4:27:31just asking it to do so. So, I'm going
4:27:33to go ahead and clear this out. I'm
4:27:35going to say go ahead and delete all of
4:27:37the temporary screenshots in the
4:27:39temporary screenshots folder. And so now
4:27:41all of those have been deleted as you
4:27:42can see. And we're basically going to
4:27:43say the most recent landing page looks
4:27:46really good. What I want you to do now
4:27:48is work in our brand assets. So our
4:27:51brand guidelines and our AIS logo. And
4:27:55this is for our community called AI
4:27:58Automation Society. So just work in
4:28:00those changes to that website clone that
4:28:02you just built. And once again, we are
4:28:04just going to stay on bypass
4:28:05permissions. I'm going to shoot that
4:28:06off. One shot prompt this thing. And
4:28:08hopefully we should get something that
4:28:10looks pretty solid. Now, what I'm
4:28:11interested to see is what it ends up
4:28:13doing with this dashboard and what it
4:28:14ends up doing with this iPhone screen
4:28:16because we haven't given it any other
4:28:18pictures. As you saw in our website, we
4:28:20obviously gave it some different
4:28:20pictures like the school games dashboard
4:28:22or me with Hermosi and Sam Ovens. But
4:28:25that's what you could do is you would
4:28:26come back into Cloud Code and you would
4:28:27say, "Hey, I gave you some more pictures
4:28:29in the brand assets. Put this one here.
4:28:31Put this one there." And it would figure
4:28:32that out for you. And of course, you
4:28:34would also have to say, "Cool. When they
4:28:35click on start for free, take them to
4:28:37this link." or when they click on see
4:28:38the demo, take them to this link. So,
4:28:40there's other little pieces that you
4:28:41would obviously have to configure as
4:28:43well, but those changes take basically
4:28:44no time. Okay, so that finished up
4:28:47pretty quickly. We've got three
4:28:48screenshots here, but I'm not going to
4:28:49click into them because I don't want to
4:28:50ruin the final reveal here. But we used
4:28:53our colors. We have our primary accent,
4:28:55our secondary, our dark background, and
4:28:57our mid background. We've got the right
4:28:58typography. We've got the right logo,
4:29:00and everything was fully translated from
4:29:02French to English. Thank goodness. And
4:29:04now it's rewritten for our community,
4:29:05which once again, we didn't actually
4:29:07give it facts about the community yet.
4:29:08This is just very simple prompting. It
4:29:10also mocked up a dashboard. So, let's
4:29:12head over to our local host. Let's give
4:29:14this a hard refresh. And boom. We now
4:29:16have our new site, master a automation,
4:29:18build faster, earn more. For the
4:29:20dashboard, it worked in like a little
4:29:22bit of a it's got members. It's got
4:29:23automations, courses. It's got it's kind
4:29:26of like a community tracker dashboard,
4:29:28and it uses our colors in there, too,
4:29:29which is cool. We've got different
4:29:31things on here, workshops, templates,
4:29:32expert community. It also changed this
4:29:35iPhone thing to member growth this
4:29:37month. So, it's keeping all of this on
4:29:38brand with the actual original reference
4:29:41site, which once again looked like this.
4:29:43However, now it has our colors and it
4:29:45has our information in here. We've got
4:29:47two paths and then we have some other
4:29:49stats down here and a nice little call
4:29:51to action at the bottom. So, cool. What
4:29:53we could do now is obviously go back and
4:29:54forth a little bit, maybe change some
4:29:56text, make things bigger, you know,
4:29:57change the images and stuff like that.
4:29:59But let's say we're at a spot where we
4:30:00like the overall feel and vibe of the
4:30:02website. But now, how do we really up it
4:30:04to the next level to make it feel
4:30:05unique? Well, what we're going to do is
4:30:07unlock the final hack, which is
4:30:10individual components. And what I mean
4:30:12by that is taking inspiration from
4:30:13different places, but for very
4:30:15individual components, for small pieces,
4:30:17not entire websites. So, what we can do
4:30:19is we can go to a website called
4:30:2121st.dev,
4:30:23which has some of the best website
4:30:25components you might be able to find.
4:30:26It's got shaders. It's got backgrounds.
4:30:28It's got home screens. It's got buttons.
4:30:30It's got, you know, mouse highlights.
4:30:32It's got so many different things that
4:30:33you can do. So, here you can see I've
4:30:34got buttons and I could make them have a
4:30:36rainbow outline. I could make them
4:30:37shiny. We could toggle, you know, dark
4:30:39mode or light mode. There's lots of
4:30:40different things we could do here. Or I
4:30:42could just click on backgrounds in here
4:30:43and I could look at other ways that we
4:30:44could have our background. So, maybe we
4:30:46want these little kind of drop down
4:30:47pills instead. Or maybe we want these
4:30:49hero waves in the background. I think we
4:30:51should actually do this instead. So,
4:30:52what I'm going to do is just copy this
4:30:53prompt right here. This will basically
4:30:55copy a chunk of code for us to give to
4:30:57Claude Code. And I'm just going to say,
4:30:59I want you to work in this background
4:31:02element right behind the hero text. And
4:31:04after I give it that prompt, I just
4:31:06paste in what we grabbed from 21st. And
4:31:08it should be able to use all of this and
4:31:11understand how to put that into our
4:31:13site. So, I'm just going to go ahead and
4:31:15shoot this off and we will see.
4:31:16Actually, one thing that I forgot to
4:31:18mention is in this case, because we're
4:31:21working with an animation, the
4:31:22screenshot might not always work the
4:31:24best. So, sometimes you might want to
4:31:26tell it not to do the screenshot flow.
4:31:28So, I'm basically actually just going to
4:31:29copy all of this text once again. I'm
4:31:32going to clear this out. I'm going to
4:31:33paste it back in. But then I'm also
4:31:35going to say
4:31:38because this is an animated background,
4:31:40do not use the screenshot tool to
4:31:42compare. just work in the code and then
4:31:44I will let you know if we need to make
4:31:45any changes. So hopefully with that
4:31:47mentioned even though it's going to read
4:31:48the claw.mmd it won't do a bunch of
4:31:50screenshots here because I've actually
4:31:52tested this out and I've had you know
4:31:53different background elements come
4:31:54through and because they're dynamic
4:31:56sometimes the screenshot doesn't fully
4:31:58capture it. So it gets stuck in this
4:31:59loop of thinking I haven't built this
4:32:01good enough. I'm going to keep trying
4:32:02and it like overengineers and it just
4:32:04doesn't really work. So sometimes you
4:32:05may want to turn off the screenshot
4:32:08tool. All right, so that just finished
4:32:09up. It didn't take a bunch of
4:32:10screenshots, so it didn't take forever.
4:32:12Let's go to the website. Let's give it a
4:32:14refresh and see. Okay. Okay. So, we've
4:32:18got a background. It looks a little bit
4:32:20um distracting. It also looks a little
4:32:22bit cheap. It looks like too pixelated.
4:32:24So, what I'm going to do now is just
4:32:25iterate. I'm going to tell it that
4:32:28I think that it's a little bit
4:32:30distracting as far as it makes the hero
4:32:33text right behind it a little bit tough
4:32:35to read. Also, in the hero text, I'd
4:32:38like it if the earn more was maybe a
4:32:41blue or a different color. I think that
4:32:43doesn't really feel good to have that be
4:32:44orange. It would be good if there was
4:32:46maybe some sort of background behind the
4:32:48hero text so that we could see it and it
4:32:50would still stand out and contrast
4:32:52against the background animation, but
4:32:54the background animation looks super
4:32:56fuzzy and super pixy. If you could make
4:32:58that look a little bit more professional
4:33:01and clean, that would be great. And if
4:33:03you guys were curious why I was just
4:33:04like staring at that and talking is
4:33:05because I was dictating and I wanted to
4:33:07be able to look at what I was talking
4:33:08about. So we've given some feedback now.
4:33:11Let's see if it can go ahead and make
4:33:13those changes. And once again, like
4:33:14we're being pretty vague here and it
4:33:17would be up to the creativity of the
4:33:18model to understand what we're asking
4:33:20for and be able to make these changes.
4:33:22Now, if you were on plan mode, it might
4:33:24be able to do a little bit better job of
4:33:25asking you some questions and maybe
4:33:27helping you get to a better solution
4:33:29first before it starts coding. But for
4:33:31the sake of the video, let's see how
4:33:32well it does with this prompt. All
4:33:34right, that just finished up. And you
4:33:36can see that that looks much, much
4:33:37better. This is definitely more what I
4:33:39was looking for when we copied over that
4:33:42animation into this website. So from
4:33:44here, we would just keep going through
4:33:46and we'd keep being really nitpicky
4:33:47about what we want to change. We'd add
4:33:48our own pictures in. We'd maybe want to
4:33:50change some of these buttons to be more
4:33:51dynamic. We'd want to maybe animate some
4:33:53of this other stuff, which we could
4:33:54easily do just by asking Claude Code to
4:33:56do so. So from here, the question is,
4:33:58how do you actually get this onto a real
4:34:00landing page? Because right now, we're
4:34:02still developing all of this code and
4:34:03we're previewing this in our local host.
4:34:06So what we're going to do is we're going
4:34:07to use a combination of GitHub and
4:34:08Verscell to do this. Cloud code is where
4:34:10we're working right now. All of these
4:34:12folders, all of these files are local,
4:34:13meaning if I pulled up my laptop, I
4:34:15wouldn't be able to access them. And
4:34:16when we're building our website, which
4:34:18is obviously this website right here,
4:34:20this is all made up of a bunch of code
4:34:22in our cloud code project. So what we
4:34:24need to do with that is we sync that
4:34:26code to GitHub and GitHub has version
4:34:29control. We can see all of our commits,
4:34:30other people can work on it, stuff like
4:34:32that. We basically host our code or our
4:34:34project in the cloud and we set up a
4:34:36really cool autodeploy between Verscell
4:34:38and GitHub. And Versell is basically
4:34:40just where we deploy our code to a live
4:34:43site. So basically what this means is
4:34:45whenever we tell cloud code, hey this
4:34:47looks good, push these changes to
4:34:48GitHub, GitHub grabs the new changes and
4:34:51then Verscell automatically grabs those
4:34:52from GitHub and then updates the real
4:34:55working version of our site. And I will
4:34:57show you guys that. But let's first of
4:34:58all do this pipeline. So the first thing
4:35:00that you're going to need to do is go to
4:35:03GitHub, create an account if you don't
4:35:04already have one, and you're going to
4:35:05need to create a new repository. So I'm
4:35:07going to create a repository right here
4:35:09called AIS test website. I'm not going
4:35:13to worry right now about a description
4:35:14or all of this and I'm just going to go
4:35:16ahead and create that repository. Now,
4:35:18what you also could do is you could tell
4:35:20Claude Code, hey, create me a GitHub
4:35:23repository and it could actually do
4:35:24that. But right now, I just wanted to
4:35:26show you guys so you can get a feel for
4:35:27GitHub if you've never used it before.
4:35:29So, anyways, now we have this repository
4:35:30called AIS test website. I'm just going
4:35:32to copy the name of that real quick and
4:35:34I'm going to come back into Cloud Code.
4:35:35We're going to clear this out and say
4:35:37awesome. So now that this site looks
4:35:39good, we need to actually deploy this on
4:35:41our domain. I need you to help push this
4:35:43to GitHub and we're going to push it to
4:35:46a GitHub repository called
4:35:49and then I'm going to paste in the name.
4:35:50Now, so far it has not yet gotten our
4:35:53GitHub credentials. So we're going to
4:35:54have to obviously authenticate into
4:35:56GitHub first so it can push that into
4:35:58GitHub. So I just got logged in as Nate
4:36:00Herkai and now it's going to create the
4:36:03ignore and get everything set up so it
4:36:04can actually do so. Now, it's not too
4:36:06big of a deal right now because nothing
4:36:07that we'd be pushing into the public
4:36:09GitHub or, you know, onto the cloud has
4:36:12API keys or has any usernames or
4:36:14passwords or any sensitive information
4:36:15or, you know, web hook abilities. But
4:36:17that is something to be aware of once
4:36:19you actually are pushing automations and
4:36:20things like that to the cloud. Make sure
4:36:22that you're not putting any of your
4:36:23sensitive information out there.
4:36:25Awesome. So, it now says that our site
4:36:27is live on GitHub. So, if I click into
4:36:29this link, we should see that we now
4:36:30have a new commit. We have all of this
4:36:32stuff like our claw.mmd. We have our
4:36:34screenshot stuff. We have brand assets
4:36:36and now we can sync this to Verscell. So
4:36:38that would be step two is you're going
4:36:39to go to versell.com create an account.
4:36:41When you create that account, it's much
4:36:43easier if you just sign in or create
4:36:44that account with your GitHub
4:36:46credentials. And then all we have to do
4:36:48is go ahead and add a new project and
4:36:50then we're able to just choose a GitHub
4:36:52repository. As simple as that. So I can
4:36:54literally just hit import on our AIS
4:36:56test website which you guys just saw me
4:36:57set up. And then all I have to do is go
4:36:59ahead and deploy this project. Awesome.
4:37:02So I've deployed a new project to my
4:37:03project. I can go ahead and continue to
4:37:05the dashboard here. And what this now
4:37:07does is we can actually visit this by
4:37:09going to ais-test-website.vercell.app.
4:37:13I open that up and now this is no longer
4:37:15local. I could open up my phone and type
4:37:17in this. You could open up your browser
4:37:18and type this in and you guys could all
4:37:20visit this site because it's now
4:37:22deployed on the cloud. But of course,
4:37:23it's got an ugly domain. So what you
4:37:25would have to do now is you would have
4:37:27to go to your project settings. You
4:37:29would go to domains. And then this is
4:37:31where you would actually just have to
4:37:32either buy a domain right here or add an
4:37:34existing one. And it's really simple. It
4:37:36would walk you through the DNS
4:37:37configuration that you need to set up.
4:37:39And it's not too difficult, but I'm not
4:37:41going to actually do that live in this
4:37:42video. So, what I wanted to show you
4:37:43guys real quick before we end off this
4:37:45video is what actually happens if we
4:37:46realize that we want to make a change to
4:37:48our website that is on the cloud. Well,
4:37:50that's why it's good that we still have,
4:37:52you guys can't see because you can't see
4:37:53the URL, but we still do have our local
4:37:55version because if I make a change here
4:37:57and I don't like it, I don't want that
4:37:59to automatically get pushed to um
4:38:01Verscell. So, what you'd probably want
4:38:03to do is in your claw.mmd file, you
4:38:05would say ultimately what's going to
4:38:06happen is we're syncing all of the
4:38:08changes to GitHub. GitHub's going to
4:38:09automatically push them to Verscell and
4:38:11we'll be good to go. But when I'm making
4:38:13changes with you here, we're always
4:38:14going to test on a local host until I
4:38:16tell you explicitly to push that to
4:38:18GitHub or commit those changes to
4:38:19GitHub. Okay. So, this is our local
4:38:22version. And let's just say, for
4:38:23example, we wanted to make this button a
4:38:25little bit cooler. So, I'm going to ask
4:38:27in Cloud Code, could you go ahead and
4:38:29make the join the community button in
4:38:32the main hero text section, make it more
4:38:34professional. So, give it like a cool
4:38:36glow. And once you've made this change,
4:38:38let me see it in local host. Don't push
4:38:41it to GitHub until I tell you to. This
4:38:42thing is getting pretty screenshot
4:38:44happy. I may have to adjust the wording
4:38:45in the cloudmd file a little bit. It
4:38:48literally took one of the main screen
4:38:49and then it took one of where it just
4:38:51cropped the actual button, but hey, it
4:38:53looks good. Okay, so what happens is
4:38:54here's the local host. I'll refresh
4:38:56that. Now we can see the little glow
4:38:58behind the join the community button and
4:39:00here is the web app version. I refresh
4:39:02this and we don't have that change yet,
4:39:03which is great because we don't want to
4:39:05push changes if they're not good, right?
4:39:07But now what I'll do is say awesome. I
4:39:09love that change. Go ahead and push that
4:39:10to GitHub. All right, so it just pushed
4:39:12that. We have a new commit. If I go to
4:39:13GitHub and I give this a refresh, we can
4:39:15see that we should see right here two
4:39:17commits. This one was add glowing pulse
4:39:20effect to hero join the community
4:39:22button. And then if I go to Verscell and
4:39:25we go to our deployments, we should see
4:39:26that we just got a second one come
4:39:28through as well just now. And now if I
4:39:30go to the site on the web and I refresh,
4:39:33we see the actual glowing join the
4:39:35community button. All right, so those
4:39:36are the five hacks that I wanted to
4:39:37cover today. We have our claw.md file,
4:39:39which as you could tell by this video,
4:39:41yes, it's nice to have something to
4:39:42start, but you are going to continue to
4:39:44iterate upon it throughout your project
4:39:45until you get to a good spot. We've got
4:39:47the front-end design skill, which is
4:39:48just like way too easy to not use. We've
4:39:50got the screenshot loop, which you got
4:39:51to be careful about, but it is very
4:39:53helpful. We've got inspiration websites,
4:39:54and then we have inspiration individual
4:39:57components, and now it's just a matter
4:39:58of making small tweaks and iterating
4:40:01upon your website. Just a reminder, you
4:40:03can grab this claw.md file for free in
4:40:05my free school community. The link for
4:40:06that is down in the description.
4:40:10So, you know how on Apple's site we've
4:40:11got the product, we've got quotes come
4:40:13in, we've got dynamic elements like
4:40:14this. We can see sort of like the
4:40:16deconstruction of the product itself.
4:40:18And all of these features on the website
4:40:19give it a super professional and branded
4:40:21feel. And I'm sure that some of you guys
4:40:22have realized by now, but this website
4:40:23is not Apple. This is one that I built
4:40:25with cloud code using this new skill in
4:40:27like 30 minutes. So, I'm going to show
4:40:29you guys today the exact process to
4:40:31build sites just like this that are
4:40:32super professional, that have these
4:40:34really nice animations going on, and
4:40:36they just feel awesome. I'm also going
4:40:38to be giving you away the skill and
4:40:39everything that you need to do this for
4:40:41completely free so that even if you've
4:40:42never opened up Cloud Code or, you know,
4:40:44built a website before, you will be able
4:40:46to do exactly what you're seeing on
4:40:47screen. So, let's not waste any time and
4:40:49get straight into the video. Okay, so
4:40:50we're using Cloud Code in Visual Studio
4:40:52Code. So, if you don't have that, just
4:40:54go to the browser, type in VS Code, and
4:40:56download this for your operating system.
4:40:57and it's completely free. Once you're in
4:40:58here, you're going to go over to the
4:41:00left, which is the extensions button.
4:41:01You're going to type in Cloud Code and
4:41:03install this one. It will then prompt
4:41:04you to sign in with your Anthropic
4:41:06subscription. And by the way, you do
4:41:07have to be on the paid plan, so either
4:41:09the Pro or the Max in order to use Cloud
4:41:11Code. Now, once you've got that
4:41:12installed, you're going to click on this
4:41:13button up here, which is the Explorer,
4:41:15and you need to open up a folder. So,
4:41:17just go to your desktop, go to your
4:41:18documents, open up a brand new folder,
4:41:19just call it, you know, 3D website
4:41:21testing, and then open it up right here.
4:41:23So, here you can see this is the project
4:41:24that I've been working in. And I've just
4:41:25called it animated websites. And what
4:41:27I'm going to do is give you guys all the
4:41:28skills in here that you need in order to
4:41:30just replicate what I'm doing. And I'm
4:41:32giving you guys all that for free, of
4:41:33course. So that once you have that, all
4:41:35you have to do is give Cloud Code a
4:41:37video. For example, here we have a video
4:41:39of a camera. You can see what happens is
4:41:41it just basically spins around and then
4:41:43it kind of like goes into X-ray mode and
4:41:45we can see inside the camera. And now
4:41:46that Cloud Code has that, you just say,
4:41:48"Build me a website for this video." And
4:41:50what you get is something like this
4:41:52where we have Vans 1. You can see it
4:41:53looks like a product landing page. As we
4:41:55start to scroll down, the camera is
4:41:56revealed. And as we continue to scroll
4:41:58more, we get dynamic text that comes
4:42:00through as well as the actual camera
4:42:02animation starting. And the crazy part
4:42:04is I didn't give it any of the copy. I
4:42:06didn't give it a color scheme. I didn't
4:42:07give it any information about the
4:42:09product. Obviously, you could, but what
4:42:11it will do is it will just basically
4:42:13create this for you and then you can go
4:42:14ahead and make the tweaks that you need.
4:42:16So, the two skills that you're going to
4:42:17need to go grab are the front-end design
4:42:19skill, which is the one that I grabbed
4:42:20from Anthropic. It was an official
4:42:22skill, but I changed it a little bit.
4:42:23So, this is like kind of my version of
4:42:25the front-end design skill for this
4:42:26specific use case. And then you're also
4:42:28going to need to go get the video to
4:42:30website skill. So, both of these skills
4:42:32are literally just markdown files.
4:42:33They're markdown files that I've worked
4:42:35on that tell Claude Code the best
4:42:36practices for creating these animated 3D
4:42:39websites. And you can get these skills
4:42:40for completely free by going to my free
4:42:42school community. The link for that is
4:42:43down in the description. You'll go to
4:42:45the classroom, you'll go to the skills
4:42:46section, and you'll be able to find all
4:42:47of them right in there. And literally
4:42:49all you'd have to do is take them from
4:42:50your downloads folder, drag them over
4:42:52here to the lefth hand side, and then
4:42:53just basically tell Claude, "Hey, set up
4:42:55this folder." So I've got myclaude.
4:42:57Within that claude, I've got a skills
4:42:59folder. And then within that skills
4:43:00folder, here are my two skills I need
4:43:02you to use. And it should look something
4:43:05like this on the lefth hand side. Now,
4:43:07before you get too overwhelmed, if
4:43:08you've never used Cloud Code before,
4:43:09it's super simple. Right here, we have
4:43:11our actual agent that we talk to and
4:43:13have conversations with. And over here,
4:43:15we just have folders. So, I just showed
4:43:16you where the skills live, which is in a
4:43:18docloud folder within a skills folder.
4:43:20But then the other stuff in here is
4:43:22super simple. Here are all the video
4:43:23files that I've used to make the
4:43:25previous websites. And then for each
4:43:26website, it created its own folder. So,
4:43:28for the camera website, it has all this
4:43:30information. For the watch website, it
4:43:32has all this information. And for the
4:43:33Yeti website, it has all that
4:43:34information as well. But anyways, if you
4:43:36just follow the steps that I take in
4:43:38this video, it will all make sense and
4:43:39you will get a nice output. And by the
4:43:41way, if that's a bit confusing, then
4:43:42definitely go check out my skills video,
4:43:44which I will tag right up here. And then
4:43:46once you understand that, hop right back
4:43:47over here. All right, so I'm going to
4:43:48walk you guys through this process step
4:43:50by step so you can see exactly what I
4:43:52did to create these really cool
4:43:53animations. And what you need to
4:43:55understand is that all of these
4:43:56animations that you're seeing on the
4:43:58site and all the other ones that I've
4:44:00shown, they're just videos. And so all I
4:44:02did here was had Nano Banana generate
4:44:03two different images for me and then
4:44:05turn it into a video. So, for example,
4:44:06that first one with the Apple Watch is
4:44:08just the Apple Watch kind of opening up
4:44:10and then exposing all of its layers. And
4:44:12then when I wanted it to close back up,
4:44:14I just prompted it to do the opposite.
4:44:16And now we have the second one where it
4:44:18goes from that end frame and then kind
4:44:19of like folds back in on itself and
4:44:21reveals the Apple Watch. So, now that
4:44:23you understand how simple this really
4:44:24is, let's go over to Nano Banana and
4:44:27start making some images. Okay, so the
4:44:29way that I like to use Nano Banana,
4:44:30whether I'm doing it in a playground
4:44:32like this or over API, is with key.ai AI
4:44:34because I found that it's really fast
4:44:36and it's like really cheap. So I went to
4:44:37Nano Banana 2, the new version of Nano
4:44:40Banana, and I came in here and did a
4:44:4216x9 aspect ratio image, and I said, I
4:44:45need a professional studio grade image
4:44:46of a blender. It should be against a
4:44:48plain all black background with no
4:44:50shadows, no hands, no reflections. And
4:44:53this is essentially going to be our
4:44:54start frame when we make that actual
4:44:56video. So then I saved that output,
4:44:58dropped it in here as an image input and
4:44:59I just said the exact same prompt, but
4:45:01this time it should be filled with fruit
4:45:03and juice. So now we have our start
4:45:04frame over here and our end frame and we
4:45:07just need to animate it so that it
4:45:08actually looks like fruit and juice is
4:45:10you know being dropped and poured in. So
4:45:12then I found the best results using
4:45:14cling 3.0 which once again you can also
4:45:16access in this key.ai. So we give it the
4:45:20start frame, we give it the end frame
4:45:21and here's what I did for my prompt. I
4:45:23actually went over to Claude and I gave
4:45:25it the start and end frame and said,
4:45:26"Help me make an AI video prompt where I
4:45:28want the lid to float off. I want fruit
4:45:30and juice to be dropped in. And then I
4:45:32want the lid to be put back on. No
4:45:34shadows, no hands, no reflections." It
4:45:36spits out this prompt. I put that into
4:45:38Cling. And here's the result. I haven't
4:45:40watched this yet, so hopefully it's
4:45:41good.
4:45:45Okay. Interesting. I mean, obviously, we
4:45:48maybe want to fix that a little bit, but
4:45:49I'm fine with that for the sake of the
4:45:51demo. the fruit and juice kind of just
4:45:52magically appear from nowhere and now
4:45:54they're in the blender. Okay, so I
4:45:56downloaded that video. I'm just going to
4:45:57drag it over here to the lefth hand side
4:45:59and you can see right now that this
4:46:00video that cloud code is looking at is
4:46:02the exact one that we just generated
4:46:03together. So this is what it's going to
4:46:05use. It's basically going to take that
4:46:07video, pull out like 120 frames or
4:46:09however many frames make up that video
4:46:11and then it's going to have all of those
4:46:12saved. So just as an example, if I go to
4:46:14the frames from the watch, you can see
4:46:17it created over a hundred webp pictures
4:46:20from that actual video. And it's kind of
4:46:22like stopotion animation where each one
4:46:24just changes a little bit and as you go
4:46:26through the actual video starts to form.
4:46:28So it basically associates each of these
4:46:30frames with a scroll position. So as you
4:46:32scroll down it kind of like reveals
4:46:34itself or if you scroll backwards it
4:46:36goes the other way. So now I'm going to
4:46:38make sure that my agent is on plan mode
4:46:39which basically means it won't actually
4:46:41do anything. It'll just read things and
4:46:42it will help create a plan which is
4:46:44going to result in much better websites
4:46:45on the first try. All right, so I
4:46:47started off by saying I just dropped in
4:46:48a video called blendercling.mpp4.
4:46:51Help me create a one-page product
4:46:53landing page for this product. It should
4:46:56be modern. It should feel very
4:46:58professional. It should have smooth
4:46:59animations and design throughout. All of
4:47:02the text should be easy to read. And the
4:47:04background of the website should be
4:47:06completely black. It should be a dark
4:47:08mode. and it should blend into the
4:47:10background of the Blender Cling image so
4:47:13that it looks like it was one, you know,
4:47:15fluid web page. So, if you were really
4:47:17doing this for a product or for a
4:47:18business, you'd probably want to prompt
4:47:20it with some more information. But, I'm
4:47:21just going to show you guys for now how
4:47:22this works. I'm going to shoot that off
4:47:24in plan mode, and you will see that it
4:47:25naturally comes back and asks me some
4:47:27questions. Now, if you're wondering how
4:47:28I was able to speak right into my cloud
4:47:30code, then check the description for my
4:47:32tool. Now, one thing you'll notice once
4:47:33you read through the way that this works
4:47:35is that you will have to have FFmpeg,
4:47:37which basically just extracts frames
4:47:39from videos. It's a free tool, and if
4:47:41you don't have it installed, Claude Code
4:47:43will help you install it. It'll just do
4:47:44it for you, so don't worry about that.
4:47:46But now, you can see it's asking us some
4:47:47questions. So, the first one is, what is
4:47:48the product name and brand for this
4:47:50blender? And I'm just going to go ahead
4:47:51and say create fictional branding, just
4:47:53like it did for the Vanta 1, which was
4:47:55the camera website. Then it says, what
4:47:57kind of content sections do you want?
4:47:59I'm just going to go with the full
4:48:00premium, which is what we've been doing
4:48:01for the other example sites. But of
4:48:03course, this is where you could
4:48:04customize it a little bit to fit your
4:48:07needs. And if you've never used Cloud
4:48:08Code before, what's happening here is
4:48:11you are basically able to look at
4:48:12everything that Cloud Code is thinking
4:48:14and doing. You can see a task, you can
4:48:16see a glob, you can see what it's
4:48:17reading, and you'll see later when it
4:48:18actually starts implementing things,
4:48:20it'll create itself a task list or a
4:48:21to-do list, and you can actually watch
4:48:23it fulfill those to-dos. So, it's really
4:48:25cool. Okay, so now we at the point where
4:48:27the plan is done. We've got Obsidian
4:48:29Vortex premium blender landing page.
4:48:31We've got the brand identity here. So,
4:48:33Obsidian, Obsidian Vortex. The tagline
4:48:35is annihilate everything. The accent
4:48:37color is blood red. And we've got fonts.
4:48:40We've got the video details,
4:48:41implementation steps. So, here is the
4:48:43ffmpeg thing I was talking about. That
4:48:45will extract the actual frames. Then it
4:48:48will build the HTML. Then it will create
4:48:50all these different sections. It will
4:48:51build the CSS. It'll build the JS. And
4:48:53then it will test locally. So, before
4:48:55this actually goes anywhere on the web,
4:48:56we will test it right here. And then if
4:48:58we're good with it, we can push it to
4:48:59the web. And I'll show you guys that at
4:49:00the end of the video. So I'm actually
4:49:02just going to make one more suggestion
4:49:03and then we'll let it start building. I
4:49:04want to say this looks pretty much
4:49:06there. There's one thing that I forgot
4:49:07to mention, which is that I want the
4:49:09product video to be kind of right
4:49:11aligned. So I want it to be on the right
4:49:132/3 of the page and all the text can be
4:49:16left aligned. And this is just to show
4:49:18you that you do have control over the
4:49:20way it looks. So I'm going to go ahead
4:49:21and shoot this off. It's going to come
4:49:22back with another plan and I'm going to
4:49:24accept it and then I'll check in with
4:49:25you guys when we have our first
4:49:26iteration of this website done. And by
4:49:29the way, now that I've accepted the
4:49:30plan, I put it in bypass permissions
4:49:32mode so that it can just continue to run
4:49:33without stopping all the time to ask me
4:49:35questions. Here is the to-do list that I
4:49:37talked about. It's going to go through
4:49:38and make sure that all of these are good
4:49:40to go. All right, so in just a few
4:49:41minutes that came back and it says, as
4:49:43you can see, the site is live at
4:49:44localhost. Open it up in your browser
4:49:46and check it out. So let me go ahead and
4:49:48open this up and we will see what we
4:49:49got. So, you just saw the Obsidian
4:49:51loading. We've got Obsidian Vortex. It
4:49:52looks really clean so far. Now, let's
4:49:54see what happens when I start to scroll
4:49:56down. So, we see the blender starts to
4:49:58appear. Nice. And as I continue to
4:50:00scroll, we have we didn't build another
4:50:02blender. We engineered a force of nature
4:50:03that reduces anything to nothing in
4:50:06absolute silence. As I keep scrolling,
4:50:08we can see that the fruit is starting to
4:50:09appear. So, we've got meet the obsidian
4:50:11vortex machine from aerospace grade
4:50:13stainless steel blah blah blah. There's
4:50:15also text in the background as you can
4:50:17see that says I think it says obliterate
4:50:19everything. So that's pretty cool. We've
4:50:21got brushless motor. We've got titanium
4:50:24blade array. So that actually kind of
4:50:25pops in a little bit late as you can
4:50:27see. As I'm scrolling it pops in a bit
4:50:29late. This one comes through. And then
4:50:31we've got these stats that come up.
4:50:33Really nice. And then we have the last
4:50:35blender you'll ever own. And then I
4:50:36should have the CTA right here which is
4:50:39own the force. Pre-order the Obsidian
4:50:40Vortex. So let's just actually think
4:50:42about that. I dropped in a prompt that
4:50:45said, "Here's a video. Create a onepage
4:50:47landing page." That was basically it. It
4:50:49used the skill and it created a plan for
4:50:51us. And then we actually had a website
4:50:54where pretty much everything is perfect.
4:50:55Obviously, there's a few things that we
4:50:56need to iterate upon, but this is super
4:50:59super clean and the animation I think
4:51:01looks really good. But the the biggest
4:51:02problem that I'm noticing is right here
4:51:04when we scroll to number two feature, we
4:51:06really don't get to see it unless we
4:51:08scroll back down. So, what I'm going to
4:51:09do now is I'm going to go ahead and
4:51:10clear this out because we're at 53%
4:51:12context, so I don't want to mess with
4:51:14context rot. And I'm going to once again
4:51:16go back into plan mode. All right. So, I
4:51:18just tested out the Blender website and
4:51:21it looks really good. I do have one
4:51:22piece of feedback though. When we're
4:51:24scrolling down and we see the features,
4:51:26feature number two doesn't actually
4:51:28appear until it's basically off the
4:51:30screen. So, we need that to actually
4:51:32come into view a bit earlier. But
4:51:35besides that, all the other features,
4:51:36features 1, three, and four are working
4:51:38well. So, I'll shoot that off. And while
4:51:39it's coming up with a plan, I just want
4:51:41to make sure that that was accurate.
4:51:42Feature three looks good. And we don't
4:51:44even have feature four. So, hopefully it
4:51:45can understand, read through the code,
4:51:47and see what's going on. So, while this
4:51:49is creating a plan, I thought that I
4:51:50would real quick explain the difference
4:51:52between local host and actually having
4:51:54something on the web. Because if you've
4:51:55never built a website before in cloud
4:51:57code, that part might not click yet. So,
4:51:59what happens is we kind of have two
4:52:01different environments. We've got our
4:52:02local computer and then we've got the
4:52:04actual cloud. So remember when cloud
4:52:05code said, "Hey, cool. Your website's
4:52:07live. You can go ahead and test it on
4:52:08localhost port 51006."
4:52:12If you right now typed this into your
4:52:13browser, you'd probably get nothing. But
4:52:15I can type it in because it lives on my
4:52:17machine locally because what's going on
4:52:19is cloud code is helping us write all of
4:52:22the actual code because that's really
4:52:23what a website is. It's a bunch of code.
4:52:25Whether that's HTML or JS or CSS or
4:52:28whatever it is, it's code. And so what
4:52:30we're doing right now is we're using it
4:52:32to build code and then we're testing it.
4:52:33And then we're going back and forth on
4:52:35our computer until we're good. And then
4:52:37once we actually like the code, we push
4:52:39it to somewhere on the cloud so that it
4:52:40can actually be viewed by other people.
4:52:43So I'll cover this pipeline once we
4:52:44actually get to that stage in the video.
4:52:46But I thought that I would address this
4:52:47kind of like local thing first real
4:52:49quick. All right. You can see it came
4:52:51back with a plan to fix the Blender
4:52:52feature number two late appearance. I'm
4:52:53just going to go ahead and accept this.
4:52:55Okay. So it did that really fast. I'm
4:52:56going to click into this website and
4:52:57we'll see if that fix has actually been
4:52:59changed. We've got feature one. We've
4:53:02got feature two. There we go. And we've
4:53:04got feature three. And everything else
4:53:05still looks intact with the site. So all
4:53:08of the magic that's happening here is
4:53:09once again within the skill that I
4:53:11built. So it is the video to website
4:53:13skill. If I open this up, this is
4:53:14basically just a markdown file and it
4:53:16says turn a video into a premium scroll
4:53:18driven animated website. So this is
4:53:20where all the secret sauce really lives
4:53:22on how it actually does this. I know
4:53:24that this is a pretty beefy skill, but
4:53:26hey, I mean, it seems to be working
4:53:28pretty well. And what you can do now is
4:53:30every time you're using this skill to
4:53:31build websites, you would just say,
4:53:33"Awesome. Here's what I told you to fix.
4:53:35Here's what I like. Here's what I don't
4:53:36like. Make sure that all of this is
4:53:37reflected on the skill.md." So
4:53:40essentially, every single time that you
4:53:41build a website with this skill, the
4:53:42skill gets better and better and better.
4:53:45Now, we have our code that we like, and
4:53:47it's time to push it over here. So, what
4:53:49we need to do is we're going to use a
4:53:50combination of GitHub and Versel. GitHub
4:53:53basically just lets us store code. So
4:53:56basically the same way you would have
4:53:57maybe like a word file locally on your
4:53:58computer. If you wanted other people to
4:54:00be able to use that, you would have to
4:54:01put that on one drive or you know uh
4:54:04Google doc so that other people could
4:54:06look at the different versions and
4:54:07collaborate on it. And then what happens
4:54:09is we sync up Verscell to our GitHub so
4:54:11that we can actually deploy that code
4:54:13onto a real URL so that it's no longer
4:54:15just like a local host URL. First thing
4:54:17you want to do is go over to GitHub and
4:54:19create an account. It's completely free
4:54:20and it has been around for tons and tons
4:54:22of years. It is a industry standard for
4:54:24code. And then the second thing that
4:54:26you're going to do is go over to
4:54:27Verscell and create an account over here
4:54:28as well. So what's cool is we can
4:54:30basically have cloud code do this entire
4:54:32pipeline for us. It's super super
4:54:33simple. So if you haven't yet connected
4:54:35your cloud code to your GitHub, you
4:54:37would just come in here, maybe clear out
4:54:38the conversation and say help me connect
4:54:41to GitHub so that I can push this code
4:54:43base to my GitHub repository. And what
4:54:45it can do is it can help you basically
4:54:47use the CLI to authenticate. Meaning it
4:54:49will basically just have a popup for you
4:54:51and all you have to do is sign in. It's
4:54:52super simple. So here you can see it
4:54:54says you're already authenticated with
4:54:55GitHub as Nate Herkai. Nice. Now the
4:54:57next step is to say awesome. Now let's
4:55:00go ahead and create a new GitHub
4:55:02repository for me. You can just call it
4:55:04blender- website and push the codebase
4:55:06for the Blender website into that
4:55:08repository. Now because we're pushing
4:55:10something to GitHub and because it's
4:55:12going to go on the web, this is where
4:55:13you'd want to be careful if you had like
4:55:15API keys or anything sensitive in there.
4:55:17In this case, we have literally nothing
4:55:19to worry about, but that is something
4:55:20just to keep in mind. All right, nice.
4:55:22So, your Blender website is now live on
4:55:24GitHub. Here's everything that it did.
4:55:25And it also says if you want to connect
4:55:27this to Verscell for auto deploys, you
4:55:28can import the repo from the Versell
4:55:30dashboard. So, first let's just check in
4:55:32on this GitHub repo. We can see that
4:55:35this has been set up. We have one
4:55:36commit, which is the one that we just
4:55:38did. And that's important because every
4:55:39time we make a change, we can see
4:55:41exactly what happened to the code here.
4:55:42Now, what you'll notice is that this is
4:55:44a public repository. You could go into
4:55:45the settings and you could make this
4:55:47private and it would still be able to
4:55:48auto deploy to Verscell, no problem. And
4:55:50now that we're in Versell, all I have to
4:55:52do is come here to projects and click on
4:55:54add new. And then what you'll notice is
4:55:55right here because I've signed into
4:55:57Verscell with my GitHub, it says import
4:55:59GitHub repository. And all I can do is
4:56:01choose right here blender- website.
4:56:03Click import. And this is automatically
4:56:04just going to build up this site for us.
4:56:06So I'll hit deploy. And then we'll let
4:56:08this spin up. And I'll show you guys
4:56:10that this is now accessible on the web.
4:56:12Awesome. So, you just deployed a new
4:56:14project to Nate Her projects. Let's go
4:56:15ahead and continue into the dashboard
4:56:17here. All right. So, something
4:56:18interesting just happened and I'm glad
4:56:19it did so I can show you guys how to fix
4:56:21it. So, our project is now here on
4:56:24Versell and when I click into the
4:56:25domain, everything seems to load up,
4:56:27right? Right. We've got Obsidian Vortex.
4:56:29We see the animations come through, but
4:56:30as I start to scroll, where's our
4:56:32Blender? We still get the text coming
4:56:34through. We still get all these
4:56:36animations that we were looking for, but
4:56:37our Blender is not there. So what
4:56:39happened was when it pushed this GitHub
4:56:41repo, it excluded the frames. So if I
4:56:44come over to the lefth hand side and we
4:56:45go to our Blender project and we open up
4:56:47the frames, you can see that all of
4:56:48these were grayed out, which basically
4:56:50means everything got pushed to GitHub in
4:56:51this folder except for the frames, which
4:56:54means when the site tries to render it,
4:56:55there's no frames to actually render. So
4:56:57what I did is I said, "Okay, that didn't
4:56:58work. You need to have the frames in the
4:57:00codebase so that it can actually use it.
4:57:02Otherwise, the animation just
4:57:03disappears." So I told it to update
4:57:05those changes, make another push. It
4:57:07fixed that and then basically it came
4:57:09back and said okay I did that now the
4:57:11frames are in the GitHub repo. If I go
4:57:13to the GitHub repo you can see that we
4:57:14now have two commits. If I click into
4:57:16the commits you can see nice the second
4:57:18one added animation frames for Verscell
4:57:20deployment. And then in Verscell if I go
4:57:22to deployments you can see that we had
4:57:24two. We had this current one that I
4:57:25rolled back to so I could show you guys
4:57:27and then we have this main production
4:57:28one which I could go to and we could
4:57:31actually roll up to this one. So I could
4:57:33click on these three dots and then click
4:57:34on redeploy. And now that I have
4:57:36redeployed, if I open up this Blender
4:57:38website one more time and we come down
4:57:40and we start to scroll, we can now see
4:57:42that the frames are being rendered and
4:57:43we still have all of the animations that
4:57:45we had built. And now the site is
4:57:47actually ready to go. And just to prove
4:57:49to you guys that this really is on the
4:57:50cloud, you can see on my phone we have
4:57:52the Blender website with the animations.
4:57:54Now, obviously, this hasn't been yet
4:57:56optimized for mobile, which would kind
4:57:57of be the next step, but you can see the
4:57:59animations are still here. And if it
4:58:01hasn't quite stuck yet, here's the
4:58:03advantage of doing it like this. Now
4:58:05that we have our site on the web, which
4:58:07maybe customers would be looking at and
4:58:09interfacing with, what if we wanted to
4:58:10change something like the colors or
4:58:12maybe even change the animation, we
4:58:14don't want to be changing what's
4:58:15actually out there in production in real
4:58:16time. So, we can change the code, we can
4:58:18test on a local host, and then once
4:58:20we're finally good with it, we push it
4:58:22to GitHub and then it automatically
4:58:23syncs to our real domain. So, that way
4:58:25we have basically a testing environment,
4:58:27a staging environment, and then we have
4:58:29our actual production website. And what
4:58:30you guys saw is that this literally took
4:58:32me like 30 minutes. And obviously if I
4:58:34spend another 30 minutes, the site could
4:58:36be like five times better. And this is
4:58:37crazy because there's so many businesses
4:58:38out there, whether they're local in your
4:58:40area or you know, you look online that
4:58:42have horrible websites because they
4:58:44don't want to prioritize it or they
4:58:45don't want to pay some web design agency
4:58:47tens of thousands of dollars for a new
4:58:49website. That could also take
4:58:51potentially months. Whereas what you
4:58:52could do is you could find prospects.
4:58:54You could find some sort of niche. So
4:58:55let's say you want to design websites
4:58:56for blender companies. You could build a
4:58:58demo site similar to this, right? Like
4:59:00you could build something like this in
4:59:02one day if you sat down and you wanted
4:59:03to make it really really good. And then
4:59:05you email them the link or you walk in
4:59:07and you show them and say I can build
4:59:08this for you with your products with
4:59:10your copy and I can get that to you in 2
4:59:12days and I can charge you know $5 to
4:59:14$10,000 which is a lot cheaper than they
4:59:16might get with other vendors. And then
4:59:18of course on top of that you could do
4:59:20monthly hosting, you could have
4:59:21maintenance which is recurring revenue,
4:59:23but it also ensures for them that if
4:59:24they need any different features, you
4:59:26can fix that codebase and you can push
4:59:27that all up there because you understand
4:59:29how this all works now. So maybe you
4:59:31don't like the style of how I'm doing
4:59:33this with the different, you know, text
4:59:34coming through and maybe like this
4:59:36animation overlay thing, then just
4:59:37update the skill so that your cloud code
4:59:40doesn't actually do that. And once you
4:59:41now actually understand how all we're
4:59:43doing is we're turning videos into a
4:59:45bunch of frames and then having it
4:59:46scroll through, you'll be able to do so
4:59:48much more because it doesn't just have
4:59:49to be product spinning or you know X-ray
4:59:52vision. It can be words. It can be like
4:59:54walking. It can be whatever you want.
4:59:55You can also you can go to this website
4:59:57awards with three Ws for some
4:59:58inspiration on animations. It is super
5:00:01super cool what people are doing here.
5:00:05So Cloud Code is insanely smart. We've
5:00:07got Opus that's kind of behind the
5:00:08scenes. It's the chat model that runs
5:00:10all of the, you know, thinking and
5:00:11planning and coding, but it's limited
5:00:15because without having, you know, web
5:00:17search or without having APIs or other
5:00:19elements of bringing in live data or
5:00:21your specific data, it's going to be
5:00:23very general. So, this section I'm just
5:00:25going to show you a few tips and tricks
5:00:26when it comes to using things like APIs
5:00:29and MCP servers. So, let's go ahead and
5:00:31jump straight in. Today, I'm going to
5:00:33show you guys how you can take any
5:00:34website and turn it into LM ready data
5:00:36in seconds. We're going to be able to
5:00:38take any website and fully scrape
5:00:39everything from it. We can get all the
5:00:41screenshots, the branding. We can map
5:00:43out the entire site. We can extract the
5:00:45data in any form we want. We can pretty
5:00:46much do it all. And we're going to be
5:00:47doing all of this through cloud code
5:00:49using a tool called Firecrawl, which I'm
5:00:51sure you guys might have heard of
5:00:52before. So, within Firecrawl, there's a
5:00:54lot of different things that we can do.
5:00:55We can scrape content, so get everything
5:00:57from the page. We can map out a site, so
5:00:59get all the different URLs and
5:01:00understand the architecture of that
5:01:02website. We can crawl it, so then
5:01:04explore all of those pages. we can
5:01:06actually search. So do like a web search
5:01:07first and then scrape the data which
5:01:09means that we can turn pages into
5:01:10structured content for us to use however
5:01:13we want. Now the thing about this is
5:01:14there are a lot of different endpoints
5:01:16to hit if you were doing this through
5:01:17traditional like API calls. So we're
5:01:19going to be using the MCP server for
5:01:21firecrawl. We're going to give that to
5:01:22claude code so it can figure out based
5:01:24on Nate's natural language request which
5:01:26of these tools do I invoke and in which
5:01:28order do I use them to actually get the
5:01:30end result that he's looking for. So,
5:01:32just to start off with a quick example,
5:01:33I'm going to use Firecrawl in the
5:01:35playground, which is just on the web on
5:01:36firecrawl.dev. You guys can get here
5:01:38using the link in the description. You
5:01:40can get on the free plan, which is more
5:01:41than enough to just play around. And
5:01:42then when you're ready to upgrade, use
5:01:44the link in the description, and you can
5:01:45get 10% off a plan. But anyways, I'm
5:01:47going to go to Upai, and I'm going to
5:01:49copy this URL, paste that in here, and
5:01:52I'm going to run a scrape. But first, we
5:01:54have to choose the format. So, by
5:01:55default, it'll pretty much say, okay,
5:01:56we're going to turn this website into
5:01:58markdown for you. But what you could
5:01:59also do is get an AI generated summary.
5:02:02You get all the links. You get HTML. You
5:02:04could get a screenshot of the full page.
5:02:05You could get the branding scraped so
5:02:07you can understand like the logo and the
5:02:08images and stuff like that. So I'm going
5:02:10to go ahead and start the scrape. And
5:02:11what we'll see is it'll pop up down
5:02:12below and we'll get all of this data. So
5:02:14that just finished up. We can see first
5:02:16of all we have markdowns. So we get the
5:02:17actual like hero text. We get all of
5:02:19this stuff up here. We've got the
5:02:21process. We have the testimonials, get
5:02:23in touch, all of this kind of stuff. We
5:02:25have an AI generated summary of what the
5:02:27business aims to do. We've got a
5:02:29screenshot of the entire page. As you
5:02:31can see, we've got branding information.
5:02:33So, the OG image, the favicon, the logo,
5:02:35colors, typography, stuff like that. And
5:02:37then we also have the JSON if for some
5:02:39reason you're crazy enough to want to
5:02:41read this. So, like I said, we're going
5:02:43to get this into cloud code. So, I'm
5:02:45going to click on the docs and on the
5:02:46right hand side or on the lefth hand
5:02:48side, we can see MCP server right here.
5:02:50And then what we want to look for is
5:02:51running this on cloud code as you can
5:02:53see right here. And now this gives us
5:02:55basically just this one line to put into
5:02:57cloud code and it will be able to
5:02:59install this for us. So I'm just going
5:03:01to go ahead and copy this. I'm going to
5:03:03go into VS Code and we're going to start
5:03:04up a new project and get everything
5:03:06initialized. So if you've never actually
5:03:08been in VS Code or worked with cloud
5:03:09code, then I'm going to link a video
5:03:10right up here. It'll get you caught up
5:03:12and then you can come back over here
5:03:13when you are ready. So what I'm going to
5:03:14do is open up a folder. So I'm just
5:03:17going to open up a project called
5:03:18scraper. There's nothing in here. It's a
5:03:19completely new project. So this is what
5:03:21your guys' setup should look like.
5:03:22Lefthand side, nothing. right-and side,
5:03:24open up cloud code. And I'm just going
5:03:26to say, "Hey Claude, I want to connect
5:03:28to Firecrawl's MCP server." And you can
5:03:31do that using this command. But I'm not
5:03:33going to give you my API key. I'm just
5:03:35going to put it in av. So if you could
5:03:38create that file for me, I will put my
5:03:39API key in there. And then you can go
5:03:41ahead and initialize and connect to
5:03:42Firecrawl's MCP server. All right. So
5:03:44that's going to go ahead and get set up
5:03:45for us. You can see we now have av right
5:03:48here, which is a new file. And this is
5:03:51basically the best way for us to
5:03:53securely put in our API keys so that
5:03:55they're not being stored in like the
5:03:56conversation history. So I'm going to
5:03:58delete this. I'm going to go back into
5:03:59firecrawl. I'm going to go to my
5:04:02dashboard. And right here you can see
5:04:03there is an MCP integration, but we
5:04:05wanted to use the cloud code version.
5:04:06And now we can see API key. I'm going to
5:04:08go ahead and copy that. Paste in the API
5:04:10key right there. And then I'm going to
5:04:11do CRL S just to save it. You could also
5:04:13go here and click file and then save
5:04:15right there. And now we can close out of
5:04:17that file. My API key has been added
5:04:19tov. go ahead and set up the firecrawl
5:04:21mcb server. So now everything should
5:04:23have been set up correctly. So I'm going
5:04:25to hit control shiftp and I'm going to
5:04:27say developer reload the window which is
5:04:29just going to actually let cloud code be
5:04:31able to use this now. So just sending
5:04:32off a request to make sure that that
5:04:33actually worked. As you can see it was
5:04:35able to call the tool and it decided to
5:04:37use the scrape endpoint rather than you
5:04:39know like a map or a crawl. So this is
5:04:41the actual like full markdown of the
5:04:43website itself. Cool. So the next thing
5:04:45I want to do is give our project a
5:04:47little bit more context as to what it's
5:04:49actually doing in here. So first of all,
5:04:51what I want to do is create a fire crawl
5:04:53MCP guide so that when we ask it
5:04:55something, it understands what are the
5:04:57tools I have access to and which ones
5:04:59should I use for what scenario. So I
5:05:01said create a firewall cheat sheet as a
5:05:03markdown file in this project that you
5:05:04can look at and basically it should tell
5:05:06you about the different tools and how to
5:05:08use them. So it's going to go ahead and
5:05:09create this markdown file for us. All
5:05:11right, cool. So, it created that cheat
5:05:13sheet as you can see right here. And if
5:05:14I just open up this full screen, you can
5:05:16see that we've got a quick reference
5:05:17guide. We've got the tool overview. And
5:05:19then it goes on to actually break down
5:05:21how you use each tool. So, this should
5:05:23hopefully be good enough for now Claude
5:05:25to look at whenever we want it to do
5:05:27something else. It even gave it a quick
5:05:28little decision guide, which is pretty
5:05:30nice. Now, finally, before we actually
5:05:32start running with this thing is we need
5:05:33a claw.md file, which is basically the
5:05:36system prompts for this project. Hey
5:05:38Claude, I need you to help me set up a
5:05:39claw.md file for this project. I want
5:05:42this to basically explain that this
5:05:44project specifically is for scraping
5:05:46data. Whether that is extracting it,
5:05:48getting screenshots, crawling
5:05:49everything, mapping everything. You have
5:05:52access to the firecrawl mcp server to do
5:05:54everything that you need to do with
5:05:55websites. And you also have the
5:05:57firecrol-cheatsheet.mmd
5:05:59which explains how that MCP server works
5:06:01and when to use each tool. So I just
5:06:02shut that off. Now I just did want to
5:06:04say this is a demo, right? So, I'm doing
5:06:05this all in bypass permissions mode. But
5:06:07in practice, what I would have done is
5:06:09went to plan mode, brainstormed with
5:06:10Claude a little bit to make sure that it
5:06:12agrees with like the way that we're
5:06:14setting up all these files and then we'd
5:06:16go ahead and implement that plan once we
5:06:18are in alignment. But as you can see, we
5:06:20now have our claw.md file and this is
5:06:22basically a scraper project. We have
5:06:24some information about what this project
5:06:26does, how to actually use the tool, what
5:06:28to reference. And this document, as you
5:06:30can see, is a lot more concise than the
5:06:31cheat sheet. And the reason I wanted to
5:06:33separate this is because you don't
5:06:35actually need this entire cheat sheet to
5:06:36be in the claw.mmd file, but now claude
5:06:39knows that it's there in case it ever
5:06:40needs to use it. Okay, so let's think
5:06:42about a cool use case that we might
5:06:43actually want to do with something like
5:06:45firecrawl. So let's say we've got this
5:06:47remote job website and I search for
5:06:49content and there's about 1,700
5:06:51different job opportunities here and
5:06:53there's also I'm assuming not all just
5:06:55on one page. So there's 2, three, four,
5:06:57maybe even up to 60. So, I'm just going
5:06:59to go ahead and copy the URL of this
5:07:01first page right here. We're going to go
5:07:03into Claude and ask it to help us out
5:07:04with this. Hey, Claude. So, I found this
5:07:06website and I've got about 1,700 job
5:07:09opportunities that I want to look at,
5:07:10but I need help using the Firecol MCP
5:07:12server in order to get all of these
5:07:14listed out. I want these as structured
5:07:16data so I could maybe just throw them
5:07:17into a Google sheet. Now, in this case,
5:07:19I am going to go to plan mode because
5:07:20this might take a little bit of thinking
5:07:22as far as understanding the structure of
5:07:23the site and maybe using more than just
5:07:25a scrape. It might have to use a map or
5:07:27a crawl or something else. So, we'll see
5:07:28what it decides to do here. So, hope you
5:07:30guys see now why I did this. It first
5:07:32decided to scrape. It understood the
5:07:33website and then it decided to map. And
5:07:35now it's creating more of a
5:07:36comprehensive plan about what to find.
5:07:39It's also asking me some questions,
5:07:40which is going to make this job work a
5:07:42lot better. So, it asks if we want all
5:07:441782. I'm actually just going to go
5:07:46ahead and say like 200 because I don't
5:07:47want this to take forever. For data
5:07:49fields, I'm going to grab all of them.
5:07:50For description, let's just do a
5:07:52summary. And I'm going to go ahead and
5:07:54submit those answers. and it's going to
5:07:55keep working on the plan. All right, so
5:07:56looks like that is all done. We're going
5:07:57to go ahead and auto accept this plan.
5:08:00So I'll check in with you guys when
5:08:01that's done. Now, right here is the
5:08:02beauty of Agentic Workflows because it
5:08:04tried to, you know, execute the plan,
5:08:06but once it got into it, it realized
5:08:08that something didn't work. So it said
5:08:09that the extract actually returned empty
5:08:11results and the site might require more
5:08:13sophisticated handling. So now it's
5:08:15trying out the firewall agent. So just
5:08:17super cool the way that it's able to,
5:08:18you know, run into an issue and then fix
5:08:20it. Okay, so that just finished up and
5:08:22it was able to get 200 job listings for
5:08:24us. A few things happened in there, but
5:08:26it it was able to just correct itself
5:08:27and change up the plan and we did get
5:08:30our final output. So, let me open up
5:08:31this CSV. You can also see that it
5:08:34dropped it in this project and we could
5:08:35look at it over here, but it's not
5:08:37really very nice to look at. So, here's
5:08:38the actual Excel file. We've got title,
5:08:41company, job type, location, salary,
5:08:43experience, category, posted, how long
5:08:45ago, apply URL, description, and tags.
5:08:48And we do indeed have 200 of these. So
5:08:50now if we wanted to applied all of
5:08:51these, we've got all the URLs and we
5:08:53have all this info that we need in order
5:08:54to go and do that. So think about how
5:08:56long that would have taken you to build
5:08:57an automation in something like Nitn in
5:08:59order to go scrape 200 of these job
5:09:01postings. Or if you were to just do this
5:09:03manually, it would have taken a lot
5:09:05longer. And once again, I didn't have to
5:09:06think about any of the configuration. I
5:09:08just gave Cloud Code the MCP server and
5:09:10let it run. We're going to do two more
5:09:11really quick use cases. The first one
5:09:13I'm going to do is grab Cloudbot or
5:09:15Moltbot and drop it in here and say,
5:09:17"Please grab screenshots of this page
5:09:18and help me understand the branding."
5:09:20And I'm assuming that this is going to
5:09:21use the Fire Crawl MCP server. I hope it
5:09:23does. And then I'm going to grab this
5:09:24website, which is coffee. And I'm going
5:09:27to open up a different agent on the
5:09:28right hand side. And for this one, I'm
5:09:29just going to tell it to map this site.
5:09:31Go ahead and map out this site for me
5:09:33and show me what it looks like. So now
5:09:36we've got two different Cloud Code
5:09:37agents working at the same time. They're
5:09:39both doing different tasks and they're
5:09:40using different fire crawl tools. and
5:09:42then I'll check in with you guys when we
5:09:43get both of these back. All right, so
5:09:44the map is already done. You can see
5:09:46that it comes back and it says, "Okay,
5:09:48so here are all the main pages and it
5:09:49gives me the links. Here are all the
5:09:50different categories. So best sellers,
5:09:52coffee, instant, matcha, all this other
5:09:54stuff. We've got different collections.
5:09:55We've got different locations. We've got
5:09:58tons of different URLs for products,
5:10:00brew guides, all this kind of stuff."
5:10:01And so now that it has that context, I
5:10:03could have it go actually crawl those
5:10:04things if I wanted to or, you know,
5:10:05extract all of that to a database or
5:10:08whatever we want to do with it. And now
5:10:09it looks like this one is finishing up
5:10:11over here with the Moltbot
5:10:13documentation. So the first thing we
5:10:14have is a screenshot. So if I open this
5:10:16up, we can see right here that we do
5:10:18have a screenshot of that whole landing
5:10:19page for Moltbot. And then we also have
5:10:21the branding like the color palette, the
5:10:23typography, spacing, and components.
5:10:26We've got the logo. And you can see that
5:10:28all of this was able to be done with
5:10:30Firecrawl super easily. So I wanted to
5:10:32show you guys all of that stuff that we
5:10:34just did together, what that actually
5:10:35costed me. So, I'm going to refresh my
5:10:37dashboard here when we're looking at the
5:10:38usage. And you can see that that took me
5:10:40about 30 credits out of my 500 that I
5:10:43get for free. So, 6% of my 500 credit
5:10:46limit. Now, that's really the main
5:10:47difference when it comes to pricing.
5:10:48You've got these different plans. You've
5:10:50got a different amount of pages that you
5:10:51can scrape. You get a different amount
5:10:52of credits. But the other big one is the
5:10:55concurrent requests. So, with the free
5:10:56plan, you can only be doing two at a
5:10:58time. With this hobby plan in the
5:11:00middle, you can be doing five. If you
5:11:01scale that up, you can be doing more and
5:11:03more. And it's not really a huge deal
5:11:04because what would happen is cloud code
5:11:06would basically just cue them up and
5:11:07wait and retry. But if you did want to
5:11:10do some big operations in bulk, then it
5:11:12may be nice to have more concurrent
5:11:14requests running. And remember, you can
5:11:15use the link in the description to get
5:11:1710% off your firewall plan.
5:11:22You can see right here, all I said was,
5:11:23"Hey Claude, I want you to take this
5:11:24YouTube video and repurpose it into a
5:11:26LinkedIn X and Instagram post." Then I
5:11:28dropped in the link to the YouTube video
5:11:29and shot it off. Not only did it create
5:11:31all these assets, but it also found bugs
5:11:33in its own code and fixed those. And
5:11:35then we have this folder over here
5:11:36called drafts. And if I open it up, you
5:11:38can see that we have building beautiful
5:11:40websites with cloud code, which is the
5:11:41video I gave it. And then in here, we
5:11:43have Instagram with our actual post text
5:11:45and five visuals. We've got LinkedIn
5:11:47with our post text and a visual. And
5:11:49then same exact thing for X. So that
5:11:51exact workflow right there took me from
5:11:52having one long form YouTube video to
5:11:54having a finished LinkedIn post, a
5:11:56finished Instagram post, and a finished
5:11:59X post. And if I wanted to ask it to
5:12:01generate posts for six other social
5:12:03platforms, it could because it can use
5:12:05all of them and understands how they all
5:12:07work. So today I'm going to be showing
5:12:08you guys how you can basically 9x your
5:12:10content game using a combination of
5:12:12cloud code and potato. So right now when
5:12:14you're creating content, you know, it
5:12:15takes a lot of time and when you put all
5:12:17that time into creating, let's just say
5:12:18a YouTube video, it'd be really nice to
5:12:20be able to repurpose that content into
5:12:23different platforms as well. So what can
5:12:25help you do is it can create the source.
5:12:27So it can look through transcripts,
5:12:28websites, PDFs. It can find inspiration.
5:12:31It can then create visuals for you. So
5:12:32infographics, carousels or videos. And
5:12:35then it can actually go ahead and
5:12:36schedule that stuff. So it can post it
5:12:38to nine platforms. It can create the
5:12:40stories. It can create the, you know,
5:12:41content calendar. And we can do all of
5:12:43that using potato and automate it with
5:12:45cloud code. So you guys saw a demo
5:12:47earlier, but I'm literally going to set
5:12:48up a brand new account today. I'm going
5:12:50to walk you through the exact steps that
5:12:51you need to do. And basically all we
5:12:53have to do is get our API key, add our
5:12:55MCP config for potato and then just
5:12:57connect our accounts and we're already
5:12:59ready to start creating content in less
5:13:00than 5 minutes. All right, so the way
5:13:02that I like to use cloud code is within
5:13:04an IDE called Visual Studio Code. Now
5:13:06you could use this in anti-gravity, you
5:13:08could use it in the terminal, you could
5:13:10use it in cursor, but I like to use
5:13:11Visual Studio Code. So if you don't have
5:13:13this, then just go to your browser, type
5:13:14in Visual Studio Code. You can download
5:13:16this for both Windows or Mac or whatever
5:13:18operating system that you're on. Now
5:13:19once you're in here, this is what it
5:13:21should look like. and I'm going to walk
5:13:22you through everything you need to click
5:13:23on and everywhere you need to type. So,
5:13:25don't get overwhelmed. If you'd rather
5:13:26watch like kind of an intro video and
5:13:28then come back, then I'll tag this one
5:13:29right up here and then hop on back over
5:13:32here. And by the way, if you've been
5:13:33watching my channel for a while, then
5:13:34you've known about Blotato. I showed it
5:13:36in Naden in this video. And also in all
5:13:38of my kind of like faceless shorts
5:13:40videos, we use Blot to do the auto
5:13:42posting and scheduling. But now, I'm
5:13:44just showing you how it's actually a lot
5:13:46easier to use with Cloud Code. So,
5:13:48that's exactly why once we're in here,
5:13:49we're going to go over to this lefth
5:13:50hand side and click on the extensions
5:13:52button. And all you have to do is type
5:13:53in cloud code. It'll be this one right
5:13:54here that's verified from Anthropic. And
5:13:56then you'll just go ahead and install
5:13:57this. When you install it, it will
5:13:59prompt you to sign in with your paid
5:14:02Cloud subscription. Now, this does have
5:14:04to be the Pro or Max plan because if
5:14:06you're on just the free, you don't have
5:14:07access to Claude Code. Now, once you've
5:14:10installed this, what it will do is give
5:14:11you this little orange button in the top
5:14:12right which lets you open up Claude
5:14:14Code. And this is kind of like your
5:14:16typical Claude or Chat GBT interface
5:14:18where you get to talk to an agent right
5:14:20here. And now what we need to do is open
5:14:21up a project or a folder. So I'm going
5:14:24to go over to this top left button that
5:14:26says explorer. And it will say you have
5:14:28not yet opened a folder. Go ahead and
5:14:29open one up. And that's where we'll be
5:14:31working inside for this specific, you
5:14:33know, AI social media poster project. So
5:14:35I just went ahead and I created a brand
5:14:37new one. I just called it potato.
5:14:39There's absolutely nothing in here. And
5:14:40this is what your screen should look
5:14:42like. And now what we want to do is just
5:14:43basically close out of the welcome
5:14:45thing. We can go ahead and double click
5:14:46and then hit the cloud code button. And
5:14:49now we just have our files which will be
5:14:50on the left. We don't have any yet. And
5:14:52then we have our cloud code agent right
5:14:54here that we are going to be able to
5:14:55talk to. So what I'm going to do in the
5:14:56chat is paste in this prompt that says
5:14:58create me a new skill called repurpose
5:15:00YouTube video. It's going to create an
5:15:01AI social media manager that makes
5:15:03social media posts for LinkedIn,
5:15:04Instagram, and X. The user will input a
5:15:06YouTube video URL and wants it I
5:15:09misspelled this here to be turned into a
5:15:11LinkedIn post, Instagram post, expost,
5:15:12and each one should have a visual that's
5:15:14optimized for that platform. So, Volt is
5:15:16basically going to take this video and
5:15:18do everything for us. I end this prompt
5:15:20by saying, "Ask me clarifying questions
5:15:21one at a time until you are 95%
5:15:23confident that you can complete the task
5:15:25successfully." And I kind of use this
5:15:27templated prompt from Sabrina. So, shout
5:15:28out, Sabrina. You guys can all copy and
5:15:30paste this exact prompt for my free
5:15:32school community or you can just copy it
5:15:34by looking at it right here. So now that
5:15:36this is running, it's going to start
5:15:37going through that process. The first
5:15:39thing that it's doing is it's
5:15:40researching about Blotato to figure out
5:15:41what it's actually able to do and it's
5:15:43basically going to help us build out
5:15:44this flow where we drop in a video. Blot
5:15:47extracts the transcript, adapts the
5:15:48content for these different media
5:15:50platforms, and then it creates
5:15:52everything. And then we're able to
5:15:53review it and then just basically
5:15:55approve it manually. Now, before I start
5:15:57going through this flow of answering
5:15:58questions, I wanted to explain what is a
5:16:00skill because you'll notice I asked it
5:16:02to create a new skill. Just think of a
5:16:04skill like a recipe. If you tell your
5:16:06agent to write a LinkedIn post, it would
5:16:08look at the LinkedIn post skill and that
5:16:10would have the name of the dish, the
5:16:12ingredients, the steps, and then the
5:16:14finished output. That way, the agent
5:16:15could read the recipe and make sure that
5:16:17every single time you ask it to make
5:16:18that dish, it comes out perfect. So,
5:16:21because we're turning this process into
5:16:22a skill, every single time we use it,
5:16:24it's only going to get better and more
5:16:26consistent. But anyways, now we're going
5:16:27to come back into Cloud Code and answer
5:16:29the clarifying questions. So, it's
5:16:31asking me what program language to build
5:16:33this in. I don't really know what I want
5:16:34to do here, and you may not either. So,
5:16:36what I'm just going to say is whatever
5:16:38you think is best. It decides to go with
5:16:40Python because it's the cleanest for
5:16:41this kind of APIdriven tool. So, then it
5:16:44asks, should the tool autopublish or do
5:16:47preview and approval? I want you to
5:16:49always make sure I review it before you
5:16:50ever publish anything on my social media
5:16:52accounts. And if you guys are curious
5:16:54about how I'm talking and words are
5:16:55appearing, then check out the
5:16:56description for the tool. Anyways, for
5:16:58the AI generated post copy, do you want
5:17:00to use Claude to write the tailored
5:17:02posts or should the tool just reformat
5:17:04the extracted YouTube content without an
5:17:06LLM? I definitely want to use Claude to
5:17:08rewrite the text content for the
5:17:10different platforms. Cool. So, we'll
5:17:12eventually have to give it an Enthropic
5:17:13API key as well as the Blot API key.
5:17:16Now, it asks about the tone of voice
5:17:18that we want. So, in this case, if you
5:17:20had like a custom GPT or Claude project
5:17:22already that helps you write LinkedIn
5:17:23posts or Instagram posts, you could just
5:17:25go ahead and grab the instructions from
5:17:26that and put that here. But for now, I'm
5:17:28just going to keep this really simple
5:17:29for the sake of the demo and just say
5:17:31platform adapted, professional on
5:17:33LinkedIn. Um, and on X, you can be
5:17:35casual and maybe even a little bit
5:17:37humorous. Now, it's asking for
5:17:38Instagram. What type of post should it
5:17:40create? I would like you to create a
5:17:42educational carousel that looks like I
5:17:46am writing tweets. And now it asks about
5:17:48the visuals for LinkedIn X. For
5:17:50LinkedIn, let's do a key takeaway
5:17:53graphic. It should be clean and it
5:17:55should have some text that explains the
5:17:57value in the video. And for X, yeah,
5:17:59let's just do something eye-catching.
5:18:00Maybe a cool quote. Next question is,
5:18:03how do you want to review the stuff?
5:18:04Let's just go ahead and save everything
5:18:07to the drafts folder so I can take a
5:18:08look at it myself. When publishing,
5:18:10should you be able to selectively
5:18:11approve which platforms to post to?
5:18:13Yeah, I will tell you one by one which
5:18:16ones have been approved. I think I
5:18:17nearly have everything. One last
5:18:19question. Should the tool let you edit
5:18:20the draft text files before publishing
5:18:23and have the publish command pick up
5:18:24your changes? Absolutely.
5:18:28Okay, so it looks like we're done with
5:18:29the question stage now. Now, you might
5:18:31notice that there's still some stuff
5:18:32that we might want to give to this
5:18:34platform, like maybe some more
5:18:35information about our business. It
5:18:36doesn't really know anything about what
5:18:37we do and maybe things like the colors
5:18:39we like to use or our logo or something
5:18:42like that so that it can be on the
5:18:44visuals. So, now what it's going to do
5:18:46is it's going to build up those
5:18:48different tools. As you can see, it
5:18:49makes a to-do list and it's going to go
5:18:50through one by one and finish all that.
5:18:52And also, what you're going to notice is
5:18:54on the lefth hand side, we're going to
5:18:55start to get files and folders in our
5:18:57project set up. And that's really
5:18:59important because if your cloud code
5:19:02workspace isn't organized and it doesn't
5:19:03understand where files are and if you
5:19:05don't understand where files are then
5:19:07it's just going to get messy and it
5:19:08might be hard to manage the context.
5:19:10Okay, so that has finished up. You can
5:19:12see that we have a new project
5:19:13structure. So over here we have a pi
5:19:15cache. We've got claude with our
5:19:17command. So this is essentially the
5:19:18exact same thing as skills. So this is
5:19:20the repurpose YouTube video skill that
5:19:22it created. You can see that it created
5:19:24some actual Python scripts to draft to
5:19:26post to publish. And so maybe we would
5:19:28want to clean this up and put this in a
5:19:30different folder called scripts or
5:19:31something like that. But the action item
5:19:33on us now is to actually set up our API
5:19:36keys. So if I go into thev, you can see
5:19:39that we have a potato API key and an
5:19:41anthropic API key that we need to set
5:19:43up. So the first step would be to use
5:19:45the link in the description and go to
5:19:47and that will help you get 30% off for 6
5:19:49months. Now once you get that set up,
5:19:51all you'll have to do is go over to the
5:19:52bottom left and go to your settings and
5:19:54then click on right here API. And this
5:19:56is where it will ask you to just
5:19:58basically make sure that this is a paid
5:19:59feature. So if you enable it, you will
5:20:01be on a paid plan. And then you'll go
5:20:03ahead and copy this API key right there.
5:20:05And then in thev, you'll paste this in.
5:20:08And then you will save it. And then it's
5:20:10also asking for an anthropic API key.
5:20:12I'm actually going to go ahead and use
5:20:14open router instead because you can
5:20:15access all the models there. So I went
5:20:17into open router. I created a new key
5:20:18and I'm going to copy this and paste it
5:20:20into Visual Studio Code. And I'm just
5:20:22going to tell Cloud Code that I am using
5:20:25Open Router with Cloud models instead of
5:20:27Enthropic, but you can use whatever you
5:20:29want here. So, I just cleared the
5:20:30context and we're about to do a test
5:20:32run. But before that, I just wanted to
5:20:34show you guys something that we can do
5:20:35that's pretty cool real quick. So, I'm
5:20:37going to go over here and I'm going to
5:20:38drop in a new folder. And I'm going to
5:20:40call this brand
5:20:42assets. Now, what I'm going to do is
5:20:44drag in a profile picture of myself in
5:20:46the brand assets. It's this provo
5:20:47picture right here because I wanted to
5:20:49be able to use this in the tweet style
5:20:51infographics or carousels that we told
5:20:53it we want to make. So, what I can do
5:20:55now is go to YouTube. I've got this
5:20:56video I made a few days ago about
5:20:57building websites in Cloud Code. Copy
5:20:59the link. Come back into Visual Studio
5:21:01Code and say, "Hey, Claude, I want you
5:21:03to take this YouTube video and repurpose
5:21:05it into a LinkedIn X and Instagram post.
5:21:08I've given you in the brand assets
5:21:10folder a profile picture of myself to
5:21:12use in these, you know, different visual
5:21:14posts. Let me know when you've got some
5:21:15stuff ready to review and make sure
5:21:17you're updating your skill document with
5:21:19your findings from this first test run.
5:21:21So now it's going to read through the
5:21:23skill. It's going to execute these
5:21:25different Python scripts right here. And
5:21:27if it runs into any issues or anything
5:21:29that we told it, like using our profile
5:21:31picture, it will update that skill
5:21:32document with. And here's an example of
5:21:34it already needing to make an adjustment
5:21:36is because it said YouTube is blocked by
5:21:38the web fetch tool. Let me try
5:21:39alternative approaches. So that just
5:21:41finished up. You can see that it started
5:21:42off by reading the skill. It goes
5:21:44through and it tries different things.
5:21:45It made its to-do list and it was able
5:21:47to create the actual textbased posts,
5:21:49but what happened was it actually failed
5:21:52on the visuals. So, what it did is it
5:21:54added a known issues and finding section
5:21:56to the actual skill itself. But, we're
5:21:58going to go ahead and try it again and
5:22:00we're going to see if it can fix it. So,
5:22:01I just said try to create the visuals
5:22:03again. Make sure they are images, not
5:22:05videos. And we aren't worried about
5:22:07posting yet. We just want you to create
5:22:08the assets. So, it's once again going to
5:22:10dive into everything. It is going to
5:22:12investigate the templates and then it's
5:22:14going to come back hopefully with
5:22:15something that we can review. So visuals
5:22:17have been created this time and
5:22:18apparently they're looking great and you
5:22:20can see once again it's updating the
5:22:21skill document so that that never
5:22:23happens again. Okay, so these have been
5:22:25created successfully. We've got our
5:22:26LinkedIn with a whiteboard infographic.
5:22:28Let me go ahead and open that up real
5:22:29quick. It's putting all of this stuff in
5:22:31the drafts folder and you can see we've
5:22:33got Instagram, LinkedIn and X and this
5:22:35is all for the YouTube video which is
5:22:37called building beautiful websites. So
5:22:38it's keeping our stuff organized. So for
5:22:40LinkedIn, here's our visual. We've got a
5:22:42whiteboard that says building beautiful
5:22:43websites with cloud code, three key
5:22:45steps. We've got cloud.md, front-end
5:22:47design skill, and then adding your brand
5:22:48assets to a folder. And you can see at
5:22:51the bottom, it also says full
5:22:52walkthrough on YouTube at Nate Herk. And
5:22:55then we also have the actual post right
5:22:56here, which is the textbased copy of the
5:22:59LinkedIn post. So let's say we like that
5:23:01one. Let's go ahead and look at
5:23:02Instagram. We've got the same thing.
5:23:03We've got the post right here with the
5:23:05different, you know, slides, and then we
5:23:07have the actual visuals. So, here is
5:23:09number one. Your cloudMD file is
5:23:11everything. It is a system prompt that
5:23:12runs before every session. We've got the
5:23:14next one, which is the front-end design
5:23:16skill. And then pretty much all of these
5:23:17I'm assuming are the same. We've got
5:23:19brand assets. We've got you don't need
5:23:21to be a developer. And then we've got
5:23:22the difference between vibecoded and
5:23:24professional. So, the one thing I will
5:23:26say about these are that I think this
5:23:27would look a lot better if we had our
5:23:29profile picture as well as like a blue
5:23:31check mark verified badge. So, that's
5:23:33something that we'll probably want to
5:23:34change. And then real quick, just to
5:23:35look at the X post, we've got the actual
5:23:37text itself, which is um very casual and
5:23:40it's kind of more like a meme. And then
5:23:42for the visual, we just have a very
5:23:43simple quote. But as you guys know, I
5:23:45want to make those carousels have the
5:23:47profile picture in there. So I just
5:23:49asked Claude Code to put our profile
5:23:51picture in the carousel slides as well
5:23:53as adding a blue check mark. So we'll
5:23:55see if we can get the job done. All
5:23:56right. So look how cool this is. It
5:23:58fixed that workflow. So it now should
5:24:00have new carousel slides for us. But
5:24:02what it did is it actually had to resize
5:24:04our image because it realized that the
5:24:06potato API wouldn't take it if it was
5:24:08too big. So this was the original and
5:24:10then it resized it to make it smaller,
5:24:12but it still obviously looks the same.
5:24:14And now every single time that we run
5:24:15this for Instagram carousels, it should
5:24:17be able to make it the way we want it.
5:24:19So let's take a look. All right, here is
5:24:21the new Instagram carousel. We've got
5:24:22our name, we've got the verified badge,
5:24:24as well as our profile picture. And so
5:24:26now it would just be a matter of
5:24:27optimizing the actual content that's put
5:24:29in here if we don't think that this was
5:24:31prompted well enough and maybe adding
5:24:32one more at the end which would be a CTA
5:24:34that says like follow for more or
5:24:36something like that. But keep in mind
5:24:38all I said was take this YouTube video
5:24:39and repurpose it. We didn't give it any
5:24:41context about our business about
5:24:43previous Instagram or LinkedIn posts. We
5:24:45didn't give it anything but literally
5:24:46just said make us content. And the only
5:24:48reason I'm telling you this is because
5:24:50think about how much better this will be
5:24:52as you start to add more business
5:24:53context, add more brand guidelines, and
5:24:56then iterate and refine. We've ran this
5:24:58workflow twice, I think, at this point,
5:25:00and it's gotten better each time. What
5:25:01would happen if we ran this 10 times,
5:25:03and every time we gave it more feedback
5:25:04and more feedback so that by the time
5:25:06we're ready to host it, so that if we
5:25:08want it to run automatically, every time
5:25:09I post a new YouTube video, it
5:25:11automatically gives us this stuff. By
5:25:12the time we do that automatically, it's
5:25:14already like a really rock solid or
5:25:16battle tested skill. And by the way, in
5:25:19Blotato, if you go to my videos, you can
5:25:20see all of the ones that you've
5:25:21generated. And you can also go to the
5:25:22API dashboard to see all of the requests
5:25:25that you had been making to Blot. But at
5:25:27this point, the only thing left to do is
5:25:28schedule these out or just instantly
5:25:30post them now that we've reviewed them.
5:25:32So, what you do is you go to your
5:25:33settings and you have to log in with
5:25:34your different accounts. And it's
5:25:36literally like, let's say we wanted to
5:25:37log in with Instagram. We'd click on
5:25:39this and it would just bring us to a
5:25:40sign-in page in Instagram and it would
5:25:42connect everything very easy for us. And
5:25:44then after we've done that, you can see
5:25:45here I've only connected to my X
5:25:48account. It lets you copy your account
5:25:50ID. So basically it associates an
5:25:52account ID in here with potato to
5:25:53actually post on your behalf. But Claude
5:25:55code using the right API endpoints
5:25:57should be able to grab all those IDs for
5:25:58us. So there's really nothing manual
5:26:00here. So just as an example, let's make
5:26:02sure that it can actually post to X for
5:26:04us. Cool. That output looks great. Can
5:26:06you go ahead and post our content on X
5:26:08for us? All right, so that post is live.
5:26:10If I open this up, we should see on X
5:26:13that I did just make this tweet, which
5:26:15I'm going to delete right now. But just
5:26:17wanted to prove to you guys that that
5:26:18endpoint does in fact work. So at this
5:26:21point, now that we know this works, we
5:26:23could just build different skills within
5:26:25this kind of potato environment. So we
5:26:27could build one for getting inspiration.
5:26:28We could build one for creating, you
5:26:30know, Tik Tok videos, whatever we want
5:26:31to do. But before you start scaling this
5:26:33up, it's really important to have some
5:26:34structure to this project because we've
5:26:36got, you know, our claude with our
5:26:38skills, we've got our brand assets,
5:26:39we've got our drafts, but we also have
5:26:41some scripts right here that are just
5:26:42kind of in the middle of nowhere. And we
5:26:44also don't have a claw.md file yet. So,
5:26:47I'm just going to go ahead and do
5:26:48slashinit, which basically just reads
5:26:50through the current project structure
5:26:52and creates a claw.md file around what
5:26:54we have right here. And I assume at this
5:26:56point everyone's aware of what the
5:26:57cloudmd file is, but if you're not, it's
5:27:00basically the overall system prompt for
5:27:01this specific project. Meaning every
5:27:03time before you shoot off a message to
5:27:05cloud code or before cloud code reads
5:27:07it, it's going to read the claw.md file
5:27:09first to understand the direction, what
5:27:12tools it has at its disposal, what rules
5:27:14it needs to follow, things like that.
5:27:16Which means you don't want to keep your
5:27:17claw.md file very long. I think best
5:27:20practice is to keep it under 150 lines.
5:27:22Otherwise, you're just going to fill up
5:27:23your context much quicker. So now you
5:27:25can see that we have a claw.mmd file
5:27:27that goes over the overview, the
5:27:28commands, environment variables,
5:27:30architecture, patterns, things like
5:27:32that, which now, as you can see, gives
5:27:34our project a little bit more structure
5:27:35right here. But I'm still not satisfied.
5:27:37What I want to say now is we have four
5:27:40Python scripts that don't have a home.
5:27:41Could you throw those into a folder,
5:27:43maybe call it scripts or something like
5:27:44that, and make sure that our other
5:27:47skills and cloudmd files are aware of
5:27:50this and can reference it in the future?
5:27:52And that's just the way that I decided
5:27:53to set it up. But you could also say,
5:27:54"Hey, you know, we've got a ton of files
5:27:56here. Can you help me figure out a
5:27:57strategy to clean this up so we can
5:27:59continue to scale this project?" So, you
5:28:01can see that it made a new folder. It
5:28:02threw all of the Python scripts in
5:28:03there, and now it's updating other files
5:28:05in here to make sure that the whole
5:28:07project understands where everything is.
5:28:09So, that is going to do it for today. I
5:28:10think that you guys should be in a
5:28:11really good spot now set up with Cloud
5:28:13Code, set up with Plot to really improve
5:28:16your content game.
5:28:19Google just dropped what some are
5:28:21already calling the most powerful
5:28:22workspace CLI on the internet. So if
5:28:24you've got a ton of stuff that lives in
5:28:26the Google environment just like I do,
5:28:28then you're going to love this because
5:28:29now any of my cloud code projects can
5:28:31access everything. And all I had to do
5:28:33was install one simple thing. So here
5:28:36you can see I said what can you do with
5:28:37GWS which is Google Workspace CLI. So it
5:28:41can search, list, upload, download,
5:28:43move, copy, share anything in my Google
5:28:45Drive. It can do anything in my Gmail.
5:28:48It can do anything in my calendar. It
5:28:49can do anything with Google Docs. Same
5:28:51thing with Sheets. Same thing with
5:28:53Slides. And it also has multi-step
5:28:55workflow recipes. So these are basically
5:28:57like skills. These are chain command
5:28:59patterns for common tasks like creating
5:29:01docs from templates, reading sheet data,
5:29:03and creating a report doc, finding free
5:29:05time and scheduling a meeting. And there
5:29:06are over a hundred of these that it
5:29:08actually has. So out of the box, when
5:29:09you give Claude Code the GWS CLI, you
5:29:12can do anything across any of the tools.
5:29:13And you also have access to over a
5:29:15hundred skills. So I don't know how many
5:29:17times you guys have tried to use
5:29:18something like Claude or Naden to build
5:29:20you a Google doc. And you do this over
5:29:22API. And it ends up just looking like
5:29:24something like this. It literally just
5:29:25looks like raw markdown. And it's
5:29:26obviously horrible. And sometimes to go
5:29:28along with a YouTube video, I make
5:29:30resource guides that look like this. But
5:29:31obviously they have to be formatted.
5:29:33I've got like a header up here and I've
5:29:34got links and different things in this
5:29:37format. But now I can literally just
5:29:38take the link to a YouTube video. I can
5:29:40drop that into cloud code and say create
5:29:42me a YouTube resource guide. It's going
5:29:44to go ahead and download that transcript
5:29:46from the video. And now what it's doing
5:29:47is it's creating the Google doc not via
5:29:49API call, not via MCP, but via a bash
5:29:52command. Meaning it's literally running
5:29:54a terminal command in order to talk to
5:29:56Google and make this. So it just
5:29:58actually created the doc. Here's the ID.
5:30:00And now it's going to populate it with
5:30:01what I need. And now it finished this
5:30:02up. It gave me the link. I'll click into
5:30:04this. And we can see boom, we have an
5:30:06actual resource guide. It's got the
5:30:08image inserted up here as a header. It's
5:30:09got a link that goes right back to my
5:30:11YouTube channel. It breaks down the
5:30:12market traditional automation. It goes
5:30:14through all this stuff and then even has
5:30:15my CTA at the bottom as you can see
5:30:17after all these horizontal lines to join
5:30:19the plus group. So that was obviously
5:30:21just one quick example, but there's so
5:30:23many different benefits here using this
5:30:24Workspace CLI. The first one is that you
5:30:26have one interface. So basically, like I
5:30:28said, it was one GWS CLI that Cloud Code
5:30:30now has access to and it can access my
5:30:32Gmail, my drive, Docs, Sheets, Calendar,
5:30:34Admin, and more. It's also JSON first
5:30:36with structured responses. So, our AI
5:30:38agent is really, really good at working
5:30:39with it. We have auto discovery, meaning
5:30:41the CLI is pretty much always going to
5:30:42stay up to date automatically. Pretty
5:30:44much zero maintenance because we
5:30:45authenticate and then we're going to be
5:30:46good to go. It has built-in skills for
5:30:48triage, for prep, for generations. Like
5:30:50I said, there's 100 others. And it's not
5:30:53much overhead because it's basically
5:30:54just one tool. It's not the same as like
5:30:56having all these different API endpoints
5:30:58or all of these different MCP configs
5:31:00and tools that would take up more
5:31:01context. But I know you're probably
5:31:03wondering, what is a CLI? It stands for
5:31:04command line interface. And what we're
5:31:06typically used to is a GUI or a
5:31:08graphical user interface where we can
5:31:10see buttons, we can see form fields, and
5:31:12we can click on things and that's how we
5:31:14navigate, but computers are more
5:31:16navigating by text and by commands and
5:31:18by typing. So that's really all that a
5:31:20CLI is. And this is an open- source
5:31:22Google Workspace product, and obviously
5:31:24it's completely free. So I'll leave a
5:31:26link to this GitHub repository down in
5:31:27the description if you want to check it
5:31:28out, read more about it. But I'm also
5:31:29going to walk through some of the key
5:31:30details right here. The first thing that
5:31:32I wanted to show you is if you go down
5:31:33here to the skills, this is where we can
5:31:35actually see all of the different kind
5:31:37of recipes they call them for pre-made
5:31:39multi-step workflows that it has. As you
5:31:41can see, creating events from sheets,
5:31:43creating presentations, creating meat
5:31:45space, label and archiving emails.
5:31:46There's so many different patterns that
5:31:48you might use from this pre-built
5:31:50library. Now, if we keep scrolling down,
5:31:52what you'll also notice is that right
5:31:53here it says this is not an officially
5:31:55supported Google product. Now, that
5:31:57doesn't mean that it's unsafe. This is
5:31:58an actual Google product, but the reason
5:32:00why it's not officially supported is
5:32:02because right now it's more of like an
5:32:03open- source beta. It's kind of a
5:32:05developer playground rather than like an
5:32:07enterprisebacked software. And you can
5:32:09see right here that it also says this
5:32:11project is under active development.
5:32:12Expect breaking changes as we march
5:32:14towards v 1.0. So this thing's already
5:32:16really good out of the box and it's only
5:32:17going to get better. And you can see,
5:32:18like I said, when Google Workspace adds
5:32:20an API endpoint or method, GWS picks it
5:32:23up automatically. So you might as well
5:32:24chuck it into cloud code right now and
5:32:26start getting used to it. Okay, so I
5:32:27just uninstalled this so I can walk you
5:32:29guys through step by step how this
5:32:30actually works. It's super easy. What I
5:32:32do is I basically copy the link to this
5:32:34GitHub repository as you can see. And
5:32:36I'm going to basically just give it to
5:32:38Cloud Code and say, "Hey, I want to
5:32:40install this GWS CLI, read through the
5:32:44documentation, and help me install
5:32:45everything that I need to install, and
5:32:47then we're going to get set up." So,
5:32:48this is basically going to do all the
5:32:49research for me, and then all I have to
5:32:51do is follow its instructions. So, it
5:32:53read the docs. It's looking at what we
5:32:55already have installed. It basically saw
5:32:56that I already had some of the
5:32:57prerequisites. So if you don't have
5:32:59those, you'll have to install those. And
5:33:01then it told me that we needed to
5:33:02install the CLI. So it did that. And now
5:33:04we have two options. So the first one is
5:33:06to install G-Cloud CLI so that we have
5:33:07automatic setup and off. Or we could do
5:33:09it manually by creating our own project
5:33:12and whatnot. So let's just go ahead and
5:33:13try option A. Okay. I thought this was
5:33:15going to be just like a simple command
5:33:17that it ran and then we were good. But
5:33:18it's actually like some other thing to
5:33:20install. So let's actually go back and
5:33:21try manual and I'll just show you guys I
5:33:23guess the harder way. Okay. So I'm going
5:33:24to go to this link. go to our Google
5:33:26Cloud Console and make sure you're
5:33:27signed in with the right account up in
5:33:29the top right. And I'm just going to go
5:33:30ahead and create a new project just to
5:33:32show you guys what this would look like.
5:33:33So, new project. I'm going to call this
5:33:35one Claude Code GWS. And we're just
5:33:40going to go ahead and create this
5:33:41project. So, this is spinning up right
5:33:43now as you can see. And now that it has
5:33:45been created, I'm going to select it so
5:33:46we're inside of it. And then I'm going
5:33:47to go up here and type in APIs and
5:33:50services. Click on that. And we have to
5:33:51set up our OOTH consent screen. So, I'll
5:33:53click on this. and it's going to say get
5:33:55started. Click on that. We have to give
5:33:58our app a name. And then we have to
5:34:00choose an audience. So I'm just going to
5:34:01do internal because I only need this
5:34:03right now for my own organization. If
5:34:05you want to do external, it'll basically
5:34:07have you do testing or published. And if
5:34:09you do testing, just make sure that you
5:34:11add your email as a test user. And then
5:34:14all you have to do after you put in your
5:34:15contact information is hit I agree. And
5:34:18then you go ahead and create that. Now
5:34:19once that has been done, you're going to
5:34:21go to create a client ID. So, I'm going
5:34:23to go back into APIs and services. I'm
5:34:25going to go to credentials and then I'm
5:34:27going to go ahead and do a create
5:34:28credential oath client ID. Now, in here,
5:34:30we're going to choose a desktop app. I'm
5:34:32going to just call this GWS and go ahead
5:34:35and hit create. And now, we have our
5:34:37client ID and our client secret. And so,
5:34:39what you're going to do is download this
5:34:40as a JSON file. Now, you can see here
5:34:42that it says to download that file and
5:34:44save it to your global.config/GWS.
5:34:48So, basically, if you can't find this,
5:34:50just say, "Hey, can you give that to me
5:34:51in a full path?" And then you can paste
5:34:52that into your finder or your file
5:34:55explorer and it will take you there. It
5:34:56will probably look something like this.
5:34:58And then you just drag in that
5:34:59credential thing. I called mine client
5:35:01secret. And cloud code will be able to
5:35:03look at this globally now. And so what
5:35:05you'll notice is that we didn't in this
5:35:06project yet enable these APIs. So let me
5:35:09just show you what happens without that.
5:35:10So it says the last step is to run GWS
5:35:13off login. So I just said, hey, I
5:35:15finished option B. The credentials are
5:35:16called client secret. And then I told it
5:35:19to run the O login. So that should
5:35:21basically open up a tab for you, but if
5:35:23it doesn't, then you can ask for it to
5:35:25give you that URL so that you can
5:35:26actually authenticate in. So you would
5:35:28basically choose your account that you
5:35:29want to use. And then you just have to
5:35:31basically confirm that it can access all
5:35:32of these different things as you can
5:35:33see. And then when you hit allow, you
5:35:36should be properly authenticated. After
5:35:37that, it's going to come back and say,
5:35:38"Okay, cool. Let me see if everything
5:35:40works." Now, this hasn't been perfect on
5:35:42the first try every time, but if you
5:35:44just go back and forth a little bit,
5:35:45say, "Hey, that didn't work. Hey, this
5:35:46is what I'm seeing." It will be able to
5:35:48get you there. It's going to be your
5:35:49best friend for something like this
5:35:51because remember it can read all of the
5:35:52actual documentation. And now it says
5:35:54that the author is working, but we have
5:35:55to enable these APIs in our Google Cloud
5:35:58projects. So basically just clicking
5:35:59open these one at a time and all you
5:36:01have to do is hit enable. So it's super
5:36:03simple. You just have to do this like I
5:36:05said for all of these different services
5:36:06that you actually want to be able to
5:36:08use. So that's why I did this on a new
5:36:09project cuz I just wanted you guys to
5:36:10see that. But if you already have one
5:36:12that has all these enabled, then you can
5:36:13just use that project and generate that
5:36:15OOTH client ID. So there you go. You can
5:36:17see that this works. I said, "Can you
5:36:19find my Google doc that I made in April
5:36:20of 2025?" And I went ahead and pulled
5:36:22links to all five of these because
5:36:24obviously that was a very vague request.
5:36:26And now we could take action pretty much
5:36:27anywhere in Google Workspace super
5:36:29simply with this CLI. But like I said, I
5:36:32just got this set up today and I've been
5:36:33playing around with it a ton in my
5:36:34executive assistant project and it's
5:36:36been awesome. It can literally do
5:36:37anything. So here I'm asking it to grab
5:36:39my unread emails from today and based on
5:36:41what it knows about my business and my
5:36:42priorities, give them a score. And if
5:36:44the priority score is below five, just
5:36:46mark it as unread automatically. All
5:36:47right. All right. So, here you can see
5:36:48it said, "Got 30 unread emails. Here's
5:36:50my priority score based on your business
5:36:52context." And as I scroll down, you can
5:36:54see that it's getting different ratings.
5:36:55And based on what I'm seeing right now,
5:36:57this actually looks pretty good. So then
5:36:59I started playing around with Google
5:37:00Slides because I use Gamma right now,
5:37:02but at some point I could imagine that
5:37:03if this gets good enough, then I
5:37:05wouldn't need Gamma anymore. And this is
5:37:06a free option compared to Gamma
5:37:08subscription. So I had it create me a
5:37:09slide deck and it was okay. I threw in
5:37:12my brand guidelines. I threw in my logo
5:37:13and I said, "Hey, can you see this? you
5:37:15created this using the Google slides and
5:37:17it's okay but there's some weird things
5:37:19that I need you to fix. So then it came
5:37:20back and said I cannot see the slides. I
5:37:22just know how to build them
5:37:23programmatically. So that's why there
5:37:24may be some errors with spacing and
5:37:26stuff. So then I basically just gave it
5:37:27access to ChromeDev tools so that it
5:37:29could open the page, screenshot it, look
5:37:31at it and then we built a plan to add
5:37:33visual validation to this Google Slide
5:37:35Creator skill. So now you can see as
5:37:37it's going through it actually takes
5:37:38screenshots and then it can make fixes
5:37:40based on that. So then after it fixes
5:37:42everything, it says, "Okay, cool.
5:37:43Updated the skill. Take a look at it
5:37:45now." So I'll open up this link. Brings
5:37:46me to Google Slides where I have this
5:37:49slide deck. It has kind of my brand
5:37:51colors. It's got the logo up top right.
5:37:53And then as we go through, we can also
5:37:54see that the spacing is a little bit
5:37:55better. It's still not perfect
5:37:56obviously, but we have custom images
5:37:58here that were generated with Nano
5:38:00Banana 2. And even the images are kind
5:38:02of on brand with the sort of orange and
5:38:04blue color scheme. As you can see, we've
5:38:06got this one with the WAT framework.
5:38:08We've got this slide. And it even ends
5:38:10with a CTA for the free school
5:38:12community. So, just to see what else
5:38:13happens, I'm going to say, "Take a look
5:38:14at the slide deck and do another audit.
5:38:16How could you improve the skill in the
5:38:17future?" So, it's going to go ahead open
5:38:20up a tab as you guys just saw. It's
5:38:22going to take images. It's going to
5:38:23flick through the different slides and
5:38:24capture them. And as you can see over
5:38:26here, it now says take screenshot. And
5:38:29now, it's reading that screenshot right
5:38:31there.
5:38:32Now, it just moved on to the next slide.
5:38:34And it's going to go through and look at
5:38:35every single slide. And then, it's going
5:38:36to come back with a plan. And we could
5:38:38probably do a similar visual and
5:38:39validate flow with creating Google Docs
5:38:42as well. So now you can see it's almost
5:38:44on to that last slide. And I hope it
5:38:45fixes this last slide because what you
5:38:47can see here is that the spacing is
5:38:49really off down here. So you can see it
5:38:51came back with an audit. It came back
5:38:52with some future improvements. And one
5:38:54thing that I did notice is that because
5:38:56I made the window smaller, its
5:38:57screenshots were probably worse quality.
5:38:59So it said presentation mode screenshots
5:39:01would probably be better. But anyways, I
5:39:03just wanted to give you guys a little
5:39:04taste of how you can use the GWS CLI.
5:39:07but also use it with other tools to make
5:39:09the functionality even more powerful.
5:39:11So, just remember that this is very new.
5:39:12There's a lot of people out there on
5:39:14Twitter right now saying that this is
5:39:15insanely overpowered. There's also a lot
5:39:16of people that are saying that it just
5:39:18feels kind of finicky. So far for me,
5:39:19it's been pretty great. Everything that
5:39:20I've asked it to do or find or schedule,
5:39:23whatever it is, it's been doing that
5:39:24pretty much perfectly. But there are
5:39:26some people saying that it's asking them
5:39:27to reauthenticate multiple times. So, if
5:39:29that's a little frustrating, I guess
5:39:31just keep in mind that it will only get
5:39:32better and we're not even to version one
5:39:34yet. So I definitely recommend that you
5:39:35come to this GitHub, read about it, but
5:39:37more importantly get this thing
5:39:38installed in your cloud code setup and
5:39:40just start using it, using it, using it.
5:39:45All right, are you ready for a really
5:39:46fun project? We are going to build our
5:39:49own executive assistant. So think about
5:39:51everything that you've done so far in
5:39:52this course. We built some workflows.
5:39:54We've talked about the folder
5:39:56architecture. We've talked about claw
5:39:58MD. We've done a lot of different
5:39:59things, right? Let's turn this into an
5:40:01actual system that can know everything
5:40:04about our business and that we can use
5:40:05all the time. Essentially, a second
5:40:06brain. Now, before you guys hop into
5:40:08this next one, I wanted to preface
5:40:09something just to make sure that there's
5:40:11no confusion. So, what you're going to
5:40:12notice is that I've got some skills
5:40:13running. And if you just think back to
5:40:15earlier when we were building a
5:40:17workflows using the WAT framework,
5:40:18workflows agent tools. Skills are
5:40:21basically workflows. They're basically
5:40:23the exact same thing. We build out, you
5:40:25know, markdown SOPs, natural language
5:40:27instructions that cloud code can read
5:40:29and execute on natural language. So, as
5:40:31I start to talk about skills and as we
5:40:33start to learn more about that, just
5:40:34think of them as the workflows from
5:40:35earlier in this course because workflows
5:40:37have tools that it can call on. And
5:40:38guess what? Skills have Python scripts
5:40:41that they can call on. So, they're
5:40:42literally the exact same thing. I just
5:40:44like to start off calling them workflows
5:40:45and tools because I think that it's just
5:40:47a little bit more intuitive. So,
5:40:48hopefully you guys get that, but it'll
5:40:50click as soon as you get through these
5:40:51next couple videos. So, let's not waste
5:40:53any time. I hope you guys are excited.
5:40:54Let's get into it.
5:40:58This right here is my Claude Code
5:41:00executive assistant. Let me show you a
5:41:01few things that it can do. Okay, so I'm
5:41:03going to start off by saying pretend
5:41:04that it's morning because it's not right
5:41:05now and use the morning coffee skill to
5:41:07help me plan my day. So, I'm going to
5:41:08shoot that off. Now, while that's going,
5:41:10I'm going to open up a new tab and I'm
5:41:12going to shoot off this message that
5:41:13says, "Spin up a sub agent to do
5:41:15research on the new Cloud Code voice
5:41:16feature and then create both a LinkedIn
5:41:18post and a Twitter style carousel for
5:41:20me." It's created a to-do list and it's
5:41:22getting going on that. I'm opening up
5:41:24another window and I'm saying do a pulse
5:41:25check on the team to see if we're on
5:41:27track for the week and for the quarter.
5:41:28And finally, one more just for fun.
5:41:30Create me a visualization for a YouTube
5:41:32video where I want to explain why having
5:41:34Claude as an executive assistant is
5:41:36awesome. And now it's using a
5:41:37visualization skill. So what you're
5:41:39watching right now are four different
5:41:40agents. 1 2 3 4. All doing things for me
5:41:43in parallel. And not just for a cool
5:41:45demo, but these are actually things that
5:41:46I do every single day or every single
5:41:48other day. So, the first agent's done,
5:41:50which is our morning coffee. And we can
5:41:52see this is what is on the calendar for
5:41:53today. That's crazy that it said March
5:41:555th. It's actually the 4th, but still
5:41:57this is all correct. So, it looked
5:41:58through my calendar. It looked through
5:42:00my project management and our quarterly
5:42:01goals. And it pulled in urgent action
5:42:03items. It also pulled in my video
5:42:05pipeline. So, it sees what I'm scripting
5:42:06and what's in the backlog. And it uses
5:42:08all of this context as well as
5:42:09everything on this lefth hand side which
5:42:11are, you know, current priorities, me,
5:42:13OTAAS, my team, work, we've got
5:42:15projects, we've got decisions, we've got
5:42:17all this stuff in here that it's able to
5:42:18look through. So, we've basically given
5:42:20it access to everything going on in my
5:42:21business and now it's able to just plan
5:42:23out my entire day and I can go ahead and
5:42:24say yes and it will just block off my
5:42:26whole calendar. And I really love doing
5:42:27this because if I don't have to have
5:42:29that decision fatigue of what should I
5:42:30do with my next 15 minutes, my next hour
5:42:32and it's just done for me, I'm way more
5:42:34productive. The second one is done. So,
5:42:36we had a sub agent doing research. We've
5:42:37got a LinkedIn post and we've got our
5:42:38carousel. So, right here is the LinkedIn
5:42:40post which is in my tone of voice as you
5:42:42can see. And then if I go over here to
5:42:43our projects and I go to carousels and I
5:42:46go to March 4th, we can see that we have
5:42:48seven slides. We've got slide one which
5:42:50is you can now talk to your code editor.
5:42:52Slide two cloud code voice mode is live.
5:42:54Slide three. Slide four. I think you
5:42:56guys get the point. And then here's
5:42:57slide seven which is the CTA. Now we've
5:42:59got the pulse check which is an even
5:43:01deeper dive than that morning coffee
5:43:02skill. It's looking at all of the
5:43:04initiatives that we currently have in
5:43:05progress and I can see each task and the
5:43:07current status and based on all that
5:43:08information and based on our goals, it
5:43:10gives me follow-ups. So, I'm really
5:43:11easily able to check in on the team and
5:43:13make sure that everything is actually
5:43:14progressing. And the last one is our
5:43:16actual visualization for an executive
5:43:18assistant. So, I would just need to go
5:43:19to projects. I would go to
5:43:20visualizations and we should see March
5:43:224th, we have Claude executive assistant
5:43:24PNG. So, here's what we got. We've got
5:43:26you on the lefth hand side where you're
5:43:27buried in work and on the right hand
5:43:29side we've got with Claude executive
5:43:31assistant. And I'm actually going to use
5:43:32this in the video. So, that was a quick
5:43:34demo. You guys got a little bit of a
5:43:36taste of what this executive assistant
5:43:37can do for me. Now, don't be overwhelmed
5:43:39by all these files over here. I've built
5:43:40this up by using it every day, but I'm
5:43:42going to show you guys exactly how you
5:43:43can follow this framework to have
5:43:44basically your own Herk 2, which is what
5:43:46I have right here, but it would be for
5:43:48you. And so, all four of these things
5:43:49that just happened probably took me a
5:43:51minute or two. And if I was to do each
5:43:52of those manually, it would have taken
5:43:54me at least 25 minutes. So, if you want
5:43:56to know how you can build this for
5:43:57yourself, let's get into it. So today
5:43:59you are going to be building your own
5:44:00executive assistant with claude code.
5:44:01But I wanted to start off real quick by
5:44:03just talking about what most people do
5:44:04and why that's not really the same
5:44:06thing. So with something like claude or
5:44:08chatbt, we've been way more productive
5:44:10because we've been able to save
5:44:11memories. We've been able to save, you
5:44:13know, maybe custom skills or custom
5:44:14prompts. There's still so many times
5:44:16where you've probably thought to
5:44:16yourself, man, I wish this thing just
5:44:18knew everything about what's going on.
5:44:20So in that old way, you kind of feel
5:44:21like you're repeating yourself a lot or
5:44:22you're giving it extra context or it's
5:44:24just helping you get maybe like 50% of
5:44:26the way there instead of 90% of the way
5:44:27there. But with an AI assistant, it
5:44:29knows your name, your business, your
5:44:30priorities, your team, your current
5:44:32things. It knows the decisions you've
5:44:33made. And it can also do things for you
5:44:35like check in with the team. It can
5:44:36create stuff. It can research. It can
5:44:38plan your day. And this is the visual
5:44:39that we just generated together. And I
5:44:41actually like this more because it's
5:44:42showing that I'm able to create more
5:44:43YouTube videos because of this
5:44:45assistant. So, the benefits are pretty
5:44:46clear here. You can save a lot of time.
5:44:48You never have to really repeat
5:44:49yourself. You scale your team. And you
5:44:51could potentially sell this skill
5:44:53because now you understand how to set up
5:44:55these frameworks with context management
5:44:57and you know memory. And in this process
5:44:59of building your own assistant, you're
5:45:00going to get so much better at cloud
5:45:02code, which is a really good skill to
5:45:03have. So like I said, today I'm going to
5:45:05be showing you guys exactly how I built
5:45:06this thing. And there's four main
5:45:07phases. So phase one is we need to give
5:45:09it a home. So that's kind of like the
5:45:11structure of our project, which I think
5:45:12is the most difficult or kind of like
5:45:14the most confusing thing up front
5:45:16because as you guys saw in my project,
5:45:18there's a lot of folders and in each of
5:45:19these folders, there's a lot of
5:45:20subfolders. So there's a lot going on.
5:45:22So it's really important to be able to
5:45:23make sure that you know where stuff is,
5:45:25but that Claude also knows where
5:45:26everything is because as you scale, more
5:45:28and more files will be created, more and
5:45:30more skills, more and more processes.
5:45:31After we've given it a home, we need to
5:45:33give it some life. So we need to give it
5:45:34some rules. We need to give it some
5:45:35context about you. It needs to learn
5:45:37everything about you and what you're
5:45:39currently up to. After that, we need to
5:45:40give it hands. So we're going to build a
5:45:42first skill together. And then we're
5:45:43going to see how we can build more and
5:45:44more skills and sub agents and stuff
5:45:46like that so that it actually gets more
5:45:47useful at doing things for you rather
5:45:49than just like helping you think. And
5:45:51then finally talking about how we
5:45:52actually let this thing grow. How do we
5:45:54improve it? How do we really scale it?
5:45:56And how do we make sure that this
5:45:57assistant actually gets smarter over
5:45:58time and really is leverage for us. All
5:46:01right, so phase one, let's give this
5:46:03thing a home. So, if you guys have
5:46:05watched any videos on my channel with
5:46:06Cloud Code, then you've noticed I'm
5:46:08using Visual Studio Code, which is just
5:46:09an IDE, an integrated development
5:46:12environment. It's completely free to
5:46:13download for Windows or Mac. So, go
5:46:15ahead and grab that. And then once
5:46:17you're in there, this is what it should
5:46:18look like when you open it up. Now, what
5:46:20you have to do is come over to the lefth
5:46:21hand side, click on extensions, and
5:46:23you'll type in Claude Code, and it will
5:46:24look like this. And all you have to do
5:46:26is go ahead and install this. When you
5:46:28do that, it'll prompt you to sign in
5:46:29with your enthropic subscription, which
5:46:31you do need to be on a paid plan for
5:46:33Cloud Code. You could use your API key,
5:46:35but it's better to just have a fixed
5:46:37cost rather than worrying about how many
5:46:39tokens you are racking up. All right, so
5:46:41once Cloud Code has been installed,
5:46:42you'll notice in the top right there's a
5:46:44little orange button. If you click on
5:46:45that, this is where you have the Cloud
5:46:46Code actual agent that you can go ahead
5:46:48and start talking to. But before we
5:46:50start talking to it, we still haven't
5:46:51really given it a home. So, we need to
5:46:52set up a folder, which means you need to
5:46:54open up your file explorer or your
5:46:55Finder or whatever it's called, and
5:46:57create a new folder. So, for this, I'm
5:46:59just going to be in my desktop. I'm
5:47:00creating a new folder, and I'm just
5:47:01going to call this EA demo. EA for
5:47:04executive assistant. And now, what I
5:47:06have to do is open up that folder in
5:47:08Visual Studio Code. So, I'm going to go
5:47:10on the lefth hand side and click on the
5:47:11explorer. And it says you have not yet
5:47:13opened a folder. This will basically be
5:47:15your project. And I'm going to go ahead
5:47:16and click open folder and open that one
5:47:18up that I just made. So, here is my EA
5:47:19demo. Select folder. And now we are
5:47:22actually in our project. This is the
5:47:23home for our cloud code agent because
5:47:25what I can do is open up this little
5:47:26button. Close out of everything else.
5:47:28And now we have our folder on the lefth
5:47:30hand side. There's no files in there
5:47:31yet. And then we have our cloud code
5:47:33agent right here in the middle that we
5:47:35can talk to. So now the second piece of
5:47:36giving this a home is understanding how
5:47:39cloud.md works and creating our cloud.
5:47:41MD file. So if you've built an agent in
5:47:43naden or you've built a custom GBT with
5:47:45chat GBT, you understand that when you
5:47:47do something like that, you have to give
5:47:48it a system prompt. you have to give it
5:47:50instructions. So that's exactly what we
5:47:51need to do here and we do that with a
5:47:53claw.md file. So the way that this works
5:47:55is you send something to cloud code and
5:47:58before it actually reads your message,
5:48:00it's going to load in the claw.md file
5:48:02and read the entire thing and it's going
5:48:03to do that every single time. So the
5:48:05claw.mmd file should have only the most
5:48:08important rules about this is what the
5:48:09project does, this is where you need to
5:48:11look for your rules, this is where you
5:48:12look for context, that kind of stuff.
5:48:15because if you fill it super super full
5:48:16of random information, then you're going
5:48:19to go through your tokens and your
5:48:20context window faster. So, what's
5:48:22ultimately going to happen by the end of
5:48:23this video is you will have not only a
5:48:25cloudMD file, but you'll have acloud
5:48:28file, you'll have projects, you'll have
5:48:29context, you'll have decisions, you'll
5:48:31have a bunch of different folders in
5:48:32here. But the brain cloud.mmd tells
5:48:35cloud code where does everything
5:48:37actually live. So, that's how we stay
5:48:38really organized. So, all we're going to
5:48:40do in here to start is we are going to
5:48:42come over to the lefth hand side. I'm
5:48:43going to click on new file and I'm just
5:48:45going to call it in all caps claude and
5:48:48then MD. And MD just stands for
5:48:50markdown. And so what happens is the
5:48:52cloud.mmd file pops up right here. And
5:48:54you can see that there's not currently
5:48:55anything in there. So I'm just going to
5:48:56say to claude, hey claude code, this
5:48:59folder is for you to be my executive
5:49:01assistant. So just throw a quick blurb
5:49:03about that in the cloud.mmd file. I
5:49:05shoot that off. It's basically going to
5:49:07look through the project, see what's in
5:49:08here, and then it's going to edit that
5:49:10file so that we have a little blurb in
5:49:12there. And unless you're on bypass
5:49:13permissions mode, it's going to ask you
5:49:15for permission here. So now it says
5:49:17done. Cloud.mmd has a quick description
5:49:19of your folder as an executive assistant
5:49:21workspace. If I click into the cloud.md,
5:49:23you can see this is what we've got. This
5:49:24folder is cloud code's workspace for
5:49:26acting as an executive assistant. Use it
5:49:28to help the user with scheduling, task
5:49:29management, research, drafting,
5:49:30communications, and any other EA related
5:49:32work they need. So this is going to
5:49:34evolve throughout the video. You guys
5:49:35will see that once we start to give it
5:49:37some more life, which as you guys know
5:49:39is phase two. So, like I said, cloud.mmd
5:49:42is the brain, and it's going to tell
5:49:43cloud code where to look for information
5:49:45about us, which is going to be in a MI
5:49:47file. Information about your business,
5:49:49which will be in work, information about
5:49:50your team, which will be in team, and
5:49:52then information about what you're
5:49:53currently focusing on right now, which
5:49:54will be in priorities. And it will also
5:49:56understand all of your rules, like the
5:49:58way you like to speak, um, your style,
5:50:00formatting, stuff like that. Okay. So,
5:50:03I'm going to paste in this prompt, which
5:50:05you guys will be able to access for free
5:50:06in my free school community. I'll have a
5:50:08post in there associated with this video
5:50:10and then you just need to basically grab
5:50:12that markdown file, copy it and paste it
5:50:14into here. You can see that this prompt
5:50:16is pretty beefy and it's going to
5:50:17basically walk you through and have
5:50:19Claude Code extract all the information
5:50:21out of you that it needs to get this
5:50:22project started. So, I'm going to shoot
5:50:24this off and we're going to see that
5:50:25it's going to start to ask us some
5:50:26questions about us. So, phase one is
5:50:28creating the folder structure and it's
5:50:30initializing a git repo. And now you can
5:50:32see that all of these folders and files
5:50:34have popped up on the lefth hand side
5:50:35which is very similar to how my Herk 2
5:50:37project was set up. We've got templates,
5:50:38references, projects, decisions,
5:50:40context, archives, and the claude. So
5:50:42all of this stuff is going to start to
5:50:43get filled up a little bit. And trust
5:50:44me, as you start to use this, it will
5:50:46make much more sense. So now we are on
5:50:48phase two, which is the interview
5:50:49section. So first part, what is your
5:50:52name? What's your role? What's your time
5:50:53zone? Blah blah blah. So I'm just going
5:50:55to give it some dummy information here.
5:50:56Okay, so I shot off my initial answers
5:50:58and now it's asking us more about our
5:50:59business and our work. So, obviously I
5:51:01am just kind of giving it some dummy
5:51:03information to show you guys how it's
5:51:04going to set up these folders, but this
5:51:06is where you should really take some
5:51:07time here and let it get to know you and
5:51:09give it as much detail as you want
5:51:11because ultimately you're going to want
5:51:12it to know all of this stuff either way.
5:51:14So, take your time here. Really give it
5:51:15information. If it asks you something
5:51:16and you don't know the answer to it,
5:51:18maybe just say that you don't know and
5:51:19see if it can help you brainstorm some
5:51:21stuff. But also, you could just say
5:51:22skip, you know, I'll set this up later.
5:51:24And you can see that it might not move
5:51:25on to the next section until it feels
5:51:27like it has enough information. So,
5:51:28right here, it asks if I had anyone key.
5:51:30I said yes, I have one other person and
5:51:32I didn't say the name. So, it's asking
5:51:33what is the name? And now it's moving on
5:51:35to priorities, goals, and projects.
5:51:37Remember that you'll be able to plug it
5:51:38in real time into ClickUp or, you know,
5:51:40ASA or notion or whatever you use for
5:51:42your project management and your goals
5:51:45because a big part of this is making
5:51:46sure that everything it's looking at is
5:51:47actually current. This is just kind of
5:51:49the initial onboarding to get it
5:51:50familiar with your business. Section
5:51:53five asks about communication
5:51:55preferences. So, how do you actually
5:51:56like to interact with something? because
5:51:58this can be really flavored to you. And
5:52:00remember, none of this stuff is
5:52:00permanent. You can always change it
5:52:02later. And the last section is what do
5:52:03you want help with? So like what are the
5:52:05recurring tasks that eat up your time?
5:52:07What would you hand off to an assistant
5:52:08first if you could? Are there any
5:52:09specific workflows that you want to
5:52:10automate or templatize? Now, if you
5:52:12don't know right now, that's fine. Just
5:52:14say skip because what's going to happen
5:52:15is I'm going to challenge you to use
5:52:17this as much as you can. Don't use your
5:52:19custom GBTs. Don't use your projects.
5:52:21Try to migrate everything into here and
5:52:22just use this. And over time, you'll
5:52:24realize what is recurring and what are
5:52:26processes that you can actually use in
5:52:27cloud code. So I just told it let's skip
5:52:29that for now and let's just keep on
5:52:31moving through the setup. So now that it
5:52:33has all of our information, it's going
5:52:34to build out the files based on our
5:52:36answers. And we're going to be able to
5:52:37see that right now it's writing the MI
5:52:39file and now it's writing the work file
5:52:41and now it's writing the team file and
5:52:43the current priorities. And all of this
5:52:44is going to get looped back together.
5:52:46All right, so all of this is set up. We
5:52:48have our tree view of our folder
5:52:50structure. So we can really dig into
5:52:51this if we want to see what's in the
5:52:53cloud or what is in the archives, the
5:52:55context, all this kind of stuff. We get
5:52:57a summary of how everything works. So if
5:52:59you're confused about any of this stuff,
5:53:00you will get a summary and you can also
5:53:02ask, hey, like what does this context
5:53:03file do or what does the archives folder
5:53:05do? And it will be able to explain it
5:53:06for you. But you can see it's populated
5:53:08this stuff with information about me.
5:53:09The skills that we need to build are on
5:53:11the backlog. Otherwise, if you listed
5:53:12some, it would say, okay, cool. Let's
5:53:13just start building those skills right
5:53:15now. And as far as keeping the assistant
5:53:16sharp weekly, nothing required right
5:53:18now. We have auto memory for daily
5:53:20learnings. Monthly, we'll update this
5:53:21stuff. Quarterly, we'll update this
5:53:23stuff. And as needed, we will log
5:53:24decisions in the decision log. And pro
5:53:26tip, if you want your assistant to
5:53:27remember something permanently, just
5:53:28tell it remember that I always prefer X.
5:53:30And it will save that across all future
5:53:32conversations. And then the last thing
5:53:34it's going to do is an initial git
5:53:35commit. And this is just local. It's not
5:53:37going to do a actual GitHub repository
5:53:39out on the cloud. So, this will just
5:53:41kind of locally store these files so
5:53:43that you can have some version control.
5:53:45But if you want to, you could just say,
5:53:46"Hey, instead of doing this here, let's
5:53:48actually just do it on GitHub." And then
5:53:49you would just give it your email. It
5:53:51would be able to help you sign in. And
5:53:52then it can actually just make the repo
5:53:53for you, and it can handle all of those
5:53:55future commits and pushes. So the
5:53:57benefits of that is that in GitHub, it
5:53:59basically stores all these folders, all
5:54:01these files, which means from any
5:54:02device, you could basically pull in that
5:54:04repo into cloud code and you could
5:54:05always have your executive assistant
5:54:06ready. You've got cloud backups, you've
5:54:08got rollbacks, you've got collaboration,
5:54:10and you've got branching. So it's just
5:54:12best practice to put your codebased
5:54:13projects into GitHub. So let's just real
5:54:16quick take a look at what actually
5:54:17happened. So the first thing I want to
5:54:18look at is the dotcloud. You can see in
5:54:20here right now we have rules with
5:54:22communication style. So if I open up
5:54:23this, we can see that it threw in some
5:54:25information about the way that we like
5:54:26to talk. So bullet points, everything
5:54:28concise, no m dashes, internal speak
5:54:30casual, external speak even more casual.
5:54:32If we go to the skills, we can see
5:54:34there's nothing currently in there.
5:54:35There's nothing in the archives folder.
5:54:36In the context folder, we have current
5:54:38priorities. This is everything that I
5:54:39just talked about during our setup. It
5:54:41also says when this was last updated,
5:54:43which is nice. We've got our goals and
5:54:45milestones, which it says update this
5:54:47file at the start of each quarter. So,
5:54:48that's good. We've got 2026 annual goal,
5:54:51Q1 2026, Q2 2026, informal milestones.
5:54:55We've got the MI file, which is going to
5:54:56evolve a lot over time. You can even
5:54:58give it information about your
5:54:59background, where you grew up, all this
5:55:01kind of stuff, and it can use all of
5:55:03that to tailor it even more. We've got
5:55:04the team markdown, so anything that you
5:55:06needed to know about some of the key
5:55:08people in your organization. And then
5:55:09we've got the work MD, which has some
5:55:11business and company information. In the
5:55:13decisions folder, we have a log. So,
5:55:15anything major that happens, it will be
5:55:16logged here with the date, the decision,
5:55:18the reasoning, and the context. We've
5:55:20got our projects, which it created a
5:55:21folder for each of them. So, we've got
5:55:23chocolate pistachio flavor, we've got
5:55:24website launch, we've got West Coast
5:55:26expansion, we've got winter events. And
5:55:28each of these have a readme file which
5:55:29basically is just a quick description of
5:55:31what this project actually means and you
5:55:33know the status or any other information
5:55:34about it. We've got a references folder
5:55:36where we'll be able to drop examples and
5:55:38SOPs. There's nothing in there yet. And
5:55:39then for templates we also have you know
5:55:41we're able [snorts] to drop in some
5:55:42stuff that it can reference and this is
5:55:44just an example session summary. Nothing
5:55:46is in there for this session yet. And
5:55:48then the last thing to look at here
5:55:49would be the claw.md file. So earlier it
5:55:51was just like a few lines and it was
5:55:52very basic but now it's tailored towards
5:55:54us. So, you are Jack's executive
5:55:56assistant. Be direct, concise, and
5:55:58casual. Here is Jack's top priority.
5:56:00Here is what's very urgent. This needs
5:56:02to happen ASAP. And so, this is getting
5:56:03read every time so that it can basically
5:56:05keep checking in. Hey, is this done? Is
5:56:06this done? Now, here's something really
5:56:08cool where the clawmd file can point to
5:56:10the right files. So, remember how I said
5:56:12that this got loaded in every single
5:56:14time you talk to cloud code? That means
5:56:16if we threw all of our business
5:56:18information, all of the information
5:56:19about you, your priorities, that would
5:56:20be a lot of tokens. So remember that all
5:56:22claw.mmd has to do is tell cloud code,
5:56:24hey, if you need some information about
5:56:26the current focuses, go read this file.
5:56:28So that's exactly what you're seeing
5:56:29right here. It says, hey, if you want to
5:56:31understand who Jack is, go read this. If
5:56:33you want to understand, you know, the
5:56:34business details, go read this. And so
5:56:37on and so forth. So that's how we're
5:56:38able to save tokens here, but still give
5:56:40Cloud Code all the information that it
5:56:41needs. It will also tell Cloud Code what
5:56:43tools it has access to and how to use
5:56:45them. Right now, we haven't really set
5:56:46anything up. And similar concept over
5:56:48here with our projects. It says, "Hey,
5:56:50if you want to look at this project, go
5:56:51here. This project, go here." You'll
5:56:53notice right now we're at about 87
5:56:54lines, which is pretty good. I always
5:56:56try to keep them, you know, under about
5:56:58150, maybe 200 max. That's just best
5:57:00practice. Over time, they're going to
5:57:02get larger and larger, so you should
5:57:03regularly be kind of compacting them
5:57:05down and pointing out to other files
5:57:07when possible. Now, this is a great
5:57:09place to start, but at some point,
5:57:10you're going to want to add some more
5:57:11files on your own. So, this is what
5:57:12you've got right now, but what you can
5:57:14see is in my Herk 2 version, I've got a
5:57:16few more. And one of the ones that I
5:57:17added that you'll probably want to add
5:57:18is a brand assets. So in here I've got
5:57:20fonts and I've also got some images,
5:57:23logos, brand guidelines, things like
5:57:24that. So all you do if you wanted to add
5:57:26a new folder or new files is you would
5:57:28just come in here, add a new folder,
5:57:30name it whatever you want, and then just
5:57:31start a conversation with cloud code and
5:57:32say, "Hey, I added a folder called
5:57:34brand- assets. Just update your skill
5:57:36documents or update the cloud.MD to know
5:57:38that it's there. And what I'm going to
5:57:40drop in there are logos or head shot or
5:57:42things like that." So all of that can be
5:57:44obviously customized as long as claw.md
5:57:46understands that. So that's basically
5:57:48the way that our project is set up. It's
5:57:50now got a home. It's got some life. And
5:57:52let's move on to phase three, which is
5:57:54giving it hands. Actually letting it do
5:57:56something for us. Okay. So if I was you
5:57:58guys right now setting up this executive
5:58:00assistant, the first skill that I'd
5:58:01probably build is connecting it to
5:58:03ClickUp or Notion or wherever you have
5:58:05your actual like project management or
5:58:07task management. And I'm not going to
5:58:08walk through that right now because I
5:58:10use ClickUp and I don't want to go
5:58:11through that setup in front of you guys
5:58:12because it might not apply to you
5:58:14specifically if you don't use ClickUp.
5:58:15But basically, it's super easy. You
5:58:17explain in natural language exactly what
5:58:18you want to do. You have it do research
5:58:20on the endpoints or maybe even an MCP
5:58:22server. And then all you have to do is
5:58:24go grab an API key and put it in av
5:58:26file, which you'll notice right here we
5:58:28don't currently yet have. So here's what
5:58:30you would do. I'm going to go to plan
5:58:31mode and I'm going to say, "Help me
5:58:34build a research skill. This skill is
5:58:37going to use perplexity. So I need to
5:58:38give you my API key in AENV file. So go
5:58:41ahead and create that file for me. And
5:58:42what I want the skill to do is help me
5:58:44do research. This is more than just a
5:58:46simple couple, you know, web searches
5:58:48and web fetches that you might be able
5:58:49to do already. This is research that's
5:58:51kind of deeper and it's also tailored
5:58:54towards me and my business because it
5:58:55understands the context of, you know,
5:58:56what's going on and our current projects
5:58:58and priorities. So I'm going to shoot
5:59:00that off in plan mode. Cloud code's
5:59:01going to think about it. It might do a
5:59:02little bit of research to help us build
5:59:04this plan and then it's going to come
5:59:05back with something. And what's pretty
5:59:06cool is you can see that first of all it
5:59:08explores the project to see how things
5:59:10work and it even spun up a sub agent to
5:59:12explore the structure of our project
5:59:14right here. And sub aents are really
5:59:16cool because they have their own context
5:59:17window and they might even be able to
5:59:19use their own model if you you know
5:59:20configure one to do so. So I'll touch on
5:59:22that a little bit more after we build
5:59:24the skill. So I may come back and ask
5:59:25you some questions which is a good
5:59:26thing. So here it says for perplexity
5:59:28what model do you want to use? And I'm
5:59:29just going to go with sonar for now and
5:59:31submit that off. And now it thinks it's
5:59:33good and it's come back with a plan. And
5:59:34we can see we have build research skill
5:59:37with perplexity. Jack wants a reusable
5:59:39research skill that goes beyond basic
5:59:40web search. It's going to create thev
5:59:42for our API key. Now the reason why it
5:59:44needs to do that is because if you do
5:59:46end up pushing this to a GitHub repo or
5:59:48putting that anywhere on the web, it's
5:59:49best practice to have. Eenv, which is
5:59:51where you have all of your API keys, and
5:59:53you don't give it to Claude in the
5:59:55conversation history right here. It's
5:59:56going to build the actual skill itself.
5:59:58So we'll see that once it's created. And
6:00:00the skill instructs Claude to understand
6:00:01the question, load context. So read
6:00:03through me, work, team, current
6:00:05priorities, goals, and anything else.
6:00:07Formulate queries, call perplexity API,
6:00:09synthesize, save report, and then
6:00:11present the summary. Another thing worth
6:00:12noting is that it's also going to update
6:00:14the claw.md so that it now knows about
6:00:16this skill and knows that it exists. So
6:00:18anyways, it looked at the API
6:00:19documentation. It gave us a research
6:00:21report format. I'm just going to go
6:00:22ahead and auto accept this plan, and
6:00:24it's going to start building it out for
6:00:25us. All right. So, the agent just
6:00:27finished up. It said, "Here's what we
6:00:29created. We have thev, which you can go
6:00:31drop your API key in. We've created the
6:00:33research skill." So, if I go to Claude,
6:00:34if I open up the skills, we can see we
6:00:36have a research folder. And in here, we
6:00:38have a skill.md for research. So, I'm
6:00:41not going to read this whole thing, but
6:00:42you can see now that we have this actual
6:00:43skill. And as you can see, this
6:00:45basically just says when do you use the
6:00:46skill and how do you actually run the
6:00:48process of, you know, calling the API
6:00:50and making the report for Jack. And it
6:00:53updated the claw. MD to show Claude that
6:00:55this skill now exists. So, I'm going to
6:00:57go ahead and grab my API key from
6:00:58Perplexity and paste it into the ENV. As
6:01:01you can see, I'm just going to drop it
6:01:02in right here and then save it. Okay, I
6:01:05put the API key in there and we're going
6:01:06to test it out. I said research ice
6:01:08cream events in Portland using the new
6:01:10research skill. And what we're looking
6:01:11for here is seeing if it's actually able
6:01:13to call the skill. So, it says, okay,
6:01:16let me load the relevant project context
6:01:18and then I'll run the research. The
6:01:20content has been loaded. Now I'm going
6:01:21to break this up and I'm going to call
6:01:23perplexity. So here you can see it did
6:01:25three different searches. It searched
6:01:27for Portland ice cream dessert events
6:01:29and then it searched for Portland vendor
6:01:31application process and finally it
6:01:33searched for Portland dessert scene and
6:01:34connections. So it comes back with a
6:01:36very concise output which is exactly
6:01:38what we told it in the way that we like
6:01:40to communicate. We got this information
6:01:42right but then it also says that I saved
6:01:45the full report to this folder. So, if I
6:01:48go to research, which it created a new
6:01:50folder right here, we can see that we
6:01:52have March 3rd Portland ice cream events
6:01:55research. And this is a markdown file
6:01:56that has much more detail about what it
6:01:58actually found as far as these different
6:02:00events and these different, you know,
6:02:01vendor things that we might be
6:02:02interested in. And it includes all of
6:02:04the links to all the sources that it
6:02:06actually pulled here. And what's great
6:02:07about this and why you guys can probably
6:02:09now see that the folder management is so
6:02:12important because every time you do
6:02:13research, every time you generate
6:02:14content, every time you make updates to
6:02:17projects, all of that still lives in
6:02:19this project, which is why it gets
6:02:20smarter and smarter about what you're
6:02:22doing the more you use it because I
6:02:24could basically clear out this context
6:02:25and I could pick up right where I left
6:02:27off because it's able to go read this
6:02:29file, this research report. So, that was
6:02:31a quick preview of a skill. Now, there
6:02:33are going to be some times where you
6:02:34might actually want to use a sub agent
6:02:35instead of a skill. Now, as you guys
6:02:36just saw in the demo, when you're
6:02:37talking to the agent and you call a
6:02:38skill, it basically just uses the skill
6:02:40right here in this context window with
6:02:42this model with that, you know,
6:02:44conversation awareness. But a sub agent
6:02:46basically gets called on by this main
6:02:48worker here and it has fresh memory,
6:02:50fresh context, and you can even use a
6:02:51different model. So sometimes if you
6:02:53want to do research, but you want it to
6:02:54be cheaper, you could just tell it to
6:02:56delegate that to a sub agent and that
6:02:58researcher sub agent could maybe use
6:02:59haiku instead of opus 4.6. So that's
6:03:02kind of the benefit of a sub aent.
6:03:07So, let me just show you guys real quick
6:03:08how that would actually work. I'm going
6:03:09to go back into plan mode and I'm going
6:03:11to say I need to create that research
6:03:15skill. So, basically everything you just
6:03:16did with the research skill, but I want
6:03:18to create that as a sub agent and I want
6:03:20this to be using haiku instead of opus
6:03:24just in case I need to do some cheaper
6:03:26research. So, create that sub agent for
6:03:28me. Put that within theclaw folder in a
6:03:31folder called agents. So we now have the
6:03:33plan to create that research sub aent
6:03:35and you can see that it's going to use
6:03:36model equals highq and I'm just going to
6:03:38go ahead and accept these changes. So
6:03:40what it did is it created the agent.mmd
6:03:42which we'll go ahead and take a look at
6:03:43and I believe that it also updated the
6:03:45research skill so that if I say that I
6:03:47want it to be like cheaper or to use a
6:03:48sub aent that the research skill will
6:03:50actually invoke the sub aent which is
6:03:52pretty cool. And then it also added an
6:03:54update to the cloud.MD file which now
6:03:56should have an agent section. You can
6:03:58see there's been a new folder created.
6:03:59We have agents under our do.claude and
6:04:01in here we have a research agent. So if
6:04:03I open this up, you will see that we
6:04:04have you are a research agent for ice
6:04:06cream Fridays. You receive a research
6:04:08query, call the perplexity API and save
6:04:10a structured report. So I'm trying to
6:04:12keep this as simple as possible, but
6:04:14there is something that I want you guys
6:04:15to notice about both the agent and the
6:04:18skill that was just created here. So
6:04:20these were created as complete markdown
6:04:21files which is fine. But actually the
6:04:23way that we need to set up both of these
6:04:25is using a YAML front matter because
6:04:27that's where we can actually have the
6:04:28configuration. So I'm not going to dive
6:04:30super deep right now into exactly why
6:04:32this is but at a high level it makes
6:04:34cloud code better at understanding what
6:04:36that skiller agent does when to use it
6:04:38and it's going to save you tokens as
6:04:40well. So for now I'm going to manually
6:04:42build this YAML front matter by
6:04:44literally just going like this going
6:04:46back into RS code and then for the skill
6:04:50I'm just going to paste that up here and
6:04:51then change the name and change the
6:04:53description myself and I'm going to do
6:04:54the exact same thing then with the
6:04:55agent. But what I would actually
6:04:56recommend you do is in your project give
6:04:59cloud code this document. You know give
6:05:01it the example give it even the URL and
6:05:03say hey look up on this page what is the
6:05:06best practice for creating sub aent
6:05:08files or creating skill files so that
6:05:10every time you build me a new sub aent
6:05:12or you build me a new skill it does it
6:05:14in the best practice. So really the
6:05:16reason I showed you guys this and the
6:05:17point I'm trying to make here is cloud
6:05:18code is smart but it's way smarter if
6:05:20you let it do its own research and then
6:05:22save that research so that in the future
6:05:24everything gets better and better. So
6:05:25now you can see I've updated both the
6:05:27agent and the skill with the YAML front
6:05:29matter and now it's going to work a lot
6:05:31better. It would still work the other
6:05:32way but trust me this is just going to
6:05:34save you tokens and you're probably
6:05:36going to get more consistent results.
6:05:37All right so now let's go ahead and test
6:05:38that out. I'm going to go ahead and
6:05:39clear this conversation and I'm going to
6:05:41say I need you to do some really quick
6:05:43research for me. I want to keep this one
6:05:45cheap, so just use the research agent
6:05:47and find out, you know, how ice cream
6:05:49events are going in Los Angeles. All
6:05:51right, so this is pretty cool. It read
6:05:53through the research skill. It read
6:05:54through the research agent. And then it
6:05:56also read through my plans. So it said,
6:05:57"Hey, Jack wants quick, cheap research
6:06:00about LA. This ties directly into the
6:06:02West Coast expansion priority." So now
6:06:05it has all that context. It calls the
6:06:06research agent with the model Haiku. And
6:06:08now it's going to go ahead and do that
6:06:10research for us. All right, awesome.
6:06:11That research is done. If I open up the
6:06:13research folder, you can see that right
6:06:14here we've got our LA report about some
6:06:17ice cream stuff. So, just keep in mind
6:06:19the research report is still really
6:06:20good. It still has lots of sources. It
6:06:22still has lots of detail. It wasn't the
6:06:24research that was worse. It was just
6:06:26when all of that raw data got passed
6:06:28back to Claude Code. It was using Haiku
6:06:30to summarize it and to write this MD
6:06:31file rather than having Opus do that.
6:06:33Okay, so we've set up our executive
6:06:35assistant. We have given it information
6:06:36about us and our business. We have
6:06:38connected it with a skill and an agent.
6:06:40And now phase four is basically just
6:06:42about letting it grow. Now I could walk
6:06:44through some more step-by-step examples
6:06:45with you guys, but really this is about
6:06:47customizing it for your tech stack,
6:06:49your, you know, project management
6:06:50system, your team, and the way that you
6:06:52work. So with my cloud code assistant
6:06:54here, because I have all these different
6:06:55agents now and these different rules and
6:06:57these different skills, it's just
6:06:58because I used this every day instead of
6:07:00using like Claude or Chatbt on the web.
6:07:03It gets smarter and smarter over time.
6:07:04So right now you're at day one and if
6:07:06you use this every day, a month from
6:07:08now, this thing is going to look crazy
6:07:09different. there's going to be way more
6:07:11docs, way more decisions, way more
6:07:12skills, and it's just going to be
6:07:14smarter and smarter, and you're going to
6:07:15be like living in that thing. So,
6:07:17really, what I would challenge you to do
6:07:18from here is to only use this. Just like
6:07:21for the next week, try only using this.
6:07:24Take all of your cloud projects, take
6:07:25all of your custom GBTs or, you know,
6:07:27your gems, whatever it is, take those
6:07:28instructions, put them into this
6:07:30project, and say, "Hey, turn this into a
6:07:31skill for me." Use that skill, use new
6:07:33skills, and every time that you're using
6:07:35the skill, say, "Hey, this is what I
6:07:37didn't like. This is what I did like.
6:07:39Update it. you know, let's let's just
6:07:40keep making it better and better. And
6:07:41that same exact theory applies to the
6:07:43entire system, the cloud files, the
6:07:45reference files, the project files,
6:07:47everything. And then, like we talked
6:07:48about earlier, put this thing on GitHub,
6:07:50have version control, and make sure that
6:07:52if you switch to a different laptop for
6:07:53some reason, you can still use your
6:07:55executive assistant. But anyways, that's
6:07:57all I got for this one. Now, I'm sure
6:07:58you guys are all excited to keep
6:07:59building on top of this executive
6:08:01assistant. So, what's next is really
6:08:02just to go master skills.
6:08:06All right, this is where it starts to
6:08:07get like really really fun. We're about
6:08:09to dive into skills, sub aents, agent
6:08:12teams, and a few other things. But this
6:08:13is where we get so so powerful with our
6:08:15Claude code setups.
6:08:19I have genuinely never been as
6:08:20productive as I am right now because of
6:08:22Claude skills. This image really does
6:08:24sum it up. I feel like I'm working on
6:08:25multiple different computers and
6:08:26multiple different tasks at the same
6:08:28time without sacrificing quality. It
6:08:30just comes down to one simple word and
6:08:32that is leverage. with Claude skills or
6:08:34any agent skills for that matter, you
6:08:35have way more leverage than if you were
6:08:37doing this by yourself. So, in this
6:08:38video, I'm going to be breaking down
6:08:40what skills are, how they work, and how
6:08:41you can build really, really good ones,
6:08:43even if you've never heard of the
6:08:44concept or built a single skill ever
6:08:46before. So, let's hop into a live demo
6:08:48real quick and get going. All right, so
6:08:50here we are in Claude Code in my Herk 2
6:08:52project, which is kind of just like my
6:08:54personal assistant. Now, if this stuff
6:08:55over here looks overwhelming, don't
6:08:56worry about that right now. Just worry
6:08:58on what I'm actually asking the agent to
6:09:00do. So, I've got this skill that I
6:09:01called morning coffee that helps me plan
6:09:03my day every morning. So, it's not
6:09:04morning right now, but I'm going to run
6:09:06this so you can see how it works. So, as
6:09:07Cloud Code's figuring that out, what I'm
6:09:09going to do is open up another agent.
6:09:11And in this one, I'm asking it to run a
6:09:12pulse check on all my projects and
6:09:14commitments to see how things are going.
6:09:15I'm opening up another one to create me
6:09:17an Excal diagram of the difference
6:09:18between local AI models and closed
6:09:20source models. And let's just do one
6:09:22more that's going to scrape the comments
6:09:23from my recent YouTube videos and give
6:09:24me an analysis on what I need to
6:09:26improve. So, what you're looking at
6:09:27right now is four different agents
6:09:28running in parallel doing things for me.
6:09:30And that took me probably 30 seconds to
6:09:32ask them to do that. And because I built
6:09:34all of these skills, all of these agents
6:09:36have all of the context about my
6:09:37business, what's going on with our
6:09:38projects, my YouTube channel. It has
6:09:40literally everything it needs. And now
6:09:42all of those agents are done. Here was
6:09:43my February 26th morning coffee. I had
6:09:46three things on the calendar. So, what
6:09:47it's going to do is it's going to look
6:09:48at my ClickUp. It's going to see what
6:09:49else I've got this week and look at my
6:09:51tasks and then help me plan the rest of
6:09:52my day. So, this is the plan that it
6:09:54suggested. All I'd have to do is say,
6:09:55"Yep," and it would block off everything
6:09:57for me. And for me, that's huge because
6:09:58I don't have decision fatigue anymore of
6:10:00what I have to work on. The second agent
6:10:02came back with a pulse check, 2 days
6:10:03until the end of the month. Here's where
6:10:05everything stands. So, obviously, I'm
6:10:06going to blur all of this out, but it's
6:10:07basically catching me up on all of the
6:10:09different main initiatives that we're
6:10:10doing this month and this quarter and
6:10:12making sure that everything is on track.
6:10:13And right here, you can see there's a
6:10:14few things that I need to follow up on
6:10:16manually. And this might have slipped
6:10:17through the cracks because I'm so busy
6:10:18making YouTube videos if I didn't have
6:10:20this personal assistant to check up on
6:10:22me using the skill. The third agent came
6:10:24back and has finished the excal job
6:10:26diagram and I pasted it in and it looks
6:10:27like this. So if I needed to make a
6:10:29video about this, I wouldn't have had to
6:10:30take my own time to create this
6:10:32visualization. So for the comment
6:10:33analysis that came back and we can see
6:10:35all of the comments, all the views and
6:10:37things that I need to address either in
6:10:39future videos or in the comments. We got
6:10:40some confusion. We've got some cost
6:10:42things that I need to cover. We have to
6:10:43stop demoing toy examples for tool
6:10:46videos. Seems like you guys really want
6:10:47to see some anti-gravity stuff. I
6:10:48promise that'll come soon. And then
6:10:50we've got top three priorities. So, I've
6:10:52been recording now this video for about
6:10:536 minutes. Just think about if I would
6:10:55have done all four of those things
6:10:56myself, how much context switching I
6:10:58would have done and how long that would
6:10:59have taken me. Okay, so now that you
6:11:01guys have seen an actual demo and
6:11:03hopefully you're a little bit excited to
6:11:04learn about cloud skills if you haven't
6:11:05used them yet, what actually are they?
6:11:07So, skills are reusable instructions.
6:11:10You write them once, you save them as a
6:11:12skill. You can trigger them anytime, and
6:11:13you're going to get way more consistent
6:11:14results because it's going through that
6:11:16same process every time. So this visual
6:11:18right here was actually an AI generated
6:11:19image that I used with a skill. It was
6:11:22this one right here in my cloud code
6:11:23called Excalaw visuals. But sometimes
6:11:25with AI generated images, they don't
6:11:27spell words right. As you can see here,
6:11:29this is all messed up. This stuff is
6:11:30messed up. So I also have one, as you
6:11:32saw in the demo, to create Excal
6:11:34diagrams. And this one creates the
6:11:36actual Excal that I could move and edit.
6:11:38And the words are always perfect because
6:11:40it's actually just typing. And those two
6:11:42skills alone have saved me so much time.
6:11:43And what I'm going to do, by the way,
6:11:44for all you guys, is in my free school
6:11:46community, the link for this is down in
6:11:47the description, I've added a new
6:11:48classroom section called agent skills,
6:11:50where I'm just going to be dropping a
6:11:51ton of these skills that you guys can go
6:11:53grab for completely free. So, before we
6:11:55dive into this today, just real quick,
6:11:56why should you care? And there's three
6:11:58big reasons. You can be way more
6:12:00productive as a person because you can
6:12:01automate things like you just saw me do,
6:12:03and you can legitimately build a
6:12:04personal assistant that can do almost
6:12:06anything for you. The second one is team
6:12:08leverage. So, you can turn existing SOPs
6:12:10into automations really, really easy.
6:12:12And if you build something new, not just
6:12:14you can use it, but your entire
6:12:15organization can. So, everyone as a
6:12:17group is getting way more productive,
6:12:19which will almost undoubtedly result in
6:12:21growth of the business. And also
6:12:23monetization. We're entering this new
6:12:24world where skills are having a big
6:12:26moment and you're able to capitalize on
6:12:28a lot of this stuff. Now, I'm not saying
6:12:29this is going to be a viable business
6:12:30model for a long, long time. So, you
6:12:32shouldn't bank on it, but it is
6:12:34something to be aware of just like when
6:12:35people were selling edit workflow
6:12:36templates and things like that. But once
6:12:38again, it just comes down to one word,
6:12:40which is leverage. This isn't just
6:12:41theory. This is something that we're
6:12:42seeing with clients. This is something
6:12:44that we're seeing internally in my own
6:12:45business. This speed of work that we're
6:12:47able to achieve right now feels insane.
6:12:49But that is going to become normal. And
6:12:51if you can't do that, you instantly
6:12:52become way too slow and way too
6:12:54expensive for the business. And they
6:12:55might not keep you around. We're
6:12:56actually making it a priority to make
6:12:58sure all of our employees are using
6:13:00cloud code. Because now I have all these
6:13:01different skills that I can just run
6:13:03with a simple slash command or a simple
6:13:04natural language prompt and get in one
6:13:06day a week's worth of output. Because
6:13:08once again, one person can figure out
6:13:10the best way to do something and turn it
6:13:11into a skill that the entire team can
6:13:13use. But they don't just generate text.
6:13:15They're basically automations. They can
6:13:17run scripts. They can call APIs. They
6:13:18can create things. They can have sub
6:13:19agents. And they can be called on from
6:13:21agents as well. So this is truly AI
6:13:23automation. And just to really hammer it
6:13:25home, they're basically SOPs for your AI
6:13:27agents. The same way where you would
6:13:29train a human employee by letting them
6:13:30read through an SOP to understand the
6:13:32process and then they'd be able to do
6:13:33it, you just train an agent on it. You
6:13:35give them the skill, they read it, and
6:13:37then they do it. And the coolest part
6:13:38about it is the more you use the skill,
6:13:40the better and better it gets. So, we've
6:13:42talked about a lot of these benefits,
6:13:43right? But what actually is a skill?
6:13:45Well, it's just a folder and it lives
6:13:47somewhere in your project. The most
6:13:48common example is probably going to be
6:13:50in your do.claude/skills/skill
6:13:53name and then you've got like a skill MD
6:13:55or a markdown file. So, right here in my
6:13:57Herk 2 project, you can see up top we've
6:13:59got.claude. I open up.claude, we have
6:14:02agents, rules, and skills. Right now,
6:14:04we're just talking about skills. If I
6:14:05open that up, we can see all the
6:14:07different skills that I've created in.
6:14:08So let's say for example the Excal
6:14:10diagram skill. If I click into that, we
6:14:12have a skill.md. And when I open that
6:14:14up, you can see we have the name of the
6:14:16skill, the description, and then we have
6:14:17the actual workflow. Step one,
6:14:19understand the concept. Step two, plan
6:14:21the layout. Step three, generate
6:14:22elements. And this entire skill
6:14:24basically teaches my agent how to build
6:14:26these Excal diagrams for me. So like I
6:14:28said, that is the anatomy of the skill.
6:14:30We've got front matter which is kind of
6:14:32between these two dotted lines and that
6:14:34is in something called YAML which you
6:14:37don't need to worry about what that
6:14:38means. This is just the way that it's
6:14:40kind of indicated sort of like your
6:14:41markdown or your JSON or Python. Now up
6:14:44here we'll have a name and a description
6:14:46which tells Claude Code what the skill
6:14:48is called and what the skill does. So as
6:14:50you can see this one is called Excalraw-
6:14:52Diagram and there's a brief description
6:14:53about what it actually does or when to
6:14:55use it. And then we have the
6:14:57step-by-step rules, which are basically
6:14:58the instructions. And this is what
6:15:00Claude actually does once it decides
6:15:02that this is the right skill for the
6:15:03job. Now, the interesting thing about
6:15:05skills is that sometimes you need way
6:15:07more data. So, for example, let's say
6:15:09we're writing a LinkedIn post. We have a
6:15:11skill for that, right? But what needs to
6:15:13go in the skill is other information
6:15:15sometimes like a company tone of voice
6:15:17or maybe your LinkedIn, you know, tone
6:15:19of voice, a target avatar, current
6:15:21priorities, a logo. Maybe there are
6:15:23other things that you want to put into a
6:15:25skill besides just like the step-by-step
6:15:26instructions that will make it better.
6:15:28So, the question is, where do these
6:15:30things go? Well, there's typically two
6:15:32options, but essentially, as long as
6:15:33you're pointing to the right path in the
6:15:35skill.md, you're fine. So, let me
6:15:38explain what I mean by that, and then
6:15:39I'll show you what I mean by that. So,
6:15:40first option is to have it
6:15:42self-contained. So, in
6:15:43yourcloud/skills/skll
6:15:46name, you can have the skill. MD, you
6:15:48can have your scripts right there, and
6:15:50you can have your references right
6:15:51there. Or option B is that they're not
6:15:53directly nested right under that skill.
6:15:56So here we have.claude/skill/infographic
6:15:59and the skillmd and we still have our
6:16:00scripts and our references in the same
6:16:02project but it's just not nested
6:16:04directly under that skill. And so I know
6:16:06that might have made no sense. So let me
6:16:07show you exactly what I mean by that.
6:16:09Okay. So here we have a skill called
6:16:10idea mining. And so what happens in here
6:16:12is basically use when someone asks for
6:16:15content ideas, video ideas, what to make
6:16:16next or to run idea mining. And so in
6:16:19here I gave it some context, right? My
6:16:21channel has this many subs. It's about
6:16:22AI automation. My content pillars are
6:16:24naden, rag agents, cloud code, voice AI.
6:16:28And then what I gave it is a bunch of
6:16:29references. I gave it channel data,
6:16:31which is YouTube channel.md. I also gave
6:16:33it the raw data, which is a JSON file. I
6:16:35gave it a competitor list for me. And I
6:16:37gave it an actual script to run analysis
6:16:40on my YouTube channel. And so in this
6:16:42case, what you can see is that I went
6:16:44for option B where I'm storing those
6:16:46reference files and those scripts not
6:16:48directly nested in this skill. So
6:16:51basically what it could look like is
6:16:52within the skill itself, we could have a
6:16:55folder called, you know, references. We
6:16:57could also have a folder in here called
6:17:00scripts. And then within both of these
6:17:02subfolders, we could have more things
6:17:03like, you know, in the references, I
6:17:05could have channel data. And in the
6:17:07scripts, I could have
6:17:08YouTube-analysis.js
6:17:10js or py whatever it is. Basically the
6:17:13idea is it doesn't matter where those
6:17:15actual reference files or scripts live
6:17:17as long as you point to the right spot
6:17:18in the md file. So in my case where
6:17:20these actually live is in a different
6:17:22folder. So here for the channel data I
6:17:24would basically just go all the way down
6:17:25to references and then I could go down
6:17:27to right here YouTube channel.md. So
6:17:30cloud code reads the skill and then it's
6:17:32able to find this if it needs it. Same
6:17:33thing for the scripts. It would go down
6:17:35here to scripts and then it would find
6:17:37analyze YouTube. py and it would just
6:17:40pull this in if it needed it. So
6:17:41hopefully you guys are with me. However
6:17:43you want to set it up works. And I think
6:17:44that's the most overwhelming thing about
6:17:46cloud code right now is that everyone
6:17:47uses different kind of folder
6:17:49architecture. But don't worry guys, I'm
6:17:51totally on top of this. I have a skill
6:17:52that I built out called skilluer. And
6:17:55this one I'll be giving away for free
6:17:57once again in my free school community
6:17:58right here. And all you'd have to do is
6:18:00load in the skill builder and then it'll
6:18:02help you build out everything you need.
6:18:03And I'll be showing a live demo of that
6:18:05in a few minutes here. So the skill.md
6:18:07is the actual brain itself and the
6:18:09supporting files are the tools that it
6:18:11can use. That doesn't mean every single
6:18:13time the skill is invoked that those
6:18:14reference files will all be called. And
6:18:16just in case you guys were wondering if
6:18:17you've watched some of my previous cloud
6:18:18code videos where we've used the WAT
6:18:20framework to build automations. This is
6:18:22very very very similar in that
6:18:24framework. The W the workflows were the
6:18:26markdown file SOPs. That's basically the
6:18:28skill. The tools were the actual Python
6:18:31scripts and that's basically just the
6:18:32scripts that you might write or the
6:18:33references that you would add in. So, if
6:18:35you've already been building some WAT
6:18:36stuff, you will pick up skills super
6:18:38super quickly. And the cool thing about
6:18:39skills is that you don't have to build
6:18:40all of them. Obviously, as you're
6:18:42working with cloud code and you're
6:18:43finding that you're doing things
6:18:44repetitively, you can go ahead and build
6:18:46a skill for it. But there's an official
6:18:48library from Anthropic of Skills.
6:18:50There's a community of everyone that's
6:18:51open sourcing their skills and giving
6:18:53them out. And there's a marketplace
6:18:54where you can share and sell or you know
6:18:57download skills from people. And then
6:18:58you would take that skill or that
6:19:00essentially a prompt and you would add
6:19:02your own flavor to it. The one thing I
6:19:03would say is just be careful and make
6:19:05sure that no one's trying to, you know,
6:19:06give you a skill that has any malicious
6:19:08intent in there. And all these skills
6:19:10can work across different products. So,
6:19:12cursor, anti-gravity, codeex, because
6:19:15it's so based in Markdown and it's
6:19:17essentially just a prompt, tons of
6:19:18different AI models can use them. Okay.
6:19:20So, how does Claude know when to use a
6:19:22skill? Well, there are two ways to
6:19:24actually trigger them. The first one is
6:19:25you can be explicit, which basically
6:19:27means you can do a slash command and say
6:19:29the skill name and it will just directly
6:19:31fire off that skill. Or it could just be
6:19:33natural language. So if I had a school
6:19:35post skill, I could say
6:19:36slashschool-post.
6:19:38Or I could natural language just say,
6:19:39"Hey, help me write a school post about
6:19:41X." Claude would find that skill and
6:19:43then invoke it. So when you ask Claude
6:19:45to do something, it will first read
6:19:46through the cloud.MD file. It will
6:19:48analyze your request and it will search
6:19:50through the skills and see which one do
6:19:51I have that helps with this query. If it
6:19:54finds one, it will invoke it. But if it
6:19:55can't find anything, then it will
6:19:57basically just use its general
6:19:58knowledge. So not every single request
6:19:59that you give to cloud code will invoke
6:20:01a skill. Now a really important part of
6:20:03that is understanding how skills stay
6:20:05lightweight because if you've been using
6:20:07cloud code you know that context
6:20:08management is a huge deal. And if you
6:20:10had all of these skills to look through
6:20:12and all of these skills are I don't know
6:20:13hundreds and hundreds of lines then if
6:20:15cloud code was searching through all of
6:20:17these every single time that would
6:20:18surely eat up a ton of your tokens. So
6:20:20what's used is something called
6:20:22progressive context loading which
6:20:24basically means we have three levels.
6:20:26Level one is the initial search where
6:20:28cloud code only looks for the name and
6:20:30the description. So right here you can
6:20:31see let's say we ask for an Excal
6:20:33diagram. It would basically search
6:20:35through all the skills but it would only
6:20:36read the YAML front matter. So it would
6:20:38read the name and the description. And
6:20:40typically this front matter is only
6:20:41going to be you know maybe roughly 100
6:20:43tokens. So it stays very lightweight.
6:20:45And then moving down to level two let's
6:20:47say it identifies okay cool this is the
6:20:49right skill for the job. Then it would
6:20:51run the full skill.mmd and it would read
6:20:54through everything. And so that's when
6:20:55it would start to actually understand
6:20:56what goes on in the skill. And that
6:20:58might be anywhere from a thousand to a
6:21:01couple thousand tokens. And then level
6:21:03three is once again a decision. Only
6:21:05load in the extra files when needed. So
6:21:07if I need to look at any scripts or
6:21:08references or templates or I need to
6:21:10pull in some brand assets or more
6:21:12context, I'm only going to do that if
6:21:14the specific requests requires it. And
6:21:16so hopefully now you're starting to
6:21:18understand a little bit more about under
6:21:19the hood what's actually going on when
6:21:21you ask cla code to do something for
6:21:23you. And you can always go to cloud code
6:21:25docs and go to the skills section and
6:21:27just read about how this stuff works.
6:21:28It's really, really simple. On the doc
6:21:30itself, it will tell you just make sure
6:21:32to keep the skill.md under 500 lines.
6:21:34Move detailed reference material to
6:21:36separate files. And so, I know this may
6:21:38seem like it's just a lot of information
6:21:40being thrown at you. So, let me just
6:21:41kind of contextualize this and slow it
6:21:43down and reassure you guys. You're never
6:21:45ever ever going to write a perfect skill
6:21:48the first try. The way that I build my
6:21:49skills is I have Claude Code do
6:21:51something with me. I walk it through the
6:21:53steps, you know, each time. And then
6:21:55when we're done, if we've went from
6:21:56point A to point B, I say, "Cool. This
6:21:59is something I do once a day. Let's turn
6:22:01this into a skill. Ask me more questions
6:22:02so we can make sure you have all the
6:22:04information you need." And once again,
6:22:05I'm going to show you guys opening up a
6:22:07brand new project and setting up a skill
6:22:09from scratch so you understand the full
6:22:10process, but I just had to give you guys
6:22:12some context first. Now, we have this
6:22:14thing called the feedback cycle, which
6:22:15basically means you invoke the skill,
6:22:17you actually watch the agent work, you
6:22:19give feedback, and then it fixes the
6:22:21skill, and then you do it again. And so,
6:22:22the first couple times you run a skill,
6:22:24you may feel like, eh, this feels very
6:22:25AI generated. But by the time you've run
6:22:27that skill 10, 20, 30 times, every
6:22:30single time it gets better. And so,
6:22:32that's why it's actually important to
6:22:33watch the agent work the first couple of
6:22:34times because that's how you're able to
6:22:36identify opportunities to speed it up
6:22:38and save tokens by doing things like
6:22:40this. So here's an example of the pulse
6:22:42check skill that we actually ran
6:22:43earlier. Now this skill gets invoked
6:22:44when I ask for a pulse check or checking
6:22:46in on commitments. And what it does is
6:22:48it reads through some context of how
6:22:50OTAAS work, which is important for it to
6:22:51understand every single time it reads
6:22:53the skill, which is why I put it here
6:22:55rather than a reference file. And what
6:22:56it has to do is it has to do a live
6:22:57lookup on my ClickUp to understand
6:22:59what's going on. So what I did is I
6:23:01hardcoded in these list IDs because when
6:23:03I was watching it, I realized every
6:23:05single time it was doing this, it was
6:23:07calling the ClickUp MCP and it was
6:23:09gathering all these lists and it was
6:23:11searching and parsing the results and
6:23:12then it would extract the ID and that
6:23:14just was taking so long and it was
6:23:16costing me a ton of tokens. So I
6:23:17realized that's always going to be the
6:23:19same. Why don't I just give it in the
6:23:21skill document the list IDs and now it
6:23:24knows how to do that instantly every
6:23:25time and it doesn't waste all those
6:23:26tokens. And on top of that, I know that
6:23:28searching through ClickUp can consume a
6:23:30lot of time and tokens. So, I built a
6:23:32specialized sub agent that in this
6:23:34skill, I say, "Hey, delegate to the
6:23:35ClickUp searcher agent with this query
6:23:37in order to do all of this searching so
6:23:39that you don't blow your own context
6:23:41window." All of that's handled over
6:23:42there and then you only get the
6:23:43information that you need. So, there's a
6:23:45lot of advanced things that you can do
6:23:46to manage your context. I'm not going to
6:23:48dive into all of that right now. We're
6:23:49just focusing on skills, but just wanted
6:23:50to give you a little taste of what's
6:23:52possible in the skill.md files. So,
6:23:54another good example of needing a
6:23:56reference doc like that is in my skill
6:23:58builder skill. I obviously use this when
6:24:00I'm creating new skills, optimizing
6:24:02skills, auditing skill quality, things
6:24:04like that. And a lot of the inspiration
6:24:05I got from this was of course straight
6:24:07from claude code docs itself about how
6:24:10to actually use and build and optimize
6:24:12skills. And so, when I was building this
6:24:14out, I I was watching the agent, you
6:24:16know, run the skill and I realized it's
6:24:17searching every single time. It's doing
6:24:19a web search and it's crawling the
6:24:21entire document, even if I just need a
6:24:23little piece of information. So, what I
6:24:24decided to do was I told it to basically
6:24:26scrape that whole thing and then I gave
6:24:28it a reference.md, which is basically
6:24:30the documentation. So, I've got my
6:24:32skill.md and what it does is it
6:24:34references that full file if it needs
6:24:36it. But really, the main idea that I'm
6:24:37trying to drive home here is that
6:24:39processing markdown files for your agent
6:24:41is so much quicker and cheaper than
6:24:44actually making API calls or HTTP
6:24:46requests, you know, executing functions
6:24:48and reading tons and tons of tokens. So,
6:24:50the goal is your skills will get to a
6:24:52place where you can invoke them, focus
6:24:54on something else for 10, 15 minutes or
6:24:55whatever, and then come back and have a
6:24:57finished result that is really, really
6:24:58good. But the first couple times that
6:25:00you are testing out a skill, I think
6:25:02it's a really good idea to just sit
6:25:03there and watch it and see what it's
6:25:04doing. And a lot of people have asked me
6:25:06like when do you know when to build a
6:25:07skill? Well, basically just go about
6:25:10your work and if you ever realize that
6:25:12you've done something already or you've
6:25:13instructed something differently, like I
6:25:15tell my claude to not use m dashes.
6:25:18Okay, well that's probably a good idea
6:25:19to put that in the prompt, right? So if
6:25:21you ever find yourself doing a process
6:25:22or repeating prompts, then that's
6:25:24probably a good use case to build a
6:25:26skill around it because skills don't
6:25:27have to be complex. They could literally
6:25:29just be a 50line markdown file. All
6:25:32right, so we're about to hop into a live
6:25:33build of a skill from scratch. But what
6:25:35I wanted to do real quick was go over
6:25:37the six-step skill building framework.
6:25:39So number one is the name and the
6:25:40trigger. What is it called and the
6:25:42natural language that would basically
6:25:44fire it off? Number two is the goal. So
6:25:47in one sentence, what will this skill
6:25:48accomplish by the end? What will be the
6:25:50output? Number three is the actual meat
6:25:52of it. That's the step-by-step process.
6:25:54If you had to do something manually,
6:25:56exactly what do you do in what order?
6:25:58What do you look at? And what decisions
6:25:59do you make? Number four is the
6:26:01reference files. What context do you
6:26:03need? Do you need images? Do you need
6:26:04understanding of current projects,
6:26:06current priorities? Do you need style
6:26:07guides? What do you need to do the job
6:26:10well? Number five is the rules. Think
6:26:12about what could go wrong and then the
6:26:13agent can help you building guardrails
6:26:14and constraints around that. And then
6:26:16number six is kind of like after you've
6:26:17built it, it's just the self-improvement
6:26:19loop. And after the live build, I'm
6:26:21going to talk about actually testing and
6:26:22iterating and what you need to do to
6:26:24make them really, really good. But for
6:26:26now, that's the sixstep skill building
6:26:27framework. Let's hop into a live build.
6:26:29Okay, so here we are in Visual Studio
6:26:31Code, which is where I like to use Cloud
6:26:33Code. If you don't have Visual Studio
6:26:34Code, just go ahead to a browser, type
6:26:36in VS Code, and then go ahead and
6:26:38download this. This is what it will look
6:26:39like. If it's your first time using
6:26:40Cloud Code in here, you just have to go
6:26:42to extensions on this lefth hand side,
6:26:44type in Cloud Code, and then install
6:26:45this, and then log in with your paid
6:26:48Anthropic subscription. Now, after that,
6:26:50you're going to click on this top left
6:26:51button, and it's going to pull up this
6:26:52little thing that says you have not yet
6:26:54opened a folder. What you need to do is
6:26:55open up a project to work in. So, you
6:26:57could either open up one that you're
6:26:58already working on or you could go ahead
6:27:00and create a new folder and then open
6:27:01that one up. For the sake of the demo, I
6:27:03just opened up a new blank folder called
6:27:05a bunch of skills. And I'm going to show
6:27:06you exactly what to do. So, the first
6:27:08step is to go to my free school
6:27:09community link in the description. Go to
6:27:10the agent skills classroom and download
6:27:13the skill builder folder. Once you've
6:27:15got those files ready to go, first thing
6:27:17we want to do is just set up this
6:27:18workspace real quick. Initialize this
6:27:20project with a simple.cloud/skills
6:27:24structure. Cool. So, as you can see,
6:27:26that got set up. We have aclaude. We
6:27:28have a skills folder. And what I'm going
6:27:30to do is in this skill folder, I'm going
6:27:32to create a new folder called
6:27:34skill-builder
6:27:36and hit enter. And then I'm going to
6:27:38take those two files for my school
6:27:39community, the reference and the
6:27:41markdown. And I'm going to put that
6:27:42right in here. So now we have this skill
6:27:44builder set up with the reference file
6:27:47and the actual skill markdown. I'm
6:27:49asking it if it can see that new skill
6:27:50that I just added. It says yes, I can
6:27:52see it. And I'm basically just going to
6:27:53say, cool. Let's run that skill to build
6:27:55a new one together. So now you can see
6:27:56what it did is it basically is reading
6:27:58the skill right now. This is the
6:28:00instructions that we saw right in here.
6:28:02As you guys know, since that's how
6:28:04skills work, it starts to read this. So
6:28:06here we go. I built this skill to
6:28:07actually ask you questions so that it's
6:28:09way easier for you to communicate what
6:28:10you want. So the first thing is what
6:28:12problem are you trying to solve? What we
6:28:13want to do is content creation because
6:28:15in this skill, what I want to do is
6:28:16building branded infographics. What kind
6:28:19of content does the skill create? What's
6:28:20the specific use case or workflow? And
6:28:22I'm actually just going to choose other
6:28:23for this. And I'm going to say
6:28:24educational infographics. Now it's
6:28:26asking how we should trigger this skill.
6:28:28So does it want to be natural language
6:28:29or do we want to just use slash
6:28:31commands? And I'm just going to say both
6:28:32is fine. And now we're moving on to the
6:28:34step-by-step process, which is really
6:28:36important because at this point we
6:28:37haven't told it what text stack we
6:28:39actually want to use or anything else
6:28:41about our business. So walk me through
6:28:42what should happen from trigger to
6:28:44output. And it has some good guesses.
6:28:45But what I'm going to do is do other and
6:28:47explain this the way that I want it
6:28:48built. I will tell you what I want an
6:28:50infographic about. You will create a
6:28:53concept. You will make a request to
6:28:56key.ai to use nanobanana to generate the
6:28:59outline or sorry to generate the image.
6:29:01And you will also look at the brand
6:29:03guidelines that I give you so that
6:29:04everything that is created follows my
6:29:07brand colors and typography and stuff
6:29:09like that. The output format that I
6:29:11actually want is a PNG, not any of this
6:29:13stuff. Does this need to be
6:29:14conversational or fire and forget? I'm
6:29:16just going to go fire and forget. All
6:29:18right. All right. So, how does the key
6:29:19AI nanobanana integration work? Is it an
6:29:21API call? Yep, we're just going to go
6:29:22with an API call. And in these options,
6:29:24you could literally say, I don't know.
6:29:26Let's try different things. You know,
6:29:27help me figure out what's best. It's
6:29:28asking where those brand assets live, so
6:29:30I'll put them in a folder. And where
6:29:32should the generated PNG infographic be
6:29:34saved? Yeah, sure. Let's start a new
6:29:35folder called projects, and we'll throw
6:29:36all of them there. So, it's going to
6:29:38keep our project organized as well. So,
6:29:40now it's asking about brand guidelines.
6:29:42I created this folder, and I put in our
6:29:44kind of color scheme as well as the
6:29:46actual AIS logo. I have put in both our
6:29:49AIS brand guidelines and the AIS logo. I
6:29:52want to make sure that in the top left
6:29:53corner of every single infographic
6:29:55that's created, the AIS logo appears
6:29:57exactly as I've given you. But I think
6:29:59you guys get the point here. I'm going
6:30:00to answer a few more questions and then
6:30:01I'll just show you when we have a
6:30:03result. And now that we've done that,
6:30:05what you can see is it is going to
6:30:06create the skill. It's going to create
6:30:08the logo overlay. It's going to create a
6:30:10supporting reference markdown file for
6:30:12all of the API details that it's going
6:30:13to need. So that's great. It's going to
6:30:15register the skill in claw.md and it's
6:30:17going to log its decisions. All right,
6:30:19so it fully built the skill. It created
6:30:21all those files for us. We just have to
6:30:23give it a key API key so it can actually
6:30:26run this. Okay, so I threw in my API key
6:30:28and then I said test it out with an
6:30:29infographic about cloud skills. That's
6:30:31it. No other context. It invoked the
6:30:33skill right here and we will see what
6:30:35happens. Okay, this is really
6:30:37interesting. So what it's doing is it is
6:30:38generating the image and then it's just
6:30:40going to overlay the logo. So, it's
6:30:42gonna be a lot more consistent than
6:30:43giving the AI image generator nano
6:30:45banana my logo. So, I didn't even tell
6:30:48it to do that. Let's see how it looks.
6:30:49Okay. Well, I don't love this. We're
6:30:51just going to go back and ask it to
6:30:53change some things. The logo on the top
6:30:55left doesn't look great. I gave you a
6:30:57logo with a transparent background, so
6:30:58it should just be overlaid on top and we
6:31:00should be able to see the background
6:31:02behind it. The actual infographic itself
6:31:04is all right, but I actually want these
6:31:06to always be one by one aspect ratio.
6:31:09Okay, so I made some suggestions and
6:31:11it's going to try again and it's going
6:31:12to update its skill. So we'll see if
6:31:14that's better. All right, so second time
6:31:15we run the skill. Let's see if it's any
6:31:17better. All right, there we go. We've
6:31:18got the logo up top. We've got cloud
6:31:20code skills, custom AI workflow, command
6:31:22prompt, trigger, front matter, config
6:31:24triggers, AI agent delegation, document
6:31:26output. So just keep in mind all we said
6:31:28was build an infographic about cloud
6:31:30skills. And this was run number two.
6:31:32Every single time that we do this cycle,
6:31:34remember we talked about the feedback.
6:31:36We would basically watch it again, give
6:31:38more feedback, and then keep going. And
6:31:40after we run this probably five or six
6:31:41more times, this would be really, really
6:31:43good. And then every time I ask for an
6:31:44infographic, it's going to be
6:31:46consistent. And just to show you guys
6:31:47what was actually built, if we open up
6:31:49the infographic builder skill, we have
6:31:51the actual skill itself. So we have the
6:31:53front matter right here, the name, the
6:31:55description, we've got what the skill
6:31:57does, we've got context, so here's where
6:31:59it links to the actual brand guidelines
6:32:01and logos. We've got the step-by-step
6:32:02workflow right here. And we can see
6:32:04right here for full API reference and
6:32:06parameters just see the markdown file so
6:32:08that you don't have to actually go
6:32:09search the web and search through a
6:32:11bunch of tokens. You can just read this
6:32:12markdown file. All right. So we've
6:32:14talked about a lot of stuff about skills
6:32:15today and we just built one live. So
6:32:17what I want to talk about now is really
6:32:19how do you bridge the gap from like a
6:32:2090% good skill to making it pretty much
6:32:23100%. So testing, iterating, and
6:32:26debugging. There's different symptoms
6:32:27and there's different fixes. So let's
6:32:28just kind of go down this list one by
6:32:30one. The first symptom might be it does
6:32:32the wrong steps or in the wrong order.
6:32:33Well, you would just tell it to edit the
6:32:35skill.md instructions. You could get
6:32:37missing tone, style, or context. In that
6:32:39case, you're going to add reference
6:32:41files. And of course, those have to be
6:32:43pointed to correctly in the skill.md.
6:32:45You could get the same mistake happening
6:32:46over and over, then you're going to add
6:32:48a rule. If it struggles with a tool or
6:32:50an MCP or it keeps searching for the
6:32:51same things, then create some sort of
6:32:53reference dock for it. If it works good,
6:32:55but it could get better, then that just
6:32:57means you have to brute force it. you
6:32:58have to just run it over and over and
6:33:00over and keep nitpicking at what it does
6:33:02wrong or maybe not wrong but what it
6:33:03could improve on. If the skill isn't
6:33:05triggering, then check the YAML and make
6:33:06sure it is specific enough. If the skill
6:33:08triggers too often, then maybe try
6:33:10disabling model invocation and that is
6:33:12something that you can see in the claw
6:33:14docs, which basically gives you control
6:33:15over if the skill can only be invoked by
6:33:17natural language or only be invoked by
6:33:18the slash command directly or both. So,
6:33:21like I said, if you want to look at some
6:33:22more advanced stuff, then definitely
6:33:23head over here to the actual doc. But at
6:33:25this point, we've covered almost
6:33:26everything about these skills. One thing
6:33:28that I would call your attention to is
6:33:30the actual front matter reference
6:33:31because we saw the name and the
6:33:32description which is what's required
6:33:34every time. But there's lots of other
6:33:35things that you can add in there. Here
6:33:37is the disable model invocation like we
6:33:39just saw. But you can also give it
6:33:41allowed tools. You can also give it an
6:33:42argument hint. You can give it a
6:33:44specific model to use. You can give it
6:33:46specific context. You can give it hooks.
6:33:48You can give it a specific agent. And so
6:33:50all of this lets you get really really
6:33:51granular on the exact skill and how you
6:33:53want it to be used. But don't get
6:33:55overwhelmed. You really only get to that
6:33:56point once you've ran the skill a ton of
6:33:58times. Now, another thing that I need to
6:33:59hit on real quick is where do skills
6:34:01actually live because what we've seen so
6:34:03far is just building them right in
6:34:05ourcloud/skills folder. But when you're
6:34:07doing this, they only exist in that
6:34:09specific project. So whether that's my
6:34:11her two or my, you know, the one we just
6:34:13spun up, if I went to a different
6:34:15folder, that skill would no longer be
6:34:17able to be accessed by our cloud code.
6:34:19But you can also create skills that are
6:34:20actually global. And you do that by
6:34:22doing that in a different directory in
6:34:24your kind of overall home directory. And
6:34:26that's basically indicated by the little
6:34:28tilda right here. And so that means
6:34:30every project you use in cloud code, no
6:34:31matter where you are, that skill would
6:34:33exist. So for example, I have a
6:34:35front-end design skill that is installed
6:34:37globally. So that whenever I'm anywhere,
6:34:39if I need to do front-end design, it
6:34:41just is able to use it. And just in case
6:34:43you want to look at it in a different
6:34:44way, right now what we're doing is we
6:34:46have our projects, right? So, herk 2 and
6:34:49then we have doclaude and then within
6:34:51dotcloud we have skills and then your
6:34:53skill and then your MD your references
6:34:55whatever and then maybe another skill.
6:34:56But if it was global you might not
6:34:58actually see it in your project. It
6:35:00would just be within your overall home
6:35:02directory. So the reason why you might
6:35:03want to do this is if there's something
6:35:04very specific about you, your business,
6:35:06your workflows that you want applied to
6:35:08every single project no matter what.
6:35:10Maybe your company context, your company
6:35:12projects, your tone of voice, whatever,
6:35:13then you can install that globally
6:35:15instead. If you guys love nerding out
6:35:17about this kind of stuff, then
6:35:18definitely check out my paid community.
6:35:19The link for that is also down in the
6:35:21description. We've got a great
6:35:22community, over 3,000 members in here
6:35:23who are building with AI every day and
6:35:25building businesses with AI. So, I'd
6:35:27love to see you guys in this community.
6:35:31Cloud skills just got 10 times easier to
6:35:33build and stronger to use. So, in
6:35:35today's video, I'm going to explain
6:35:36exactly why that is, and then I'm going
6:35:37to live build a completely new skill
6:35:39right here in front of you guys. So,
6:35:40real quick, what is a skill? It's
6:35:42basically just a recipe. So that when
6:35:44you ask your agent to make you, for
6:35:45example, a LinkedIn post, it will read
6:35:47the recipe and it will get it right
6:35:48every single time. And when I say
6:35:50recipes, I literally just mean text.
6:35:52It's just text instructions. It's like a
6:35:53prompt. So if I go to customize and I go
6:35:55to skills and I click on, let's say, for
6:35:57example, the internal comm skill, this
6:35:59says a set of resources to help me write
6:36:01all kinds of internal communication
6:36:02using the format that my company likes
6:36:04to use. And you can see this is the
6:36:06skill itself. It is literally just text
6:36:08that you could read, that an intern
6:36:10could read, anybody could read and
6:36:11understand what's going on in the skill.
6:36:13And if you're using them in cloud code,
6:36:14you can see I've got a ton of skills
6:36:16here. So, for example, let's look at my
6:36:18um idea mining skill. This is the
6:36:20markdown file that explains to the agent
6:36:21what this skill actually does. And once
6:36:23again, it's all just text. So, what did
6:36:25Enthropic actually do that made all
6:36:27these skills better? they updated their
6:36:29skill creator skill which is literally a
6:36:31skill that teaches Claude how to build,
6:36:34test, measure, refine, just make all the
6:36:36skills better and better and better. So,
6:36:38let's actually cover why that matters
6:36:39and what happened. So, the first thing I
6:36:41need you to understand is that there are
6:36:42two different types of skills. We have a
6:36:44capability uplift skill, which basically
6:36:47is a prompt. So, it teaches Claude how
6:36:49to do something better. for example,
6:36:51design websites with the front-end
6:36:52design skill or create documents or run
6:36:55Excel formulas. Things that maybe the
6:36:57default model by itself doesn't know
6:36:59super well, but with a prompt, it does a
6:37:01much better job. And then we also have
6:37:03encoded preference skills, which means
6:37:05that Claude already understands each of
6:37:07these pieces, but it needs to follow
6:37:09them in a specific order. So these are
6:37:11way more like actual workflows, like
6:37:12actual kind of like step-by-step
6:37:14automations. So, quick example. If you
6:37:16ask Claude without a front-end design
6:37:18skill to build you a website, it could
6:37:20do it, but it might just look very
6:37:21generic. It might look AI slop as they
6:37:24call it. But if you give it the exact
6:37:25same prompt, but this time you also let
6:37:27it use the front-end design skill, it's
6:37:29going to look much better because that
6:37:30skill tells it stuff like good fonts,
6:37:32good color schemes, you know, good
6:37:34background elements, good layouts. And
6:37:36that is a classic capability uplift
6:37:38skill. Now, here's an example of an
6:37:39encoded preference skill, which is the
6:37:41one we just saw in My Cloud Code, which
6:37:42I call idea mining. And this skill is a
6:37:45little bit more sequential and there's
6:37:46different steps involved. So, first it
6:37:48will look at my YouTube comments. It
6:37:49will look at, you know, some videos in
6:37:50my niche. It will also look at AI trends
6:37:53on X and the web. It will then spin up
6:37:55two different agents. So, a YouTube
6:37:57agent that analyzes this stuff and a
6:37:58research agent that analyzes this stuff.
6:38:00And these run in parallel. And then they
6:38:02both send their output back to the main
6:38:03agent, which will score and cross
6:38:05reference. And then the main agent turns
6:38:06all that information into some video
6:38:08ideas for me, which is why I call it
6:38:10idea mining. So, what I could do is I
6:38:12could say, "Hey, Mr. AI agent, go look
6:38:14at my comments, go look at YouTube, go
6:38:15look at X, you know, analyze that and
6:38:17help me find some video ideas and every
6:38:19time it would give me different answers
6:38:20and every time it would sort of do it
6:38:22differently or I can just say, "Hey, do
6:38:24some idea mining." And it will just call
6:38:25the skill and every time I get an output
6:38:27that I like. And the reason why this is
6:38:28actually important to understand is
6:38:30because capability uplift skills might
6:38:33fade over time because for example with
6:38:35the front-end design skill, right now
6:38:37we're with Opus 4.6, right? What if Opus
6:38:405 drops and default Opus 5 is better at
6:38:43front-end design than Opus 5 with a
6:38:45front-end skill? So, at that point, you
6:38:47might just need to retire that skill
6:38:48completely. But with an encoded
6:38:49preference skill, these will probably
6:38:51stay pretty durable and accurate because
6:38:53the process is very specific usually to
6:38:55you, which Opus 5 won't be trained on
6:38:58most likely. Okay, so those are the two
6:39:00kind of different types of skills. Now,
6:39:02we can actually evaluate them. So, with
6:39:04this new skill creator skill, which is
6:39:05an official anthropic skill, this is the
6:39:07one we're talking about. It's in the
6:39:09repo right here. And if I open up the
6:39:10actual skill MD, you can see this is
6:39:12what it does. It creates new skills. It
6:39:14can modify and improve existing skills.
6:39:15It can measure skill performance. So use
6:39:18this when you want to create a skill
6:39:19from scratch, if you want to update or
6:39:20optimize one. If you want to run evals
6:39:22to test a skill, if you want to do
6:39:23benchmarks, or if you want to optimize a
6:39:25skills description for better trigger
6:39:27accuracy. So I'm going to talk about
6:39:28what each of these little elements mean,
6:39:29but I just wanted to show you that this
6:39:31is the actual skill creator skill. It's
6:39:33basically just all of Enthropic's best
6:39:35practices on how to build better skills.
6:39:38They've done things before like dropped
6:39:40a 33page PDF which walks you through
6:39:42fundamentals, planning and design,
6:39:44testing and iteration, distribution and
6:39:46sharing, all this kind of stuff,
6:39:47patterns and troubleshooting. This is
6:39:48pretty thorough. So you could either
6:39:50take time and learn this or you could
6:39:52just give your agent the skill creator
6:39:53skill and all that information is
6:39:55already in there. So what the eval do is
6:39:57it lets your agent actually evaluate the
6:40:00quality of your skill and then make
6:40:01improvements. So, let's say you have a
6:40:03skill for creating job descriptions.
6:40:05What you could do is give your agent
6:40:07tons of examples of really good job
6:40:08descriptions that you want. And then it
6:40:10will look at your skill. It will test
6:40:12out some prompts and it will compare it
6:40:13to the outputs and it will be able to
6:40:14optimize your skill for you. As we've
6:40:16talked about in the past, the more you
6:40:17use a skill, the better and better
6:40:19because you're able to give feedback on
6:40:20what you like and what you don't. So,
6:40:21this basically shortcuts that process.
6:40:23Here's a quick example that Enthropic
6:40:25actually ran with this eval. The skill
6:40:27for filling out some PDF stuff was
6:40:29having trouble finding the right spot to
6:40:30put the text. But then after they ran
6:40:32the evaluation on the skill and it was
6:40:34able to improve, now you can see all the
6:40:35text is accurately being placed whether
6:40:37that is a checkbox or just a fill in
6:40:39some sort of field. So there's two
6:40:41reasons that we need to use evals and
6:40:43they sound kind of similar but they're
6:40:45basically the opposite. So the first one
6:40:46is to catch regressions. So this means
6:40:48let's say we have a job description
6:40:50skill. As a model evolves it might
6:40:53actually use the skill worse because
6:40:55it's trained a little different and it
6:40:57you know thinks a little different. So
6:40:58this would basically be an early signal
6:40:59that you need to evolve your skill. And
6:41:02then the second one is to spot out
6:41:03growth. So once again, as models improve
6:41:06or evolve, it might be able to just do a
6:41:08better job without a skill at all. And
6:41:10that's when you would be able to run the
6:41:11evaluation, say, okay, wow, without a
6:41:13skill, it's actually better. And I'm
6:41:14just going to go ahead and delete this
6:41:15or maybe just archive it. And then we
6:41:17can also run benchmarks. So when a model
6:41:19updates or when you make an iteration
6:41:21and you change your skill, just run all
6:41:23the evals and run a benchmark which will
6:41:25give you stuff like a pass rate, a time
6:41:27and also how many tokens are being used.
6:41:29So here's an example where they said
6:41:31benchmark the PDF skill with and without
6:41:33the skill loaded and show me sideby-side
6:41:34results so I can see the uplift. And we
6:41:37get all this information about these
6:41:38different evaluation metrics. We get the
6:41:40pass rate, we get the total time and the
6:41:42total tokens. So here you can clearly
6:41:44see that with the skill you're getting
6:41:45much better results. And then the final
6:41:47piece is skill trigger tuning. So once
6:41:50you've got a project filled up with,
6:41:51let's just say 10 or more skills, you
6:41:53might notice sometimes that you get
6:41:55false triggers or you get misfires,
6:41:57meaning you wanted it to use a skill and
6:41:59it used the wrong one or you wanted it
6:42:01to use a skill and it just didn't use
6:42:03any at all. Luckily, you could also use
6:42:05them with slash commands, but it's so
6:42:06much more convenient to just be able to
6:42:07speak a natural language and make sure
6:42:09that your agent understands you. So
6:42:11using the trigger tuning, the skill
6:42:12creator will basically analyze your
6:42:14skill. It will test out different
6:42:16prompts that you might use to trigger
6:42:17that skill and then it will edit the
6:42:18description so that that skill gets
6:42:20called more accurately. And this is an
6:42:22actual evaluation that they ran. You can
6:42:24see on the lefth hand side and on the
6:42:25right hand side we have the test score
6:42:28and the train score. And the green and
6:42:30blue are basically the results after it
6:42:32has been analyzed and fixed with the
6:42:35trigger tuning. So you can see it's
6:42:36still not perfect, but it's so much
6:42:38better than where we were without this
6:42:39new skill. And what I think is really
6:42:41cool and how I want to end off this
6:42:42section before we get into a live demo
6:42:44is where this is going. And at the
6:42:46bottom, we have a quote from Enthropic
6:42:47themselves that say, "Over time, a
6:42:49natural language description of what the
6:42:50skill should do may be enough with the
6:42:52model figuring out the rest." And I
6:42:54really think that this word may should
6:42:56actually have been will. And basically
6:42:57what this means is that today when we're
6:42:59telling our agent to build skills for us
6:43:01or maybe just giving it an SOP, we're
6:43:03giving it steps, rules, and format. But
6:43:05what's going to happen in the future is
6:43:06we're going to be able to just tell it
6:43:08in way more highle natural language what
6:43:10we want and it's going to be able to
6:43:11figure out all of that and get there
6:43:12with a spec and basically just cut down
6:43:14the time that it takes for us to get a
6:43:16really good skill or you know a really
6:43:18good automation. All right, so I am in
6:43:20my Herk 2 project which is kind of just
6:43:21like my personal assistant in cloud code
6:43:24and I'm going to show you guys how we
6:43:25can actually get the skill installed. So
6:43:27whether you're in VS Code, which is
6:43:28where I am, or in the terminal or
6:43:30desktop app, whatever, you just need to
6:43:32do /plugins, you can click on manage
6:43:34plugins, and then if you just go in here
6:43:36and you can see like all of the kind of
6:43:38anthropic official ones, you can just go
6:43:40ahead and search for skill-creator.
6:43:43And right here, you can see the official
6:43:45one. Here's the GitHub. And all you have
6:43:46to do is go ahead and click install. You
6:43:48can install this for just you, you can
6:43:49install it for your project, or you can
6:43:51install it locally. And I'm just going
6:43:52to install it for the whole project.
6:43:58So now you can see that's installed and
6:43:59I'm just going to go ahead and restart
6:44:00Claude Code so that that actually
6:44:02happens. And so just keep in mind if
6:44:04you're in Cloud Code, it may not show up
6:44:06right here in your actualcloud skills if
6:44:08you did it, you know, in your project.
6:44:10So you can just verify it and say, do
6:44:11you have the skill creator skill? What
6:44:12does it do? And you can see right here
6:44:14that we do in fact have that. So I'm
6:44:15going to go ahead and switch on to plan
6:44:16mode and I'm going to see if it can
6:44:18build us a new skill. I need you to
6:44:19create a skill called YouTube weekly
6:44:21roundup where at the end of every week,
6:44:23you will look at the videos that I made
6:44:25that week. You'll analyze the comments,
6:44:26you'll analyze the views, engagement,
6:44:28things like that. And you'll give me a
6:44:31PDF report on all of the insights,
6:44:34strengths, weaknesses, threats,
6:44:36opportunities. So, that's all I'm going
6:44:37to send off. And I kept this pretty
6:44:39vague intentionally to see what it's
6:44:41going to come back with and how it's
6:44:43going to be able to plan this out for
6:44:44us. And this is where the future's
6:44:45going. And this is what Enthropic is
6:44:46talking about because most people that
6:44:48are using skills right now are actual
6:44:50just like executives and managers and
6:44:52operators. They're not engineers, which
6:44:54means we're really good at being able to
6:44:55explain what we want, the metrics we
6:44:57need to hit, and why we need that, but
6:44:59maybe not all of those technical
6:45:00nitty-gritty details. All right, so it
6:45:02came back and asked me some questions.
6:45:04The first thing I said is I want it to
6:45:05just be the last 7 days. So, it's a
6:45:06rolling 7-day window. It asked about the
6:45:08report sections that it came up with,
6:45:10and I said those look good. And for the
6:45:11PDF style, I told it to use the brand
6:45:13assets in my folder. So right over here
6:45:15I've got my brand guidelines and then
6:45:16this one is the actual logo for AIS. So
6:45:18I'm telling it to use those and
6:45:20hopefully it can throw all that on there
6:45:21and make it feel really branded. So it's
6:45:23going to keep going now with this plan.
6:45:25All right. So at this point it came back
6:45:26with a plan. And keep in mind I still
6:45:27haven't told it anything about text
6:45:29stack or anything else. It's writing out
6:45:31everything that it's going to do. And
6:45:32normally I would read through this and
6:45:33give it some tweaks potentially but I
6:45:35just want to see what this skill creator
6:45:37is able to do with a oneshot prompt. And
6:45:38I'm just going to go ahead and accept.
6:45:40And look at this. In its to-do list we
6:45:42can see that it creates all these
6:45:43things. But then the last step is to run
6:45:44the test and iterate with the skill
6:45:46creator eval process. So I'm excited to
6:45:48see what it does there. So you can see
6:45:49that it created everything and then what
6:45:51it did is it decided to test it to do a
6:45:53final iteration. Okay. So I was a little
6:45:55confused. I said, "Do you have an actual
6:45:56PDF file for me?" And it said, "Yes, it
6:45:58is in your projects folder." I was
6:46:00looking in the templates where it
6:46:01created an HTML template, but apparently
6:46:03it actually rendered that as a PDF. So
6:46:05let me go to projects. We'll go to
6:46:07YouTube weekly roundup. And right here
6:46:09we have an actual PDF, which this
6:46:11doesn't look great. Obviously, this is
6:46:13not a PDF, but if I actually open it up
6:46:15from my files, it is a PDF. So, here we
6:46:18have the logo, we have weekly roundup,
6:46:20we have three videos published, and then
6:46:21we got some stats on views, likes, and
6:46:23comments. I'm going to keep going down.
6:46:25We have our executive summary. So, this
6:46:27is for it actually ran I think two weeks
6:46:29worth of data just to test this out. And
6:46:31I will say just by glancing at this, I
6:46:32don't think that this data is correct.
6:46:34So, keep that in mind. Here we can see
6:46:36the per video breakdown. Right now, we
6:46:38have nothing available in our SWAT
6:46:39analysis. And then we have competitor
6:46:40context and there's nothing available
6:46:42here. So now it's time to give it some
6:46:44feedback and see what it can do. I'm
6:46:45first of all going to clear out this
6:46:46context because it used up 62%. I'm
6:46:49going to go back into plan mode and just
6:46:50give it some honest feedback. All right,
6:46:51so the report looks great. Like
6:46:54aesthetically, you did a good job on the
6:46:56design. However, the data is all wrong.
6:46:58There was a lot of missing elements. I
6:47:00need you to really look at how you're
6:47:03actually scraping this data from my
6:47:04YouTube channel, how you're actually
6:47:05searching through the comments and
6:47:07competitor videos and make sure that
6:47:09there's actually data going into this
6:47:11report. And before I send this off, it's
6:47:12interesting because you can see here it
6:47:14sent us some JSON data, which is
6:47:15actually the raw information that it was
6:47:17able to find from my YouTube channel.
6:47:18And the thing is, this isn't super
6:47:20in-depth. So, I just don't think that it
6:47:22did a good enough job on the research
6:47:23element. And maybe this is exactly what
6:47:25we were talking about earlier over here
6:47:26where at some point the AI is going to
6:47:28be able to understand that we want all
6:47:30of this granular data, but maybe right
6:47:32now it's our job to just explain that
6:47:33really clearly. I want to see comments
6:47:35analysis. I want to see what's working
6:47:37for other people in the space. I want to
6:47:39see, you know, other trending videos in
6:47:41AI. And I want you to use all of that
6:47:43and use your brain to figure out what
6:47:44are the strengths my channel has, the
6:47:46weaknesses, and the opportunities and
6:47:47the threats. And then all of this
6:47:49information should be a pretty in-depth
6:47:51research report for me on, you know, my
6:47:53YouTube weekly roundup. So, while this
6:47:55is running, I thought that we should
6:47:55real quick look at what it actually did.
6:47:57So, in my claude, we've got my skills
6:47:59folder. If I go all the way down, we've
6:48:00got the YouTube weekly roundup. And this
6:48:02is the MD file. So, we've got the YAML
6:48:04up top with the name, description,
6:48:06disable, model invocation false, which
6:48:08just basically means that Cloud Code can
6:48:10call this based on a request. It doesn't
6:48:12have to be explicitly a slash command.
6:48:14And then an argument hint. So basically
6:48:17when cloud code decides to use this
6:48:19skill, it will send in maybe a hint so
6:48:22that the skill understands like what
6:48:23video we're looking at or you know the
6:48:25topic. It's giving some context. It's
6:48:27giving some channel benchmarks, some
6:48:28optional focus and then step-by-step
6:48:30instructions on what to actually do
6:48:32here. Now you can see what it's doing is
6:48:33it's calling on a script called fetch
6:48:36YouTube data, which if I was to look for
6:48:38that in here, I could probably go down
6:48:40to my scripts. I could see YouTube
6:48:42weekly roundup. And right here we've got
6:48:44some different things. We've got the
6:48:46prepare data. We've got the render
6:48:47report. And we also have a script that I
6:48:49already had in this project that it was
6:48:50able to find and use, so it didn't have
6:48:52to create a new one. And this one is
6:48:53called fetch YouTube data. So the
6:48:55skill.md file here basically points to
6:48:57everything that the agent needs in order
6:48:59to do this accurately. Okay, so it's
6:49:01come back with another detailed plan,
6:49:02and I'm going to go ahead and fire this
6:49:04off. And I love this. Once again, we've
6:49:05got all these to-dos, and then at the
6:49:07end, it says to audit with the skill
6:49:08creator. And this is such a good example
6:49:10of why using a project more and using a
6:49:13skill more makes it stronger because
6:49:15some of the pieces that I already had in
6:49:16this project, it's able to reuse like my
6:49:18YouTube analyzer agent, like my YouTube
6:49:20data script. And of course, it has all
6:49:22the context about my business and my
6:49:23YouTube channel in here already. So all
6:49:25of those changes have been made and now
6:49:27all that's left to do is actually run
6:49:28the skill. So I just called the skill.
6:49:30You can see that it's reading it right
6:49:32here. And now what [clears throat] it's
6:49:32doing is it's going to refresh channel
6:49:34data. It's going to use three agents in
6:49:35parallel, prepare the report, populate
6:49:37the data, and then render the PDF and
6:49:39show it to me. So, I'll check in with
6:49:40you guys when we get that output. All
6:49:42right, that finished up. We've got some
6:49:43quick hits. Top competitor move, biggest
6:49:44opportunity. Apparently, Jack Roberts is
6:49:46my biggest threat. If you see this,
6:49:49Jack, keep crushing it. Okay, so here is
6:49:51the report. These stats right off the
6:49:52jump look a little bit more accurate. I
6:49:54might want to tell it to make this logo
6:49:55a bit bigger, but it did what we asked.
6:49:58Seven videos published, and like I said,
6:49:59these stats look more accurate. We've
6:50:01got the executive summary here with some
6:50:02key takeaways of doubling down on dollar
6:50:05outcome titles, make a dedicated
6:50:06anti-gravity tutorial, fill the chat GBT
6:50:09to claude migration, watch Jack Roberts
6:50:11closely, and then address VS Code versus
6:50:13anti-gravity confusion. We've got the
6:50:15per video breakdown. So now you can see
6:50:17the actual metrics from all the videos,
6:50:18including the one that I literally just
6:50:20dropped like an hour ago. And so all of
6:50:22these look like they're doing okay. This
6:50:23one might not be doing the best, and
6:50:25similar with this one. But I really like
6:50:26the way that this actually looks. The
6:50:28layout's pretty good. It is very clean
6:50:31and professional. For the SWAT analysis,
6:50:33we actually have it on the second page.
6:50:35It still looks good. That's obviously
6:50:36just an easy spacing issue that we can
6:50:38fix. So, we have some strengths here. We
6:50:40have some weaknesses. We have threats.
6:50:42And we have our opportunities. Top
6:50:44comments and audience signals. Selling
6:50:46shovels in a gold rush, bro. Well
6:50:47played, man. 26 likes. Hi, Nate. 10 days
6:50:50into joining your plus community. I got
6:50:51my first potential client. It's all
6:50:52thanks to you and your community.
6:50:53Awesome. And you can see that we see
6:50:55other comments. We see what video they
6:50:57came from and how many likes. And we're
6:50:58also getting video requests, we're
6:51:00getting pain points. And so that really
6:51:01helps me stay in tune with what you guys
6:51:03are saying. Wow, it just keeps on going.
6:51:05We've got competitor context. So all of
6:51:06these channels, all of these videos, all
6:51:08of these stats, and that comes along
6:51:10with some notable gaps. And then
6:51:11finally, we get what's trending in AI
6:51:13this week. So what are skills, the most
6:51:15powerful AI agent I've ever used, all of
6:51:17this stuff with the channel, the views,
6:51:18the views per day, and the topic. So
6:51:20this is amazing. And I was able to build
6:51:22this in 20 minutes. And now what I would
6:51:24do is just keep running it. And every
6:51:26time say, "Hey, I liked this. I didn't
6:51:27like this. Use the skill creator to make
6:51:29this better and better.
6:51:32So, here I am in Cloud Code and I'm
6:51:34going to go ahead and shoot off this
6:51:35message that's asking Cloud Code to turn
6:51:37my YouTube video into a tweet style
6:51:39carousel. What you see is that it's
6:51:40invoking the carousel skill. And within
6:51:42this carousel skill, I actually tell it
6:51:44to delegate work to a carousel planning
6:51:46sub agent. So, you can see now it said,
6:51:48"I got the transcript. Let me plan the
6:51:49carousel slides." And I'm using an agent
6:51:51to do that. And so this sub agent here
6:51:53basically got the input from this main
6:51:55session and it's running on its own with
6:51:57its own context and maybe a different
6:51:58model whether that be haiku sonnet or
6:52:00opus. And while we let this finish up,
6:52:02I'm going to open up a new session. And
6:52:03in this one, I said spin up two research
6:52:05sub aents. One for AI enablement for
6:52:07SMBs and one for enterprises. And this
6:52:09main session now is going to delegate to
6:52:11two different sub aents. Here's the
6:52:12first one for SMBs and here's the second
6:52:14one for enterprises. These sub agents
6:52:16are now running in parallel doing their
6:52:17own research and then they're going to
6:52:18send all of their findings back to this
6:52:20main session. So in that first session,
6:52:22we now have that carousel done. We've
6:52:24got slide 1, slide 2, 3, 4, 5, 6, 7, and
6:52:278. And then in the second session, we
6:52:29see that both research briefs are back.
6:52:31We've got SMB AI enablement, we've got
6:52:33enterprise, and then we have the
6:52:34overlap. So these outputs that we just
6:52:36got were thanks to using cloud code sub
6:52:38aents. So in today's video, I'm going to
6:52:39explain all of this as simply as I can,
6:52:41and we're going to be looking at what
6:52:42they are, why they matter. I'm going to
6:52:44build one live in front of you guys,
6:52:46talk about how we actually do that, and
6:52:48how you can actually master them and
6:52:49make them better and better. So I don't
6:52:50want to waste any time. Let's get
6:52:51straight into the video. Okay. So what
6:52:53are sub aents? The idea about sub aents
6:52:55is that you have your main session. So
6:52:57as you guys saw in cloud code, this is
6:52:59kind of the agent we're talking to. When
6:53:00you open up a new session, you are
6:53:02talking to this little crab right here.
6:53:04And as you're talking, you fill up your
6:53:06context and it has its own kind of
6:53:08memory right there. But that main
6:53:09session can actually delegate work and
6:53:11send off prompts to different
6:53:13specialized agents. So we could have a
6:53:15code reviewer or a builder or a debugger
6:53:17or a testr runner, an architect or a
6:53:19researcher. And if we wanted, all six of
6:53:21these could be running in parallel.
6:53:23You're essentially turning clawed code
6:53:24into a project lead. And it's able to
6:53:26just do the planning, the delegation,
6:53:28and then not have to have its context
6:53:30polluted with different specialized
6:53:31tasks. And the more specific goal you
6:53:34give it and the more specific training
6:53:35you give it, the outputs are shown to be
6:53:37much much better. Especially if you
6:53:38think about it like this, a sub agent
6:53:40will be asleep, right? and it wakes up
6:53:42and it's completely stateless because
6:53:43basically the main agent said, "Hey,
6:53:45wake up. I need you to do something."
6:53:46This agent now wakes up with fresh
6:53:48context, a different chat model, its own
6:53:50tools, and its own purpose. So, it's
6:53:52like if I was throwing a party and I
6:53:53wanted some really good cupcakes. I
6:53:55probably wouldn't just walk into a
6:53:57general store and just grab, you know,
6:53:58some mass-produced cupcakes, I might
6:54:00want to go to like a niche bakery and
6:54:02give them a custom request. Now, you
6:54:04might have also heard of agent teams,
6:54:06which kind of sounds similar to sub
6:54:08agents, but they're not. And there's one
6:54:09key distinction that I wanted you guys
6:54:11to just understand real quick. Don't
6:54:12worry, I will have a full video about
6:54:14agent teams themselves, but I just
6:54:15wanted to do one quick distinction,
6:54:17which is that sub agents are focused
6:54:18workers. They run in parallel, but they
6:54:20can't talk to each other. And that's the
6:54:22key difference. It's basically one-way
6:54:24relationships. The main agent sends an
6:54:25input prompt and then the worker works
6:54:27and then sends the results back to the
6:54:29main agent. The individual workers
6:54:31basically cannot communicate. But with
6:54:33agent teams, that's where it gets really
6:54:34cool is they actually can. So they share
6:54:36a task list and they can communicate to
6:54:38each other. They can even assign each
6:54:40other tasks. So that's where it truly
6:54:41becomes like an actual team ecosystem.
6:54:44So let's actually go through an example.
6:54:45So let's say I said, "Hey, I need you to
6:54:47refactor and add some tests to X, Y, and
6:54:50Z." What the main session would do is it
6:54:53would basically hear your request and it
6:54:54would search through the agents it has
6:54:56and say which ones can do the job best.
6:54:58I'm going to give one task to the
6:54:59refactor agent. I'm going to give the
6:55:01other task to the test writer agent.
6:55:02They're both going to be working and
6:55:03then once they're done, they're going to
6:55:05bring all of those results back to me,
6:55:07the main agent, and I will do whatever I
6:55:09need to do with it and communicate back
6:55:10to you. Okay, so why use sub agents?
6:55:12I've got five reasons here. And the
6:55:14first one is to preserve context because
6:55:16you can keep the main thread clean. We
6:55:18all know that context rot is real and so
6:55:21your job as a cloud code operator is to
6:55:23figure out how can I be as efficient as
6:55:25possible with my context window. And one
6:55:27of those unlocks is realizing that you
6:55:28can delegate tasks to sub agents to
6:55:30process tons of data, do a ton of
6:55:32research, and then only send back the
6:55:34little bits of information that the main
6:55:35session actually needs. Now, the second
6:55:37piece is to enforce constraints because
6:55:39maybe your main worker, your let's just
6:55:41call it a parent worker needs to be able
6:55:43to do tons and tons of different things
6:55:45in terms of like tool calls. But
6:55:47sometimes when you want to do something
6:55:48a little bit more risky, so maybe GitHub
6:55:50actions or some other code-based
6:55:51actions, you might want to delegate that
6:55:53to a sub agent that is specialized on
6:55:55that tool or that process and it has
6:55:58limited availability when it comes to
6:56:00tools. So it can't do anything super
6:56:01harmful. It just gives you more control
6:56:03over that process. Now the third one is
6:56:05to reuse configs. Meaning all of your
6:56:07agents can be shared across different
6:56:08projects or even shared with your team
6:56:10because maybe your organization has a
6:56:11really specific process for project
6:56:13management or quarterly planning. And if
6:56:15you have a really good agent that helps
6:56:16you do that, then it can be reused
6:56:18across the entire organization. So the
6:56:21whole team gets better and better. And
6:56:22what you'll notice is that there's a lot
6:56:24of principles between sub agents and
6:56:25skills that are very similar except for
6:56:27like the context and model things. Now
6:56:29number four is specialization. Kind of
6:56:31like I alluded to earlier, the more
6:56:33specific and niche an AI is trained to
6:56:36be, the better it results are going to
6:56:37be. And that's as simple as it is. If
6:56:39you've been following my channel for a
6:56:40while and you've been building a lot in
6:56:41Naden, it's the same idea of when we
6:56:43would delegate work to a child agent.
6:56:45And then finally, we have control costs
6:56:48because once again, each sub aent can do
6:56:50their own model. So maybe you're always
6:56:51wanting to use Opus in your parent
6:56:53workflow, but sometimes you want some
6:56:54quick research done and you can do that
6:56:56with Haiku to save yourself time and
6:56:58tokens. So from there, the big question
6:57:00is when do you actually delegate?
6:57:01Because you could also get into this
6:57:03weird spot where you're doing sub agents
6:57:05for everything and you're kind of just
6:57:06overengineering. So stay in your main
6:57:08conversation if you need back and forth,
6:57:10if you need quick changes, if you need
6:57:12that shared context, that conversation
6:57:13history, and for low latency. And then
6:57:16you want to be able to delegate to a sub
6:57:17agent if you want them to be
6:57:18self-contained, if you have tool
6:57:20restrictions, or you just want summaries
6:57:21returned. Now, I know I put verbos
6:57:23output here, but that was just the image
6:57:25generation. Basically, you can control
6:57:27if you want to see the sub agents steps
6:57:28as they're thinking and doing things, or
6:57:30if you just want the concise summary
6:57:32done at the end. So, I thought that I
6:57:33would share this tweet with you guys
6:57:34real quick. Boris Churnney, the guy who
6:57:36created Claude Code, shared that he does
6:57:38use a few sub agents very regularly.
6:57:40Build validator, code architect, code
6:57:42simplifier, on call guide, verify app.
6:57:44And also, if you go to the cloud code
6:57:46docs and you read about sub agents, you
6:57:47can see that there are some examples
6:57:49here which basically give you the full
6:57:50actual like MD of this type of agent.
6:57:53So, this is the code reviewer. And if
6:57:55this doesn't make sense at all, don't
6:57:56worry. I'm going to explain what we're
6:57:57looking at here. We've got a debugger,
6:57:59we've got a data scientist, and we've
6:58:01got a database query validator. So those
6:58:03are just some example sub aents you
6:58:05might be thinking about and seeing just
6:58:07to kind of get the you know blood
6:58:08flowing up to the brain. So here's where
6:58:10your understanding of skills is going to
6:58:12help out a little bit because the agents
6:58:14are basically invoked the same exact way
6:58:16that skills are invoked which means that
6:58:17the main agent will read through the
6:58:19descriptions of the agents. So if I said
6:58:22hey can you help me write tests for my
6:58:23API? The agent would look through all of
6:58:25its different agents. It would see I've
6:58:26got a refactor, I've got deploy, I've
6:58:28got a test writer and a debug and I'm
6:58:30going to use the test writer here. And
6:58:31then it would basically send off that
6:58:33prompt to the test writer agent. So the
6:58:35clearer that your descriptions are, the
6:58:37better that the delegation is going to
6:58:38be so that you don't get any false
6:58:40triggers and things like that. And when
6:58:41I say description, I literally just mean
6:58:43the actual description. So for example,
6:58:44the code reviewer agent expert code
6:58:46review specialist proactively reviews
6:58:48code for quality, security, and
6:58:51maintainability. Use immediately after
6:58:53writing or modifying code. So, like you
6:58:54guys just saw, we have the YAML metadata
6:58:56where we would have things like the name
6:58:58of the agent, the description, tools,
6:58:59and there are a few other um options
6:59:02here, other parameters like model for
6:59:04example, and then below it, we would
6:59:05have the actual instructions that cloud
6:59:07code would actually read the full list
6:59:09of once it's decided, hey, Mr.
6:59:10Testrunner agent, I'm going to use you.
6:59:13So, literally same way that skills work,
6:59:14and all they are really is just a
6:59:16markdown file. So, I think understanding
6:59:17that really simplifies the concept. And
6:59:19very similar to skills once again
6:59:21because it's just a file they live in
6:59:23your project. So they can be either in
6:59:25your project, they can be personal, or
6:59:26they can be temporary. So personal means
6:59:28if you create an agent, you can use it
6:59:29anytime you're using cloud code
6:59:31regardless of the project. But project
6:59:33means that you would only have it in
6:59:34that specific project. So for example,
6:59:36in this project, this is my herk 2. So
6:59:38basically my executive assistant and
6:59:40I've got my do.cloud right here. In the
6:59:42dotcloud, I've got agents. And here's
6:59:44where I have my six agents right here.
6:59:46And I can only invoke these if I'm in
6:59:48this project. So if I switched open to a
6:59:49different folder, I couldn't use those
6:59:51agents. But if I installed them
6:59:52globally, then I would be able to, but I
6:59:55wouldn't actually see them right here.
6:59:56I'd see them more in like my home global
6:59:58directory for cloud code. And then
7:00:00temporary agents, you can spin up ones
7:00:02just for your session explicitly or with
7:00:03certain plugins, but I've really never
7:00:05used these before, so not going to cover
7:00:07them today. Okay, so we're about to hop
7:00:08in and build our own custom sub aent.
7:00:10But I did want to show you guys that
7:00:11there are some sub agents that cloud
7:00:13code has natively built in that when
7:00:15you're using cloud code you might have
7:00:16actually seen before and you would know
7:00:18if you're using a sub agent because it
7:00:19would have this little command that says
7:00:21agent. So the first one that we have is
7:00:23explore and this one basically searches
7:00:25and analyzes through your codebase. This
7:00:27one runs on haiku and it has readonly
7:00:29permissions. The second one that we have
7:00:31is a planning agent. So a lot of times
7:00:32when you're in plan mode you'll see that
7:00:33this agent gets invoked and this one
7:00:35does research and helps you plan. The
7:00:37model for this one is inherited by the
7:00:39parent model. So if you're on opus a lot
7:00:40of times the sub aent will be running on
7:00:42opus and this one's also read only and
7:00:44then there is a general sub agent which
7:00:45cloud code will invoke when we need some
7:00:47multi-step stuff and this one once again
7:00:49the model is inherited but this general
7:00:51agent can use its hands a bit more and
7:00:53it actually has all tools available.
7:00:54Okay, so now that we understand all of
7:00:56that let's go ahead and jump into cloud
7:00:58code and build our first sub agent. All
7:01:01right, I am in cloud code and I am using
7:01:03cloud code in VS code. Now, what's
7:01:05recommended by Enthropic to actually
7:01:06build sub agents is to use the / aagents
7:01:09command. So, if I come in here and I
7:01:10type in slash agents, you can see that I
7:01:12can click on this right here. And when I
7:01:14click on that, it says that I have to
7:01:15continue in my terminal to actually use
7:01:17this. Part of the limitation with
7:01:18extensions in IDE is that some of the
7:01:21actual slash commands you can't use. But
7:01:23it's not a big deal. You can just open
7:01:25up the terminal. It will run the command
7:01:26like that. You can see. And now we're
7:01:28basically using cloud code in VS Code,
7:01:30but in the terminal. And so once it's
7:01:32ran the / agents command, you can see
7:01:34that we can look at all of our agents.
7:01:35So we have 11. We have the project
7:01:37agents right here. And then we also have
7:01:39the built-in agents that are always
7:01:40available. And so I could choose these
7:01:42agents and I could edit them or do
7:01:43whatever I want. But what I can also do
7:01:44now is I can click on create new agent.
7:01:46So when I hit enter, this is basically
7:01:48going to help us configure the agent and
7:01:50it's going to understand all of the
7:01:51different preferences even like where we
7:01:53want to put this. So it's a lot more
7:01:54helpful than you trying to build your
7:01:56own. So for this demo, I'm going to go
7:01:57ahead and do a project agent, but once
7:01:58again, you could make them personal. I'm
7:02:00going to generate with Claude instead of
7:02:01doing it manually. And now all I have to
7:02:03do is describe what this agent should do
7:02:05and when it should be used and we need
7:02:07to be comprehensive for the best
7:02:08results. Okay, so I just kind of talked
7:02:10into this thing for a little bit and got
7:02:12a brief prompt. I would probably
7:02:13recommend that you make these a little
7:02:15bit more robust, but of course you can
7:02:17always go back and forth later as you're
7:02:18testing them and every time you use the
7:02:20agent, it should be getting better and
7:02:21better because you're going to give it
7:02:22feedback. Now what I'm going to do is
7:02:23I'm going to hit enter and it's
7:02:24basically generating that agent from the
7:02:26description. So, we're going to see that
7:02:27it's going to fill out a new MD file
7:02:29over here, and then we'll be able to
7:02:31test it out. Now, what's cool about this
7:02:32is not only is it just building the
7:02:34agent from my description, but it also
7:02:36is looking through what's in this
7:02:38project. So, it understands me and my
7:02:39business and my goals, and it's going to
7:02:41actually make the agent a little bit
7:02:42tailored towards that. So, this agent
7:02:44that I described was like an AI trend
7:02:46hunter finder, and it's going to
7:02:49understand that I like have a YouTube
7:02:50channel, and that I'm trying to make
7:02:51content on like Nitn, Claude Code, stuff
7:02:53like that. Now, the first option we get
7:02:54is what tools do you want it to use? So,
7:02:56I'm just going to go ahead and choose
7:02:57all. But here's where you could go ahead
7:02:59and say like I only want it to have
7:03:00readonly and you know MCP. That's it.
7:03:03But like I said, I'm just going to
7:03:04choose all and continue. We also now get
7:03:06to choose the model. So, I'm just going
7:03:07to go ahead and go with sonnet. And then
7:03:09you can choose a color for your sub
7:03:10agent, which is pretty cool. You only
7:03:11really see this in the terminal. So,
7:03:13like let's say I'm in the terminal and I
7:03:14say, "Hey, run the AI trend hunter." We
7:03:16would see that this color I'm pointing
7:03:18at my screen like you guys can see. we
7:03:20would see that this color pops up and
7:03:21that's how we know that that agent's
7:03:22being used. We have the ability to
7:03:24configure memory which I know I haven't
7:03:25touched on yet. I will touch on that a
7:03:27little bit later, but for now let's just
7:03:28go ahead and do the recommended enable
7:03:30option. And now what we get to do is we
7:03:32get to see how the agent is configured.
7:03:33So this is the name, this is the
7:03:34location, tools, model, memory, and we
7:03:37can see the description which tells
7:03:38Cloud how to use the agent and then we
7:03:40can see the system prompt. Now it did
7:03:42make this description kind of long. So,
7:03:43we might want to change that cuz that
7:03:44might be using up too many tokens, but
7:03:47that should probably guarantee that it
7:03:48will get invoked at the right times. But
7:03:50anyways, you can see you are an elite AI
7:03:52trend hunter. Here is your mission.
7:03:54Here's what you're looking for. It tells
7:03:56it about how to use its memory and how
7:03:57to update that and how to search for
7:03:59past context, which is helpful for a
7:04:01tool like trends because I don't need to
7:04:03get an update on something that it told
7:04:04me about, you know, like 2 hours ago.
7:04:06And if you're getting an error that says
7:04:07like it can't create it because that
7:04:09folder already exists, then what I would
7:04:10do is I would rename the current folder
7:04:12that's called agents. So just call it
7:04:14like agents one and then hit enter. It
7:04:16should be able to create the agent. Move
7:04:17all your agents into that new folder and
7:04:19then just delete the old one. I think
7:04:20that's just a bug. I've gotten that a
7:04:21few times before, but it still works.
7:04:24But then over here, you also notice that
7:04:25we have agent memory. So right now we
7:04:27have a folder for the AI trend hunter.
7:04:29There's not a memory file in there yet.
7:04:31But for example, our content hunter does
7:04:33have a memory file and it will get
7:04:35created after we run the AI trend hunter
7:04:37one time. So what I'm going to do is I'm
7:04:39just going to run the actual um AI trend
7:04:41hunter agent in the terminal so you guys
7:04:43can see what that color thing looks
7:04:44like. Now what's interesting here is
7:04:46I've already given it the API keys that
7:04:48it needs in this project for like
7:04:50perplexity or for X. But that's
7:04:52something else that you will have to
7:04:53configure if you're setting up an agent
7:04:54for the first time. It needs to know
7:04:56where to access those API keys if it's
7:04:58calling some sort of API or doing other
7:05:00tools like that. And as you can see, we
7:05:02can see that the AI trend hunter is
7:05:04firing off. We can see the web searches
7:05:06and the different things that it's
7:05:07doing. We have the ability to hit
7:05:08control O if we want to expand and see
7:05:10like everything that it's thinking and
7:05:11doing. And right here, we can see the
7:05:13prompts that the main workflow actually
7:05:15fed into this sub agent, which is pretty
7:05:17cool. And the other thing you'll notice
7:05:18is that right here we could hit controlB
7:05:20in order to run this agent in the
7:05:21background which basically just means we
7:05:23would be able to keep talking to the
7:05:24main session while we wait for this
7:05:26background agent to finish up just in
7:05:28case we want to keep working and we
7:05:29don't want to be blocked right now
7:05:30because this is running in the
7:05:31foreground. And by the way, the reason
7:05:33this doesn't have a color is because I
7:05:34chose no color. But I'll just check in
7:05:36with you guys when the sub agent has
7:05:37come back with a summary for us. All
7:05:39right, so sub agent came back. We have
7:05:41some hot signals. We've got some warm
7:05:43signals. We've got top video ideas and a
7:05:46key pattern. And so once again, the
7:05:47value prop here is that the sub agent
7:05:49used about 40,000 tokens, but the main
7:05:51agent only had to actually eat up this
7:05:53many tokens, which is nowhere near
7:05:5540,000. If I run a quick context
7:05:57command, we should be able to see that
7:05:59this session right now has only used
7:06:0029,000. So that just proves that the
7:06:02main agent didn't have to look at all
7:06:04those 40,000 tokens. Now, what else just
7:06:06happened here is we can see that a new
7:06:07file was created in the agent-memory,
7:06:10which is the AI trend hunter. This is
7:06:12the actual file that the sub aent is
7:06:13going to read when it gets woken up. It
7:06:15can see good sources to use. It can see
7:06:17recurring topics. It can see patterns
7:06:19and video angles that worked and also
7:06:21the latest scan. So yes, it wakes
7:06:23[clears throat] up on a completely fresh
7:06:24set of conversation, but it also does
7:06:27get some information to make it a little
7:06:28bit more like recent. Okay, so that was
7:06:31a quick live build of an agent. Now, we
7:06:34used the / agents command to do so,
7:06:36which lets us view agents, create new
7:06:38agents, edit configs of agents, delete
7:06:41them, and see agents that are active.
7:06:43However, the one thing that's not great
7:06:44is when we do the slash agents and we
7:06:46want to like let's say for example edit
7:06:47the AI trend hunter. If we choose edit,
7:06:50it basically makes us open this in an
7:06:52editor and edit the actual text file
7:06:54itself. So, it doesn't really let us use
7:06:55its AI brain in a really nice way to
7:06:58optimize our sub aents. And there's so
7:07:00many little nitty-gritty things that go
7:07:01into how you use them and different
7:07:03agent patterns and things like that. So,
7:07:04what I did is I built you guys a skill
7:07:06called agent builder which is trained on
7:07:08pretty much the complete official docs
7:07:10of cloud code sub aents. And so what
7:07:12that would let us do is you can see
7:07:14right here we have an agent builder. And
7:07:16when I invoke this, it's going to help
7:07:17me like run an audit on that agent that
7:07:19we just built and see how we can
7:07:21basically make it better. So this agent
7:07:23builder skill can be found in my free
7:07:25school community. The link is down in
7:07:26the description. You would go to
7:07:27classroom and then you'd come to agent
7:07:29skills and you'd be able to download
7:07:30that skill and just put it right into
7:07:32your own project. So you can see after I
7:07:34invoked it, it said, "Awesome. I've read
7:07:36the AI trend hunter agent and I've ran
7:07:37the audit. Here's what I found." So
7:07:39there's no current tool restrictions. If
7:07:40there's no max turn set and the
7:07:42description is bloated, here are some
7:07:43other things that would be nice to have.
7:07:45And then here are some things that look
7:07:46good. So then it says, "Hey, do you want
7:07:47me to change these exact things?" And it
7:07:49will actually help you brainstorm over
7:07:50how to make your agent file much, much
7:07:52better. All right, so now that you
7:07:54understand that, let's talk about what
7:07:55else actually goes into making our sub
7:07:56agents really good. So the first thing
7:07:58is auto delegation. And unfortunately, a
7:08:00lot of these things just mean that you
7:08:01have to use them a few times. And the
7:08:03more you use it, the better you
7:08:04understand how to improve it. So, you
7:08:06want to be making sure that in your
7:08:07natural language when you're explaining
7:08:08like the goal that you're looking for,
7:08:10it understands to delegate that to a sub
7:08:13agent or maybe multiple sub agents.
7:08:15Then, we have the idea of foreground
7:08:16versus background. Kind of like I showed
7:08:18you guys, foreground blocks the main
7:08:19chat and the background lets them run
7:08:21behind the scenes so you can keep
7:08:23talking to the main session. And then
7:08:25again, we have the strategy of taking a
7:08:26huge amount of context and having the
7:08:28sub agent process that rather than the
7:08:30main agent. We have this idea of
7:08:32composable skills and agents being able
7:08:34to let them work together. So kind of
7:08:35like that demo that you guys saw. So if
7:08:37I hop back into cloud code, you guys
7:08:39remember I used the carousel skill in
7:08:40that demo. And in this MD, it basically
7:08:43says gather the input and then you plan
7:08:46the carousel. And you do this by
7:08:47delegating to the carousel planner
7:08:49agent. So anytime I invoke this skill,
7:08:51the carousel planner agent gets called
7:08:53on. And I've got another one where it
7:08:55actually delegates to two different
7:08:56agents. If I go to my idea mining and I
7:08:58go to this skill MD, you can see that
7:08:59the first agent we call is the YouTube
7:09:01analyzer and then the second agent that
7:09:03we call is the researcher agent. So once
7:09:05again, this skill calls on different
7:09:07agents. But what else you can do is you
7:09:08can have agents invoke skills. So you'll
7:09:11really start to see as you build up this
7:09:12library of different skills and agents,
7:09:14they can all sort of come together in a
7:09:16really cool way. Tiny agents win. So
7:09:18don't have one that plans and builds and
7:09:20tests and documents and researches and
7:09:22reviews. have one that does each of
7:09:24those individual steps and then the main
7:09:26session can just reconcile all of that.
7:09:28Utilize parallel research. So maybe you
7:09:31need to do some research on O, some on
7:09:32database, some on APIs and then
7:09:34synthesize all that. But just remember
7:09:36they can't talk to each other while
7:09:38they're doing their individual research
7:09:39unless it's an agent team. And then kind
7:09:41of like the way you can use them in
7:09:43skills is you can chain them together.
7:09:44So once you have a specific process, you
7:09:46can say, "Hey, I need you to first use
7:09:48the code reviewer agent and then the
7:09:49optimizer agent and then this agent and
7:09:51then that agent." And that basically
7:09:52becomes an AI workflow where it's a
7:09:54deterministic order of operations, but
7:09:56some of the processes within are
7:09:58non-deterministic. It gives us way more
7:10:00control and that whole process is going
7:10:02to get refined really well. And then
7:10:04remember, start to optimize for cost and
7:10:06speed. So rather than using Opus 4.6 for
7:10:09every single thing you need to process,
7:10:11every single sub agent, start to
7:10:12understand how can I run these ones with
7:10:14Haiku and these ones with Sonnet and see
7:10:16if you can get, you know, parallel
7:10:17agents running that are giving you way
7:10:19more bang for your buck. So remember I
7:10:21touched on the fact that agents can use
7:10:22skills. You could just have them do that
7:10:24in natural language because they can
7:10:25view the whole project just the same way
7:10:27that a main agent would. But you can
7:10:29also preload those. So in the front
7:10:31matter you can tell it specifically here
7:10:32are two skills that you need to use. And
7:10:34then you can also think about things
7:10:35like sub agent compaction. So by default
7:10:38I think it's around 95% when the sub
7:10:39aents will compact. But you could also
7:10:41change that autocompact to make it you
7:10:43know 50% so that you avoid that context
7:10:45rot. So now I think you guys understand
7:10:47what I mean when I said there are so
7:10:49many different things in this doc about
7:10:51how to best use sub agents. Here are all
7:10:54of the different supported front matter
7:10:55fields. We've got name, description,
7:10:57tools, disallowed tools, model,
7:10:59permission mode, max turns, blah blah
7:11:00blah. There's more and more. And so
7:11:02that's why if you want to learn more
7:11:03about sub aents and optimize your sub
7:11:05agents, just use that skilluer skill
7:11:07that I have given you guys. As you can
7:11:09see, this is the markdown file for it.
7:11:11And then I also gave it a reference MD
7:11:13file which is basically even more
7:11:15information about this documentation of
7:11:17how to build sub aents. So this should
7:11:18really help when you are doing this.
7:11:23I just pasted in this message and I'm
7:11:24shooting it off to Claude and it says
7:11:26create a team called Neuroflow of three
7:11:28teammates using Sonnet. The first one is
7:11:29a front-end developer, the second one is
7:11:31a back-end developer and the third one
7:11:33is a QA agent. So this now invoked a
7:11:36tool called team create. And you can see
7:11:37that what it's doing is now that it's
7:11:39created this team, it's spawning up
7:11:40three teammates in parallel. And these
7:11:42are all individual agents. So right now
7:11:44we can see we have our front-end
7:11:45developer, we have our back-end
7:11:47developer, and we have our QA agent. So
7:11:49what's happening is right now we have
7:11:50these three agents working together with
7:11:52our main session. They all share a task
7:11:54list. They can talk to each other. And
7:11:55I'm going to check back in with you guys
7:11:56once this is done. Okay, so this is
7:11:58really interesting. The front end and
7:11:59backend developer sent work over to the
7:12:01QA agent, and then the QA agent found
7:12:03three critical issues. So, the main
7:12:05agent said that it's going to send all
7:12:07of this work right back to those first
7:12:08two agents to take another pass at it.
7:12:09So, here's where you can see it sent off
7:12:11those messages to the front-end
7:12:12developer, the backend dev, and the QA.
7:12:14And now they're all back to work once
7:12:16again. And there we go. The second time
7:12:17the QA agent gives it a pass. All three
7:12:20of those critical issues have been
7:12:21resolved. And then it was able to
7:12:22basically oneshot this website. Now,
7:12:24obviously, there are some things that
7:12:26aren't perfect about this, and we'd want
7:12:27to go back and iterate, but considering
7:12:28in the prompt, all I said was to build
7:12:30me a landing page for a fictional AI
7:12:32startup. And we get all of this text. We
7:12:34get these animations. We get all of this
7:12:36stuff to come in dynamically. And it
7:12:37feels pretty polished. It came up with
7:12:39all the copy, the color scheme, all of
7:12:41it. This is truly one of the most
7:12:42powerful AI agent features I've ever
7:12:44used, but you have to know how to use
7:12:45it, right? Which is why in today's
7:12:47video, I'm going to explain everything
7:12:48you need to know. What they are, how to
7:12:49set them up, how to use them, when not
7:12:51to use them, everything that you need.
7:12:52So, let's not waste any time and get
7:12:54straight into the video. So, agent
7:12:55teams, as you guys saw in the demo, we
7:12:57had one get spun up called Neuroflow.
7:12:59And in that team, we had three agents.
7:13:00We had the front-end dev, the backend
7:13:02dev, and the QA. So, what happens is the
7:13:04main orchestrator, the cloud code
7:13:06session you're talking to, creates these
7:13:07different agents and manages them, but
7:13:09not in the same way that we do sub
7:13:11agents because sub agents work
7:13:12independently and then they send their
7:13:14individual result back to the main
7:13:16agent. Agent teams have a team lead,
7:13:18maybe like a project manager, and it
7:13:20creates all of these different agents
7:13:21and a shared task list. So, the huge
7:13:23unlock here is that individual teammates
7:13:26can talk to each other. So, sometimes
7:13:27there's a dependency. teammate one needs
7:13:29something from teammate two and they can
7:13:30just talk and you can get in these
7:13:32really cool loops especially when you
7:13:33have the QA agent like we just saw in
7:13:35the demo where one of the teammates will
7:13:36basically say hey this isn't good enough
7:13:38and send the work back and then the main
7:13:39agent like I said is just making sure
7:13:41that the tasks are getting done and that
7:13:42they're all high quality. So that's the
7:13:44big difference between sub agents and
7:13:45agent teams and I had to clear that up
7:13:47because I know that's probably where
7:13:48there's some confusion. So, I am going
7:13:50to break more stuff down, but what I
7:13:51wanted to do first is just show you guys
7:13:52how to set this up because I think the
7:13:54best way to learn is just to be able to
7:13:55play around with stuff. And all you have
7:13:56to do to set this up is add one
7:13:58environment variable into your project
7:14:00setting. So, let me show you how that
7:14:01works right now. All right, here I am in
7:14:03cloud code. I like to use it in VS Code,
7:14:05but you can use the agent teams feature
7:14:07wherever you decide to use cloud code.
7:14:09You can see that I am in a brand new
7:14:10project with nothing in it. So, this is
7:14:12exactly what one of your guys' setup
7:14:13should look like if you want to follow
7:14:14along. So, like I said, we need to
7:14:16enable the feature. And I'm going to do
7:14:17that in this demo on the project level.
7:14:19So, what I'm going to do is go to the
7:14:20official Cloud Code documentation for
7:14:22agent teams. And you can see right here
7:14:24that it says they are disabled by
7:14:25default because it's an experimental
7:14:27feature. So, you have to enable them by
7:14:28adding this variable into your
7:14:30settings.json. So, what I'm going to do
7:14:32is literally just copy this JSON right
7:14:33here and come into Cloud Code and say,
7:14:35"Hey, I need you to put this in our
7:14:37local settings in this project." And
7:14:40then just paste in that JSON prompt. And
7:14:42I'm going to go ahead and shoot that
7:14:43off. And that should basically be able
7:14:44to create that file for us. So you can
7:14:46see that it set everything up. We now
7:14:48have acloud folder. If I click into
7:14:49here, we have a settings.local.json
7:14:51and it has put that command in there.
7:14:54And now our project should be set up to
7:14:55actually be able to use agent teams. Now
7:14:57before we dive in and I start showing
7:14:59you guys how to do that, there is one
7:15:00thing that I recommend doing first. And
7:15:02that's basically training your cloud
7:15:03code project on how agent teams work so
7:15:06that they can actually be used as
7:15:07effectively as possible. So the easiest
7:15:08way to do this is you go to the
7:15:10documentation on agent teams. You take
7:15:12the URL and you copy it. And then I
7:15:14said, "Hey, create me a master reference
7:15:15guide for agent teams in a folder called
7:15:17docs. This will be used to help you
7:15:19build better and more effective agents
7:15:20in the future. And now it's going to
7:15:22read through that documentation." And
7:15:23now if you ever have questions about
7:15:24agent teams or if it ever needs to look
7:15:26up something while it is building them,
7:15:28it already has that locally here stored
7:15:29as markdown. So it's going to be much
7:15:31quicker. And it just created this
7:15:32documentation about enabling them, when
7:15:34to use them, display modes, task
7:15:35management, hooks, best practices, tons
7:15:38of stuff like that. And that can be
7:15:39found in the docs folder in this
7:15:41section, which is a full markdown file
7:15:44with hundreds of lines. And that's just
7:15:46a little tip. That's something that I
7:15:47like to do whenever I have like maybe a
7:15:49big MCP server or certain documentation
7:15:51that I know it might need to look at
7:15:52constantly. So, now that we're pretty
7:15:54much set up and ready to start building,
7:15:55let's talk about prompting. How do we
7:15:57actually tell Cloud Code to build us
7:15:58agent teams, but not just to build them,
7:16:00but to actually make them really good to
7:16:02give you what you want? Because the
7:16:03truth about agent teams is that they are
7:16:04more expensive and they are a bit
7:16:06slower. But you do get much higher
7:16:08quality if you use them right. Now the
7:16:10good news is you can pretty much invoke
7:16:11them just using natural language. So I
7:16:13kind of follow this pattern. Create a
7:16:15team of X number of agents using X
7:16:18model. So Hiku, Sonnet or Opus. And then
7:16:20you basically just say the agents that
7:16:22you want. You would say the first agent
7:16:25is X ro this agent should be doing this
7:16:28and it should produce me this. It can
7:16:30talk to the other agents to do X Y and
7:16:32Z. And so pretty much just listing that
7:16:34out in natural language, whether that be
7:16:37an API designer, a database engineer,
7:16:39and or a test writer. So let's take a
7:16:41real quick look at an example prompt. So
7:16:43I'm going to read this full one out.
7:16:44Now, what you'll notice is I start off
7:16:46by establishing a goal. The reason I do
7:16:48this is because when the agents wake up,
7:16:50they have no context. They basically
7:16:52only get the prompts that the main
7:16:54session feeds into them. So if we tell
7:16:56the main agent a goal to give to these
7:16:58sub aents, they understand a little bit
7:16:59better, you know, like what they're
7:17:00working towards, but also why they have
7:17:02their teammates next to them. So the
7:17:04goal here is to build a working full
7:17:06stack app with a REST API and a React
7:17:08front end. The end result should be a
7:17:09running app that I can view on a local
7:17:11host. It should have users and post
7:17:13functionality plus a QA test report
7:17:15confirming that everything works. So
7:17:16then I said, hey, create me a team of
7:17:18three teammates using Sonnet. The first
7:17:20one's a back-end dev. It should be doing
7:17:21this. The second one is a front-end dev
7:17:23and it should be doing this. And the
7:17:25third one is a QA agent that should be
7:17:27doing this. You can see that in the
7:17:29descriptions I said when you're done
7:17:31message the front-end dev. And then in
7:17:32this one I said wait for the backend
7:17:34dev's message and then you will send all
7:17:36the stuff to the QA. And then I'm saying
7:17:38what the final deliverables should be
7:17:40because the main agent spins these three
7:17:42up and then it's going to get a bunch of
7:17:43information back. So what do I actually
7:17:44want at the end of the day? I want a
7:17:46running app. I want a report about pass
7:17:49and fail tests. And then I want a doc
7:17:52which is basically what was built, key
7:17:54decisions, and how we run this moving
7:17:56forward. So we're about to hop right
7:17:57back into cloud code and live prompt an
7:18:00agent. But real quick, let's talk about
7:18:01some dos and don'ts. So do have each
7:18:04agent own specific files because if you
7:18:06don't do this and agents are sharing
7:18:08files, they might overwrite each other's
7:18:09work, which is not good. Do define the
7:18:11output. Don't use vague deliverables. Do
7:18:14name recipients. Don't just assume that
7:18:16they're going to understand who to talk
7:18:17to and why. Do have about three to five
7:18:20teammates. Don't go for massive agent
7:18:22swarms of 10 plus. That'll also be 10
7:18:24times more expensive. And do give full
7:18:26context because of the fact that no
7:18:27history is given beforehand. Now, of
7:18:30course, they can still read everything
7:18:31in the project. They can still look
7:18:32through all of those files, but no
7:18:34context is fed in initially. And I will
7:18:36show you exactly what I mean by that
7:18:38when we go in here and spin up a new
7:18:39agent team. Okay. So, I'm in that
7:18:41project that we set up together. I'm in
7:18:42a fresh session and I'm going to send
7:18:44off this prompt. I said that the goal is
7:18:46to help me clean up the workspace. We
7:18:47have three agents called research team.
7:18:49We're using Sonnet. We've got a
7:18:50researcher, a strategist, and a critic.
7:18:52And they're basically just going to read
7:18:54through this project and make sure that
7:18:55everything's accurate and make sure that
7:18:56we're set up good. Let's take a look at
7:18:58what's going on. It's creating the
7:18:59research team, right? So, it's created
7:19:01the team, and now we have a to-do list.
7:19:03Now, what it's doing is it's going to
7:19:04spawn the three teammates in parallel.
7:19:06And when it spawns these, I'm going to
7:19:08actually show you how. So, the first one
7:19:09is a researcher. If I click into this,
7:19:12you can see that this says in. So, this
7:19:13is basically saying this is what the
7:19:15main agent sent to the agent. you are
7:19:18the researcher on the research team.
7:19:20Here is what your job is and you have to
7:19:21be thorough and include anything that
7:19:23might be helpful. So this is basically
7:19:24the prompt that spun up that agent. You
7:19:27can see same exact thing happened for
7:19:29the strategist and for the critic. And
7:19:31if I clicked in, we could once again
7:19:32read exactly what they were prompted to
7:19:34do and the step-by-step instructions,
7:19:36including stuff like when you're done,
7:19:38send your five use cases to the critic
7:19:40teammate using the send message tool,
7:19:42which once again validates that these
7:19:44agents are able to talk to each other
7:19:45and send messages to each other. So now
7:19:47we can see that all three of our agents
7:19:49are running and they are all basically
7:19:51just waiting for their turn. And you'll
7:19:53notice what it does is pretty much every
7:19:54time there's a new update, it updates
7:19:56me. So here comes another live update.
7:19:57Let's see if the researcher is finally
7:19:58done. There we go. It's done. So now
7:20:01what happens is we sent a message off to
7:20:03the researcher. And let's go ahead and
7:20:04see what that message actually said. So
7:20:06this is the main agent talking to the
7:20:08researcher. It said, "Did you send your
7:20:11structured inventory to both the
7:20:12strategist and the critic? Please make
7:20:14sure the strategist also received it.
7:20:16You were asked to message both
7:20:17teammates. And then we can see that the
7:20:18researcher confirmed that both teammates
7:20:20received the inventory and now the
7:20:22critic is running. All right, so
7:20:23everything just finished up. All the
7:20:24reports are here. But real quick, I
7:20:26wanted to draw your attention to this.
7:20:28The main agent said, "Cool. Let me shut
7:20:29down the teammates and finalize." And
7:20:32I'll touch on this a little bit later.
7:20:33But now the main agent has sent a
7:20:35message to each of them, the researcher,
7:20:37the strategist, and the critic. And
7:20:38basically said, "You're done. Save your
7:20:40work." So anyways, we'll come back to
7:20:42that in a little bit. But we now have an
7:20:43output which is a new document over
7:20:45here. Agent teams patterns and it found
7:20:47a ton of stuff. There were 11
7:20:49documentation gaps identified that are
7:20:51worth reviewing against your reference
7:20:53doc. So anyways, let's just click into
7:20:54the doc real quick. We're not going to
7:20:55read this whole thing cuz I'm assuming
7:20:57it is super long. But this is the actual
7:20:58output that we just got from this agent
7:21:00team. And you can see that this thing is
7:21:02insanely thorough. So if you wanted to
7:21:04really really understand how agent teams
7:21:05work, then spin up an agent team to help
7:21:07you explain agent teams. Now, what
7:21:09you'll notice here is we were kind of
7:21:10able to see what was going on, but not
7:21:12really under the hood. We couldn't
7:21:14actually tell what the agents were
7:21:15thinking or doing. And that's because
7:21:17we're doing this in the Cloud Code
7:21:19extension. If you do this in your
7:21:20terminal, and specifically if you have
7:21:22T-Mox installed, you can actually see
7:21:23the different agents working and
7:21:25thinking, and you can individually send
7:21:27messages to them. Because right here,
7:21:28we're kind of only communicating through
7:21:30the main session, and the main session
7:21:32sends messages to the other ones. But
7:21:33one of the value props of agent teams is
7:21:35that I could individually message a sub
7:21:37agent if I wanted to. So, let me show
7:21:38you what that looks like. All right, so
7:21:40right now I'm running Cloud Code in a
7:21:42T-m terminal. Now, if you're on Windows,
7:21:44you have to take a little bit of a
7:21:44workaround, but you just have to be in a
7:21:47T-m terminal. So, I'm not going to do a
7:21:48full setup video on it right now, but I
7:21:50literally just had Cloud Code walk me
7:21:51through it, and it was super simple. But
7:21:53anyways, what I did here is I just
7:21:54pasted in this prompt, which obviously
7:21:56is like we talked about. We have the
7:21:57goal. We say create me an agent team,
7:22:00and then we have our front-end dev, our
7:22:01backend dev, and our QA. This is
7:22:03basically the same exact prompt that I
7:22:04ran in the demo. So, this isn't to show
7:22:05you the actual deliverable. What I want
7:22:07to show you guys here is the way that we
7:22:08can visually see this. So, right here,
7:22:10what it's going to do is it's going to
7:22:12spin up that agent team for us, right?
7:22:13It's setting up the task dependencies
7:22:15and it's assigning owners. And now it's
7:22:17spawning those agents. And there we go.
7:22:19We just got our front-end dev created
7:22:20right here. And this is the blue agent.
7:22:23We have another one right here, which is
7:22:24the backend dev. And this is the green
7:22:26agent. And there we go. We just got our
7:22:28QA agent, which is the yellow one. So
7:22:30now I very clearly can see what each of
7:22:32these agents is doing, which is super
7:22:34cool. And now if I wanted to, I could
7:22:35come over here and I could check on the
7:22:37team status with the main session. I
7:22:38could come up here and I could talk to
7:22:40the front-end dev. I could approve
7:22:41things or I could give it more info.
7:22:42Same exact thing with the QA or same
7:22:44exact thing with the backend dev. So now
7:22:46I literally have an agent team that I
7:22:48can watch and I can interact with any of
7:22:50them. And I can also watch them do
7:22:52research, create things, talk to each
7:22:54other. It's super super cool. So like I
7:22:56said, I'm not going to run this whole
7:22:57thing out. I just wanted to show you
7:22:58guys that this is possible. Okay, now
7:23:01that we've seen some cool demos, let's
7:23:03talk about how do we actually make these
7:23:04things better and better and understand
7:23:06a little bit more about what's going on.
7:23:08So, here are three key rules. The first
7:23:10one is that each of these agents has
7:23:12their own territory. So, they should
7:23:13have their own file and they should be
7:23:15working on their own deliverables. They
7:23:16can send them across and they can
7:23:18communicate, but they should only really
7:23:19all be editing their own thing. The
7:23:21second thing is once again direct
7:23:22messaging. They can talk to each other.
7:23:24They don't have to use the middleman of
7:23:25the main session. And then the third
7:23:27piece is that they can be working at the
7:23:28same time. It doesn't have to be agent
7:23:30one hands off to agent two and then
7:23:32agent two hands off to agent three
7:23:33because that honestly might not even
7:23:35call for an agent team. Agent teams work
7:23:37together in parallel and need to
7:23:38communicate throughout the whole
7:23:39process. So what do teammates instantly
7:23:41know when they wake up because we know
7:23:43that they don't have any context from
7:23:44the jump. What they do have is they
7:23:46inherit the permissions from the main
7:23:48session. So if you're on bypass
7:23:49permissions then all of your agents are
7:23:51going to be on bypass permissions. If
7:23:52you allow all bash commands then those
7:23:54same permissions will once again be
7:23:55inherited by the teammates. But the
7:23:58other thing to know that's very
7:23:59important is that any of your files, any
7:24:02of your MCP servers, any of your skills,
7:24:04all of the teammates can use and access
7:24:05those things. We also have a really cool
7:24:07ability to use something called plan
7:24:09approval mode. So you guys know how I've
7:24:11told you always start in plan mode. If
7:24:12you plan with your main session before
7:24:15anything actually happens, it's way
7:24:17better. What we can do is we can have
7:24:19all of those agent teammates plan first
7:24:21and they have to basically get their
7:24:23plan approved by the main agent before
7:24:25they're actually allowed to go execute.
7:24:26So, it's really cool. You could also set
7:24:28it up where you're actually the one who
7:24:29has to approve every single plan. But, I
7:24:31think it's probably better to just have
7:24:33the main session do that. Or maybe even
7:24:35one of the teammates is just the plan
7:24:37reviewer and approver. So, I wanted to
7:24:39talk about some common pitfalls or
7:24:41mistakes that you might be making and
7:24:43what the fix could be for that. So, the
7:24:45first one is if the agents keep asking
7:24:47permissions and they keep stopping for
7:24:49that, you can preapprove certain tools.
7:24:51So, that would be in your project
7:24:53settings or your local settings. you can
7:24:54allow certain commands and that way they
7:24:56won't stop to ask you something every
7:24:58couple seconds. If the deliverables
7:25:00aren't coming out feeling holistic,
7:25:02maybe they're being overwritten. So,
7:25:03make sure that you assign file owners.
7:25:05If you spin up an agent team and you
7:25:06realize that one of the agents isn't
7:25:07really doing much or is just sitting
7:25:09around, then maybe you want to
7:25:10specifically make sure you're assigning
7:25:12each agent work or some sort of
7:25:13dependency in your plan in your prompt.
7:25:16If you're burning through way too many
7:25:17tokens, just use fewer agents. If it
7:25:19seems like your agents are losing work,
7:25:21then tell them to basically store
7:25:22everything as a temporary file that they
7:25:24can then call on later. And if you're
7:25:26getting the wrong approval and it just
7:25:27seems like it's off, then maybe just try
7:25:28to have you be the one who approves
7:25:29things to start until you understand the
7:25:31flow of how these teams work a little
7:25:32better. All right, so next I wanted to
7:25:34talk about when to use agent teams
7:25:35because like I said earlier, they can be
7:25:37slow and expensive. So you really just
7:25:39want to use them when you need something
7:25:40pretty complex done and you need lots of
7:25:42different specialized agents. So think
7:25:44about using them if your specific
7:25:46process or project has multiple
7:25:48different areas. And that way you can
7:25:49have one specialize in each of those. If
7:25:51you need those things to be done in
7:25:52parallel, if you need them to be able to
7:25:54react to each other, assign tasks to
7:25:56each other, communicate with each other,
7:25:57and if something needs to be done at a
7:25:58really high quality, and you want tons
7:26:00of different steps to make sure, then an
7:26:01agent team is probably a decent idea.
7:26:03Now, if you have a process that could be
7:26:04done sequentially, meaning every time it
7:26:06basically goes 1 2 3, and those steps
7:26:09are dependent on each other, then maybe
7:26:10an agent team isn't the right call.
7:26:12Maybe that's just sub agents. If you
7:26:14need everything in one specific
7:26:15conversation history or one context
7:26:16window, then don't use teams. If you're
7:26:18just kind of working on the same files,
7:26:20don't use teams. And if it's a very
7:26:22simple task, then agent teams would be
7:26:24overkill. There might be a lot of times
7:26:25where you'd be able to use sub agents
7:26:26instead. Like I said, with sequential or
7:26:29if you need a very focused result, if
7:26:30you don't need the agents to
7:26:31communicate, and if you want to save
7:26:33some tokens, because once again, if you
7:26:34have three sessions running, that's
7:26:36basically going to be three times the
7:26:37cost. So if you have five, it'll be five
7:26:39times the cost. Which means I like to
7:26:41stay around maybe two to five agents
7:26:43max. You can keep them running parallel.
7:26:45Otherwise, you can use sub agents and
7:26:47make sure that you are shutting them
7:26:49down if you see them early on going off
7:26:52down the wrong path. Which is another
7:26:53reason why I think it's helpful to use
7:26:55the T-Mox version so you can actually
7:26:56see them in that split pane view. And
7:26:58when I say shutdown, which is kind of
7:27:00what we saw earlier, I just mean
7:27:02basically at the end of every session
7:27:04saving your work. Because remember right
7:27:05here how we saw the main session say,
7:27:07"Hey, this is a shutdown request." The
7:27:10researcher agent here could have said,
7:27:11"I'm not done yet. Let me save stuff.
7:27:13Don't shut me down yet." So when the
7:27:14teammates actually confirmed that
7:27:15they're ready to be shut down, that
7:27:16means that everything's good. And we can
7:27:18essentially cleanly save that work. So
7:27:20everything gets cleaned up and then
7:27:22we're good to close the session and shut
7:27:23down that agent team rather than just
7:27:25force killing it right away where things
7:27:27might be all out of control and not
7:27:29cleaned up yet.
7:27:32So when you give a tool as powerful as
7:27:34cloud code access to literally control a
7:27:37browser, if you think about it, you can
7:27:38actually automate anything, whether
7:27:40that's stress testing an app,
7:27:41downloading reports, or even playing a
7:27:43game like you guys saw in my other
7:27:45video. So today I walk through three use
7:27:47cases that you can use when you connect
7:27:49Playright CLI to cloud code to control
7:27:51your browser. The first one is having it
7:27:53actually QA a web app and finding bugs
7:27:55and then fixing them. The second one is
7:27:57searching for dentists and finding their
7:27:59contact information. And then the third
7:28:00one, a question that everyone's asking
7:28:02is can browser use actually control
7:28:04sessions where you need to be logged in.
7:28:05So I show off what I'm able to do in my
7:28:07school account. So I don't want to waste
7:28:08any time. Let's get straight into the
7:28:09video. All right. So I am in a fresh
7:28:11cloud code project. As you can see, I'm
7:28:13in a folder called browser automation
7:28:14demo, but there's nothing in here just
7:28:16so you guys can follow along with
7:28:17exactly what I'm doing. So, the first
7:28:18thing to do is install Playright CLI so
7:28:21that we can actually do some browser
7:28:22automation. So, I'm just going to go
7:28:23into plan mode and I'm going to say,
7:28:24hey, cloud code, I want to use Playright
7:28:27CLI in order to do some browser
7:28:29automation stuff. Whether that means
7:28:31testing out web apps or taking
7:28:33screenshots of things, whatever the
7:28:34case, I just need you to figure out how
7:28:36you can install this for me and then go
7:28:37ahead and build a plan and let's do it.
7:28:39So, I know I'm telling you guys to use
7:28:41Playright CLI, but this is literally
7:28:42what I did when I decided I wanted to
7:28:44try out some browser automations, except
7:28:45for I said, "Hey, I need you to research
7:28:47the different tools and the different
7:28:48pros and cons and help me figure out
7:28:50what I should do. I played around with
7:28:52Chrome DevTools for a little bit, but
7:28:53now I'm using Playright CLI." And one
7:28:55reason for that, while we're letting
7:28:56this load, is that if I do /context in
7:28:58here, you can see that the Chrome
7:29:00DevTools MCP takes up so many tokens
7:29:03because there's so many different tools
7:29:05and each tool has a description. So,
7:29:06that's why I was like, hm, maybe I don't
7:29:07want to use the MCP server. Let me just
7:29:09go ahead and try this Playright CLI and
7:29:11it works really well. All right, so the
7:29:13plan is done. We have an empty project.
7:29:14We're going to initialize the project,
7:29:16install Playright, create a demo script,
7:29:18test things out. I'm going to go ahead
7:29:20and accept this plan. And now I'll check
7:29:21in with you guys when that's done. All
7:29:23right, so you can see that it is done.
7:29:24It's been installed and then it also
7:29:26tried it out already. So it ran a script
7:29:28to open up this page and then screenshot
7:29:30it. And that got saved right here. So as
7:29:32you can see, it was able to do that. It
7:29:33was able to take a screenshot, which
7:29:35means that everything is working. So now
7:29:36that we have this initial stuff set up,
7:29:38I'm just going to go ahead and do a
7:29:39slashinit to basically just initialize
7:29:42the environment and give us a claw.mmd
7:29:43file. All right, so there are tons of
7:29:45great use cases when it comes to using a
7:29:47browser automation, whether that is
7:29:48looking through Amazon or applying to a
7:29:50ton of positions or, you know,
7:29:52downloading reports from websites that
7:29:54don't have an API. And what's awesome
7:29:56about this is it will build out these
7:29:58scripts that run the browser automation,
7:30:01right? And when you pair those custom
7:30:02scripts with a skill, it gets really
7:30:04powerful because then that process of
7:30:06opening up the browser and doing certain
7:30:08things becomes consistent and
7:30:09repeatable. But one thing that I think
7:30:11is really cool about browser automations
7:30:13is the ability to basically test things
7:30:15and automate QA. So an example I wanted
7:30:17to walk through with you guys is I'm
7:30:19going to use cloud code in here to build
7:30:21us a quick web app that's going to be
7:30:23like a multi-page form and I don't want
7:30:26to test it myself. I want to have it
7:30:28test it and then it has to find the
7:30:30bugs, find suggestions, and then fix
7:30:32itself. So, it's going to be us building
7:30:33a web app or a form and it's going to be
7:30:35like completely hands-off. That's the
7:30:37goal at least. So, let's hop into plan
7:30:39mode and let's get started. I need you
7:30:40to build me a form submission website.
7:30:43Now, I want this to basically be one
7:30:46page per question. So, as soon as I open
7:30:48up the page, it should ask me for my
7:30:49first name and then there should be a
7:30:50button that prompts me to hit enter.
7:30:52When I hit enter, I go to the next page.
7:30:54And then I have my last name. And then I
7:30:56have my phone number. And then I have my
7:30:57business. And I have maybe, let's just
7:30:59say, eight questions about me and kind
7:31:02of getting me onboarded into this, you
7:31:04know, fake business or whatever. I just
7:31:06want to do this to see and test the
7:31:08functionality of what you're able to
7:31:10build me here. Okay. So, that is phase
7:31:11one. Phase one, we're going to get an
7:31:13initial version of the website. And then
7:31:16before I even open it up and test it,
7:31:17we're going to go into plan mode again
7:31:19and see if just cloud code with the
7:31:21browser automation can test it, iterate,
7:31:24test it, iterate. Okay, here is our
7:31:26plan. Multi-page onboarding form. The
7:31:28user wants a polished multi-page
7:31:30onboarding form. We have the
7:31:32architecture. There's 12 total
7:31:33questions, which look good to me. We've
7:31:35got some files. It's going to make some
7:31:37implementation details here. And all of
7:31:39this looks good to me. So, we're going
7:31:40to go back into the main session and
7:31:42accept those changes. Okay, so that just
7:31:44finished up. We have the files have been
7:31:46created. We've got a server that this is
7:31:48running on. It's going to be on a local
7:31:50host. We also have 12 one per page
7:31:52questions. And what I noticed it doing
7:31:55is it was taking screenshots and I
7:31:57didn't even tell it to. So, if I open up
7:31:58screenshots, you can see we now have
7:32:00these pictures. So, let me just look at
7:32:02these. We've got number one, which is
7:32:04what is your first name? Number two,
7:32:06last name, best phone number to reach
7:32:08you at. There's even, if you can see,
7:32:10there's like one of those scroll bars at
7:32:11the top where it's like showing, you
7:32:13know, how many questions in they are.
7:32:15So, for a oneshot prompt and the fact
7:32:16that it's already flipping through and
7:32:18taking screenshots, I'm pretty
7:32:19impressed. But, we have to see how it's
7:32:21able to actually test it out. So, now
7:32:23I'm going to go back into plan mode and
7:32:24I'm going to say, "Yep, absolutely. Spin
7:32:27up a server so that you can actually run
7:32:29this." And then what I want you to do is
7:32:31use your browser use and test it out. So
7:32:34fill in the fields, click through, and
7:32:37if there's any bugs or if there's
7:32:39anything wrong with the functionality of
7:32:40the site, make note so that you can go
7:32:42ahead and fix the actual site itself.
7:32:44And I also want you to do this in a
7:32:46headed browser so that I can watch
7:32:48what's going on. So I'm going to shoot
7:32:50that off. It's going to make a plan. But
7:32:51that is one important distinction. By
7:32:53default, you could say, hey, I always
7:32:55want you to run playright in a headed
7:32:57browser. But there is headed or
7:32:58headless, meaning it could be running
7:33:00this headless where it wouldn't show us
7:33:02on screen, but it is still running on
7:33:04its own tab and clicking through things.
7:33:05Okay, here is the QA pass. The headed
7:33:07browser testing, I'm not going to read
7:33:09this. I think that this is pretty clear
7:33:10of what we want. I'm going to go ahead
7:33:12and accept it. So, what happens is for
7:33:13every kind of like bot that we want to
7:33:15run, it has to write its own script. So,
7:33:17I believe that this is this one that
7:33:18it's writing, which is the test
7:33:20onboarding.js.
7:33:22And this is basically what instructs the
7:33:23bot on what to do. And then when we
7:33:25wanted to turn this into a skill, which
7:33:27is basically like QAing our website, we
7:33:29would say, okay, when you want to QA,
7:33:31then you run the bot and then you take
7:33:33the feedback from the bot and then you
7:33:34make changes and then you run the bot
7:33:36again. And so that's how we can turn
7:33:37kind of like all of these different
7:33:38pieces into an actual process. Okay.
7:33:41Also, that wasn't even the script. The
7:33:42script that it's writing for this is QA-
7:33:45test. So now you see what I mean. Okay.
7:33:47So the window just popped up. I'm going
7:33:48to move it into view and we should be
7:33:50able to watch it fill in. It's did it
7:33:51did Nathan. Okay, it went backwards. Now
7:33:54it's on the second one. Nathan Harrison
7:33:57filling in a phone number. It's
7:33:58continuing to click through. Harrison
7:33:59Tech. We've got a email. We've got
7:34:02founder and CEO. It shows a company
7:34:05size. Look from a drop down. Okay.
7:34:08Technology referral. Primary goal.
7:34:12Website. Continue. And anything else?
7:34:15Okay. It looks like it's having a little
7:34:16bit of trouble here now. It keeps kind
7:34:18of glitching out. But once again, I'm
7:34:19not going to stop it. I'm going to let
7:34:20it figure out what it needs to do. Okay.
7:34:22So it just shut down and now it said
7:34:24okay found some issues. Let me review
7:34:26the screenshots to understand what
7:34:27happened. So what you guys didn't notice
7:34:29is that in there it took more
7:34:31screenshots. So we can see now we have
7:34:33QA and we have all of these screenshots
7:34:35that it took during that test. So it
7:34:37says okay the UI looks polished. That's
7:34:39great. It did look really good. But the
7:34:41first bug is that enter on text area
7:34:43didn't advance to review. It stayed on
7:34:44the notes page. The second bug is that
7:34:47the review page never loaded. The test
7:34:48found zero review items and the edit
7:34:50button was intercepted by a stale page
7:34:52overlay. And now it was able to use that
7:34:54feedback to fix the actual bugs in the
7:34:56site itself. And this is just really
7:34:58cool because if you've ever built
7:34:59software or websites or apps or
7:35:01whatever, there are so many bugs and you
7:35:03don't always find them until it's maybe
7:35:04too late or maybe a customer found it.
7:35:06So, the fact that you can automate QA by
7:35:08having multiple different headed
7:35:10browsers or even headless browsers spin
7:35:12up and you can say, "Hey, you test for
7:35:13X, you test for Y, you test for Z," and
7:35:15just have them running and fixing and
7:35:17running and fixing is a complete game
7:35:19changer. And now, without me even
7:35:20asking, it says, "Okay, I'm going to
7:35:22start the server and I'm going to rerun
7:35:23the test." So, we're basically getting
7:35:25it stuck in this loop of testing,
7:35:27validating, testing, validating, and you
7:35:30know, it's really, really cool. So, I
7:35:32think it's doing another quick test just
7:35:33to see um if it truly is good. But,
7:35:43okay, nice. So, second test went all the
7:35:45way through and now it should be
7:35:46shutting down and telling us what it
7:35:49learned. As you can see, it just shut
7:35:50down. So, now you can see it has finally
7:35:52passed the tests and it goes ahead and
7:35:54ends the process because now the server
7:35:56is good. So, like I said, the next step
7:35:57from there would be to turn that into a
7:35:59skill. But what else can we do with
7:36:00browser automations? Because they're
7:36:02definitely not perfect, right? So, let's
7:36:03say we wanted to do something on the
7:36:05web. Let's say we wanted it to just be
7:36:07able to go to Google, search for, I
7:36:09don't know, let's just say dentists, and
7:36:12maybe capture like some phone numbers.
7:36:14All right, here is our plan. Build a
7:36:15Playright script that automates Google
7:36:17search to find dentist offices in
7:36:19California. We will basically launch the
7:36:21browser, do the Google search, collect
7:36:23the links, visit each site, and print a
7:36:25summary. So, that looks good to me.
7:36:27We're going to go ahead and shoot this
7:36:28off and hopefully it's going to be
7:36:31decent on the first try. I'm going to
7:36:33tell you guys right now it's probably
7:36:34not. But what's going to happen is we're
7:36:36going to have it learn. So, it's going
7:36:37to open up the browser. It's going to
7:36:39fail and we're going to say, "Okay, keep
7:36:40learning. Keep updating the script and
7:36:42don't stop until you're done." I
7:36:45literally stepped away to go get water.
7:36:47And it actually I came back and I saw a
7:36:49browser open with some dentist offices
7:36:52up. Wow. So, it visited all five sites
7:36:55and it was able to find some phone
7:36:58numbers. Now, that's really cool because
7:37:00it actually learned while it was going
7:37:01and it said that Google blocked the
7:37:02automation. So, it switched to duck.go.
7:37:05So, I told it this time, don't stop
7:37:07until it actually finds five phone
7:37:09numbers. So, it's going to be a little
7:37:10bit more aggressive. And this time, I'm
7:37:12actually here to watch it.
7:37:16Okay, so here we are. It just searched
7:37:18for dentist office LA. You can see
7:37:21there's a phone number. I just saw one
7:37:22on screen, so it probably grabbed that
7:37:24screenshot. I see another one on this
7:37:25page. So, it's getting pretty lucky
7:37:26here. It's finding the phone numbers
7:37:28pretty quick. But what I just noticed is
7:37:30it's still going to the site and it's
7:37:31still clicking on the contact button.
7:37:33Even though the phone number was up in
7:37:34the top right visible, it still clicked
7:37:36on the contact button, which I thought
7:37:38is pretty cool. So, this time it got all
7:37:40these phone numbers for these dental
7:37:41offices. So, naturally, the next
7:37:43question that I'm sure you guys have is
7:37:45how could it work on, you know, things
7:37:47that I'm already logged into? So, let's
7:37:49just find out. So, I'm going to go to
7:37:50plan mode and just ask it what would
7:37:52happen. So, there are a couple of
7:37:53approaches. We could do persistent
7:37:55browser profile, meaning it can launch a
7:37:57browser using my existing Chrome user
7:37:59data, which already has me logged into
7:38:01school. We could do manual login and
7:38:03handoff in headed mode, or we could
7:38:06connect to a running browser. Okay, so
7:38:08that's pretty cool. It could open up the
7:38:09browser. I could sign in and then I
7:38:11could just basically keep that signed in
7:38:12and have it do whatever it needs to do.
7:38:14But I do want to try option one, which
7:38:16it recommended in the first place. Okay,
7:38:19so we have the plan which is going to go
7:38:20into my school community. It's going to
7:38:22go to the win channel and it's going to
7:38:24like those posts. So we've got the
7:38:26context, the approach. Obviously you
7:38:27guys know what these plans look like.
7:38:29Let's just see what it did on this first
7:38:30try and I'm going to go ahead and
7:38:32accept. Okay, so apparently when it runs
7:38:34this, I'm going to have to log in
7:38:35manually the first time, but then in the
7:38:37future it will basically save that
7:38:39session. So let's see how that works.
7:38:41All right, it opened up school.com.
7:38:43Okay, it opened up school.com.
7:38:46Okay, so what just happened is it opened
7:38:48it up and then it closed it and then it
7:38:49opened it up and it went to my community
7:38:51and closed it and now I think it's
7:38:52actually going to let me log in. So let
7:38:54me try that real quick. Okay, so I
7:38:55logged in and then the browser shut
7:38:56down. So now I said, "Okay, cool. I
7:38:57logged in." Hopefully now when it
7:38:59launches up this browser, it already has
7:39:01me logged in. Okay, so there it is.
7:39:03Nice. You can see that I am already
7:39:04logged in, which is great. Okay, so it's
7:39:07trying to find the wins channel. It
7:39:08looks like it found it. All of these
7:39:09look like they are wins. Now let's see
7:39:11what it's able to do here. If it's Oh, I
7:39:13just saw it try to like the post. Okay,
7:39:16it just liked that one.
7:39:19Interesting. So, it's like liking them
7:39:21and unliking them. So, that's one thing
7:39:22we're going to have to fix. But, it is
7:39:24scrolling down. It got to the wins
7:39:26channel and it is now trying to like
7:39:28post, but as you guys can see, it's not
7:39:30perfect. So, I hope that it's able to
7:39:31realize that it's not really liking
7:39:32them. Yeah, it's basically going like,
7:39:36unlike like unlike, and it also just
7:39:39said liking too fast. So, let's see if
7:39:42it's able to learn. Okay, there we go.
7:39:43All right. All right. So, what I'm going
7:39:44to do at this point is, you know, I've
7:39:45seen enough. I'm going to shut this
7:39:46down. And hopefully, it's able to use
7:39:49that feedback to get better. So, the
7:39:50login persisted. It found the wins
7:39:52channel. It even found the heart-shaped
7:39:54SVG buttons, but it crashed during the
7:39:56page evaluate for liking. So, it's seen
7:39:58all that and now it's going to make the
7:40:00script even better. So, I will say with
7:40:01browser automations, I think that this
7:40:03is completely normal, which is why I had
7:40:05a little bit of doubt earlier in the
7:40:07dentist example, but literally every
7:40:09time that you use the script, it's going
7:40:11to get better. So, it's about to run it
7:40:13again. Let's see how it does. This time,
7:40:14it was able to just find the wins
7:40:16channel. Now, one thing I might want to
7:40:17tell it is I might want to sort it from
7:40:19newest posts rather than just going
7:40:21straight to the winds channel. And it
7:40:23went ahead and shut down again. So, it's
7:40:25continuously trying to learn and get
7:40:27better at this script. I'm actually
7:40:28going to stop this and I'm going to try
7:40:30to course correct it a little bit. So,
7:40:31I'm going to go into plan mode. I'm
7:40:32going to say,
7:40:34so a couple issues from the first couple
7:40:36runs. The first thing is that it was
7:40:38hitting the like button like four times.
7:40:40So, it ended up not even liking the
7:40:42post. The second thing is once you
7:40:44switch to the wins channel, try to
7:40:46filter the posts by newest. So, there's
7:40:48a little kind of sandwich bar menu
7:40:51option near the channel filter. And that
7:40:54will allow you to change the view to
7:40:56newest. And then you can go through and
7:40:58make sure that all of the posts are
7:41:00liked. You can tell that they are liked
7:41:02because the thumbs up icon on the post
7:41:04will be yellow rather than gray.
7:41:08All right. So, we got the browser open.
7:41:13Nice. It was just able to switch to
7:41:15newest.
7:41:17And now it's going through and liking.
7:41:22It skipped over that one because it
7:41:23already liked it.
7:41:30Yeah, this is awesome. This seems to be
7:41:31doing the trick.
7:41:35It's liking my own posts, which is a
7:41:36little bit eh, but besides that, it's
7:41:39actually working. And it's skipping over
7:41:40the ones that are already yellow, which
7:41:41is great. It even just went to the next
7:41:43page, and it's still liking. So, this
7:41:45would go through and like every single
7:41:47post in the wins channel. So, I'm going
7:41:50to shut that down now just because I
7:41:51don't want it to go ahead and do all
7:41:52that right now. But, that worked and it
7:41:54took what, maybe four or five
7:41:56iterations. And now what I would do is I
7:41:58would build a skill out of that. So, I
7:41:59could just say, "Hey, run the school
7:42:01like skill." So anyways, I know that
7:42:03that was kind of a simple use case, but
7:42:04the point I was trying to prove is that
7:42:06yes, you can automate stuff in areas
7:42:08where you need to be logged in. And also
7:42:10like school is one of the least
7:42:12automation friendly platforms ever. So
7:42:16there's so much stuff that I do in there
7:42:17that's very repetitive. And now I could
7:42:20build out different bots and different
7:42:21scripts and skills in order to help me
7:42:23manage some of that, you know,
7:42:25repetitive stuff. Creating pages,
7:42:27switching up links, reorganizing the
7:42:28community, stuff like that.
7:42:32All right, so we're really flying
7:42:33through this course. Unfortunately, now
7:42:35we have to talk about some stuff that
7:42:36might not seem as fun. So, we're about
7:42:38to hop into some things that are a
7:42:39little bit more complex. Like I said,
7:42:41kind of boring, but really important to
7:42:43understand. So, we're going to get into
7:42:44some stuff with permissions, context
7:42:46management, GitHub, and work trees.
7:42:48Okay, welcome back to the fun stuff, the
7:42:51slides. You guys can obviously tell I
7:42:53enjoy the slides because I'm sitting
7:42:54down and normally I am standing up. So,
7:42:57let's get into it. So, we're going to
7:42:59talk about context management a little
7:43:00bit and then some GitHub stuff and some
7:43:03Git work trees stuff. So, we've talked
7:43:05about context management a lot. We've
7:43:07talked about cloudmd files a lot, but we
7:43:09haven't really spent a lot of time on
7:43:10like strategies to basically make sure
7:43:13that the window doesn't fill up super
7:43:14quick and that we keep the quality of
7:43:16our executive assistant high or the
7:43:18quality of our project high. So, let's
7:43:21get on into this one. The first tip, of
7:43:23course, is cloudmd. keep it lean and
7:43:26focused because it is injected at the
7:43:28beginning of every conversation. This is
7:43:30really really really important like
7:43:32this. Cutting down your clawb file from
7:43:34a couple hundred lines, you know, 5 6
7:43:36700 to 100 will make a huge difference
7:43:39right off the bat. So what you do is you
7:43:43treat this as the never forget rules.
7:43:45You treat this as the source of truth
7:43:47that Claude Code needs to know every
7:43:49single time before it messages you
7:43:51because there's a lot of things that it
7:43:52it's nice to have but you don't need to
7:43:54have. So all of your detailed reference
7:43:56docs that are usually the nice to haves,
7:43:59put those somewhere else. As long as
7:44:01Claude Code knows where to find them,
7:44:03you're set. So
7:44:07this was the example of my executive
7:44:08assistant, right? You are Nate's
7:44:10executive assistant. here are full
7:44:11details about me or my work or my team
7:44:14or my current priorities. That's one
7:44:16example. You could also do things like
7:44:17this. So, you have the Google Workspace
7:44:20CLI. You're authenticated as Nate. Um,
7:44:23now if you need actual information about
7:44:26how to use this, then just look up this
7:44:28doc. So, if you can't do it on your own
7:44:29and you're running into issues, then you
7:44:31know, you can go right here in the
7:44:32references folder to a markdown file
7:44:34called GWS CLI reference and you can
7:44:37find all the other operations that you
7:44:39need. So super simple. You do the same
7:44:41thing with your skills. You do the same
7:44:43thing with other rules, stuff like that.
7:44:46The next one is to use slashclear and
7:44:48slash compact strategically. So we've
7:44:51talked about this before, right? So use
7:44:52slashclear to wipe history after big
7:44:54tasks or if you're completely switching
7:44:56topics, but before you clear, what you
7:44:59can do is you can make handoff docs. So
7:45:02you can literally have it, this is
7:45:03essentially compacting, create a doc
7:45:05about what you were working on, whether
7:45:06that's going to be just for the short
7:45:07term, a temporary file or if it's going
7:45:09to be a long-term memory file and then
7:45:11when you move over to a new session,
7:45:13it's almost like you didn't even really
7:45:14move to a new session because all the
7:45:16information is right there, which is
7:45:17basically what compacting is, right? It
7:45:19summarizes and retains the essential
7:45:21information midsession so that you can
7:45:23keep certain items and you can keep
7:45:24certain conversation history.
7:45:27The other thing is to visualize your
7:45:28context a lot. So slashcontext really
7:45:31powerful tool. I've showed that a few
7:45:33times and you can see what's eating up
7:45:34tokens. And one thing that I realized
7:45:37when I was doing this is that MCP
7:45:38servers take up so many tokens and
7:45:41sometimes it's just not worth it. So I
7:45:42think that's another tip that I have on
7:45:44this slide, next slide. So I won't dive
7:45:46into that too much, but this allows you
7:45:47to identify waste and delete things that
7:45:49you're not using. If you got something
7:45:51in your context of that project that's
7:45:53taking up, I don't know, let's just say
7:45:55a percent or two of context every every
7:45:57time, get rid of it. if unless you're
7:46:00using it, right? Like it's the same
7:46:01thing of looking at your budget, like
7:46:03what's coming out of your account every
7:46:04month. What subscriptions? If you're not
7:46:06using the subscription, don't keep
7:46:08paying for it. Get rid of it. So, this
7:46:10is what it would look like, right? You
7:46:11do slashcontext. You can see everything
7:46:13takes up tokens. You can see in this
7:46:17project when I have a brand new session,
7:46:19I'm already at 24,000 tokens. So, that's
7:46:22over a tenth. I have system prompt,
7:46:24system tools, MCP tools, uh, custom
7:46:27agents, memory files, skills, all this
7:46:30stuff. Here's an example of just Chrome
7:46:33dev tools. Every single tool takes up
7:46:36tokens because every tool has a
7:46:38description of what it does so that
7:46:40Cloud Code could see that description
7:46:41and grab it. And that's where these
7:46:44tokens come from. We also have custom
7:46:46agents. Same thing. We have memory
7:46:48files. We have skills. They all have
7:46:50tokens that they take up natively. So
7:46:54this is how you can eliminate that
7:46:55waste.
7:46:57Next we have plan upfront and work in
7:47:00phases. So create a dedicated plan saved
7:47:03to a file and then work off that plan.
7:47:06Now let's say we have a we have a plan
7:47:08for a app and the app has five stages.
7:47:11Let me work with cloud code. Let's get
7:47:13us that plan doc. Okay cool. Save the
7:47:15plan doc. Now I can I can clear I can
7:47:17open up a new session and say hey look
7:47:20at this plan doc this is what we're
7:47:21doing you're on phase one help me do
7:47:23this like you know what you know what I
7:47:24mean you have that context now without
7:47:26having to stay in the same session. So
7:47:28that's one tip to do something like
7:47:29that.
7:47:31Limit your open sessions. So it's really
7:47:34cool that you can do you know like 10 to
7:47:3515 sessions at once. I tend to not do
7:47:38that anymore. I I used to for a while
7:47:40because I thought it was so cool because
7:47:42what happens is you can't juggle and
7:47:44understand how much context is going on
7:47:47in that many sessions. Like you could,
7:47:49but it's harder to
7:47:51be aware of how much context you've
7:47:53used. And the the issue there is that if
7:47:55you're not aware, you might run into
7:47:58context rot, right? And if you run into
7:48:00context rot, that means that your
7:48:01outputs are probably worse or they have
7:48:03to be double checked by a human more
7:48:05likely. And if you're running that many
7:48:07sessions, it's easier to lose track and
7:48:09it's easier to just think, "Oh, it's
7:48:10good. AI did it. We're good." So that's
7:48:13why I would not recommend doing any more
7:48:14than three to four parallel sessions,
7:48:16keeping actual tabs on what's going on.
7:48:20Sub agents for isolated tasks. So
7:48:23delegating fileheavy investigations to
7:48:26sub agents. These sub agents run in
7:48:29separate context. And there's, you know,
7:48:31a full video in this course about sub
7:48:33agents, but basically the idea of if you
7:48:35have tons and tons of data, so like
7:48:36let's say you're working in a main
7:48:38session and you have a full YouTube
7:48:40transcript, a 15-minute YouTube video
7:48:42transcript that you want to read in the
7:48:43session, don't have your main session do
7:48:46it. Say, "Hey, sub agent, read this.
7:48:48Give me the key highlights." And that
7:48:50way, the main sessions context doesn't
7:48:52get polluted with all of that non
7:48:54nonsense that it doesn't need.
7:48:58shared memory files for multi- aents.
7:49:00So, this is really cool because you can
7:49:02have multiple agents working across the
7:49:04same plan doc or task list. Now, we do
7:49:08have agent teams, which once again, this
7:49:09course has a video about, but if you
7:49:11want to basically Frankenstein your own
7:49:12way and have more control, you can have
7:49:14multiple agents look at and work off of
7:49:17the same file. The thing you have to be
7:49:19careful of though is if they're working
7:49:21in parallel, you run the risk of having
7:49:23them overwrite each other. So as long as
7:49:25you can manage that and have that sort
7:49:28of like persistent memory across
7:49:29different agents without them
7:49:30overwriting, then it is a very powerful
7:49:32technique to potentially save yourself a
7:49:34lot of tokens.
7:49:37We also have local semantic search over
7:49:39GP. So if you have large large code
7:49:41bases or projects, GP, which is one of
7:49:44the the tools we learned about, is going
7:49:46to search through things in a much less
7:49:47efficient way than using like actual
7:49:49semantic search because that's kind of
7:49:51more of like the, you know, the meaning
7:49:53search rather than just like keyword
7:49:54matching. So benchmarks have actually
7:49:56shown that you can get a 97% reduction
7:49:59in tokens when you're doing local
7:50:01semantic searches like that. So that's
7:50:03just something to keep in mind. I have
7:50:04not yet truthfully gotten to the point
7:50:06where I need to do this, but it is
7:50:08something to keep in mind.
7:50:10Okay, now let's move on a little bit to
7:50:13GitHub and work chase. So this is really
7:50:16important, right? Because you need to
7:50:18understand where things live, how you
7:50:21track version control, how you can
7:50:22collaborate, how you can move local code
7:50:24bases between devices and then work
7:50:28trees. So basically making sure that
7:50:29you're not breaking anything. So these
7:50:31are two really important concepts to
7:50:33understand. And before we get into all
7:50:34this technical stuff that might seem
7:50:35overwhelming, just know this.
7:50:39AI knows GitHub and Work Trees so well.
7:50:42They know GitHub and Work Trees so much
7:50:44better than you ever will. Unless you're
7:50:46maybe from a developer background, they
7:50:49certainly know GitHub and Work Tries
7:50:50better than I ever will. So, all you
7:50:52have to do is be very strategic and
7:50:54smart about what you want and what you
7:50:55don't want. Meaning, if you want to save
7:50:57your work, if you want to roll back to a
7:51:00different version, if you want to share
7:51:01this with someone else, just use your
7:51:03natural language and cloud code will
7:51:05help you do that with GitHub or with
7:51:06work trees. It's super super simple. It
7:51:09just kind of goes back to what I talked
7:51:10about at the very very beginning of this
7:51:11course where I was like be genuinely
7:51:13curious. That's like the number one
7:51:15mindset shift. Okay. So there is a
7:51:18difference though between Git and
7:51:19GitHub.
7:51:21Git is a version control system. It is a
7:51:23tool that tracks every change that you
7:51:25make to your projects and this can be
7:51:26locally. Think of it like a time machine
7:51:28for your project. So every time you save
7:51:30a meaningful change, which is something
7:51:31that they call a commit, Git takes a
7:51:34snapshot of your entire project at that
7:51:36moment. So later if something breaks you
7:51:38can effectively rewind. So it's like
7:51:40you're getting a checkpoint, right?
7:51:42So a few concepts, a repo or a
7:51:46repository that's basically a folder
7:51:48that git is watching. It contains all
7:51:50the files plus the full history of every
7:51:52change ever made and it's like you know
7:51:54a bank account for your code. It has
7:51:56everything, all the transactions,
7:51:58everything. Now what is a commit? This
7:52:01is a snapshot of your project at a
7:52:03specific point in time. So each commit
7:52:06includes what you're describing or sorry
7:52:08it includes what you've changed and
7:52:09describes why and basically like hitting
7:52:11save on a word doc or a video game or a
7:52:14video file
7:52:16branch. This is a separate copy of your
7:52:19project where you can make changes
7:52:20without affecting the main version. So
7:52:23almost like you clone it, you change
7:52:25things and then if you like it, you can
7:52:26merge it back in. The default branch is
7:52:29usually called main. And when you create
7:52:31a new branch like feature login, you're
7:52:33essentially saying, "Let me try
7:52:35something without risking the stable
7:52:36version." So it's almost like the way
7:52:38that you have, you know,
7:52:41workflows that are published and
7:52:42workflows that are in draft mode or
7:52:44inactive. And you would never want to
7:52:46make live changes to the published
7:52:48version. You would change things in the
7:52:49test environment. You would change
7:52:50things as a draft. And then once you
7:52:52like it, you push that to prod or you
7:52:54push that to main or whatever different
7:52:56terminology you have. It's that same
7:52:57concept, right?
7:53:00Um, and that's where push comes in,
7:53:02right? Uploading your local commits to
7:53:03the online version of your repo. So,
7:53:05syncing to cloud. A pull is downloading
7:53:08the latest changes from the online repo
7:53:10to your local machine. So, you're
7:53:11pulling it in. And then a merge is
7:53:13combining the changes from one branch
7:53:15into another. Once your feature is done
7:53:16and tested, you can merge it back into
7:53:18the main. So, that's how we do
7:53:20everything here very, very safely.
7:53:23GitHub. So, GitHub is basically the
7:53:25cloud service built on top of git. So
7:53:28while Git runs on our local computer and
7:53:30tracks our files, GitHub hosts them
7:53:33online so that you can access them from
7:53:35anywhere. You can back them up, you can
7:53:37collaborate with others, do whatever you
7:53:39need. And that's how you can maybe take
7:53:40your executive assistant that's right
7:53:42now locally, push to GitHub, and then if
7:53:44you have a laptop and you want to go on
7:53:46a vacation, you can bring your laptop
7:53:48with you and pull in the repo. So you
7:53:50still have your executive assistant
7:53:51there. So think of it like this. Git is
7:53:53the engine that tracks your changes.
7:53:56GitHub is the parking garage in the
7:53:58cloud where your project lives so that
7:53:59it's safe and accessible.
7:54:03What does GitHub what does GitHub
7:54:05actually give you? Well, cloud backup,
7:54:08collaboration, pull requests, and
7:54:11version history. So, cloud backup is
7:54:13pretty self-explanatory. Um,
7:54:16collaboration means that multiple people
7:54:17can work on the same project without
7:54:19overwriting each other's work.
7:54:21Pull requests means that you can propose
7:54:23changes and have them reviewed before
7:54:25you merge them into the main project.
7:54:27And so that's how you can have multiple
7:54:29engineers working on the same GitHub
7:54:30repo, all pushing different feature
7:54:33requests and, you know, working on their
7:54:34own things and merging it all back into
7:54:36the main environment safely. And then
7:54:39you can also see version history, which
7:54:41would basically just mean, let's say you
7:54:43have a project, you push 10 different
7:54:46commits. You can see each commit. And if
7:54:48you needed to roll back to a certain
7:54:50commit, you could. You could see who
7:54:51made the commit. You could see what was
7:54:53the purpose of that commit. You can just
7:54:54see everything, right?
7:54:57Okay. So the basic workflow would be
7:55:00create a GitHub repository for your
7:55:03project. clone a copy of it to your
7:55:05local computer. Create a branch for your
7:55:07new feature or fix. Make changes and
7:55:09then commit. Push the branch to GitHub.
7:55:12Open a pull request to propose merging
7:55:14your changes into the main and then you
7:55:16review, approve, and merge. That's
7:55:18typically the way that it works when you
7:55:19have a team working on a project. So,
7:55:22this is exactly how professional
7:55:23developers and teams work and it's the
7:55:25same workflow cloud code uses behind the
7:55:26scenes when managing the project. Once
7:55:28again, cloud code understands these best
7:55:30practices and when you need to save
7:55:32things or roll back, it will help you do
7:55:34that.
7:55:35So now that you understand all that
7:55:37stuff, work trees should probably click
7:55:40a little bit better. So a work tree lets
7:55:42you check out multiple branches from the
7:55:44same repo at the same time, each on its
7:55:46own folder on your computer.
7:55:49So the problem is that normally in git,
7:55:52you can only work on one branch at a
7:55:53time within that folder. So if you're
7:55:55building out a new feature and suddenly
7:55:56you need to go to a different branch,
7:55:58you would have to either stop what
7:55:59you're doing and save it or you would
7:56:00have to just like commit it and it might
7:56:02not be finished yet. So it's risky. So
7:56:05this problem is solved with work trees
7:56:07because you can basically just create a
7:56:08second working directory just to work on
7:56:10that branch or just to work on your own
7:56:12fix. So your original work basically
7:56:14just stays there and you just open up a
7:56:16separate one and they they're isolated.
7:56:18They don't touch each other, right?
7:56:21So, Cloud Code has built-in native
7:56:24support for git work using a command.
7:56:27The good news is you don't ever have to
7:56:29really run this if you don't want. Cloud
7:56:30code can do it, but you could also, if
7:56:32you want full control, you could start
7:56:33up your own work tree. This means you
7:56:36can run multiple cloud code sessions at
7:56:37the same time, each working on a
7:56:39completely different task without any of
7:56:41them interfering with each other. So, it
7:56:44is very, very safe. Okay, great job
7:56:46getting through the stuff that I thought
7:56:47was the most boring about learning cloud
7:56:49code. And now let's just kind of get
7:56:51back into some more fun stuff. So I'm
7:56:52going to go over some hacks with you
7:56:54guys and then just show you some cool
7:56:55things that you can do with Cloud Code.
7:56:57And I think it's pretty fun. So let's
7:56:59get into it.
7:57:03These are the Cloud Code hacks that took
7:57:04me from a complete beginner to
7:57:05mass-producing workflows and building
7:57:07websites, apps, and AI agents in real
7:57:09time. So today we're going to be going
7:57:10from beginner hacks all the way to
7:57:12advanced power user stuff. And the best
7:57:14ones are saved at the end. All right, so
7:57:15starting off with our beginner hacks.
7:57:17Number one is to run /init on every
7:57:19project. So if you already got an
7:57:21existing project with files already in
7:57:22there, the first thing you should do is
7:57:24open it up and type /init. Cloud code
7:57:26will then scan your entire codebase,
7:57:28your folders, your files, and it will
7:57:30generate a cloud. MD file, which is
7:57:32basically like a cheat sheet for that
7:57:33project. It'll map out your
7:57:35architecture, your conventions, and any
7:57:36key files that you have in there. So
7:57:38instead of having to reexplain your
7:57:39project every session, Cloud will
7:57:40basically just contextualize everything
7:57:42and initialize everything and know what
7:57:44you're working with. And if you're
7:57:45starting a new project from scratch,
7:57:46then you can have cloud code help you
7:57:48create that cloudmd file yourself just
7:57:50by explaining what's the goal of this
7:57:52project, what text stack you want to
7:57:53use, or any rules or key folders and
7:57:55files. All right, number two is to set
7:57:57up a status line. So if you're working
7:57:59in the terminal, you can type / status
7:58:01line and tell cloud code what you want
7:58:03to see, your model, your context
7:58:04percentage, cost, whatever. It basically
7:58:06generates a little script that sits at
7:58:08the bottom of the terminal. So as you're
7:58:09talking every single time, you can just
7:58:11see that status line. It's just kind of
7:58:13like a mini dashboard for your session.
7:58:14So, it's really helpful to always be
7:58:15able to see how much context you have
7:58:17left so you can avoid context rot. Hack
7:58:19number three is using voice input. So,
7:58:21cloud code just shipped a native/v voice
7:58:23command, which means you can literally
7:58:24just talk to your terminal and have it
7:58:26code for you now. So, it's still rolling
7:58:27out. It will be out for everyone soon,
7:58:29but another good hack would just be to
7:58:31use an app to actually be able to voice
7:58:32dictate anywhere. So, if you want to see
7:58:34the tool that I use, you can check out
7:58:35the description. Now, I can just talk
7:58:36and words will appear anywhere. Hack
7:58:38number four is to keep your context
7:58:39small. So don't dump your entire
7:58:41codebase into a conversation. Only give
7:58:43Claude what it needs for the current
7:58:44task. So try to break big problems into
7:58:47small focused steps. The less noise in
7:58:48the context window, the better Claude
7:58:50performs. It's simple, but a lot of
7:58:52people ignore this. Hack number five is
7:58:53to use /context to find your token
7:58:56bloat. So if you do /context, you'll see
7:58:58exactly what's eating your tokens.
7:59:00Whether that's system prompts or file
7:59:01contents, MCP servers, whatever it is,
7:59:04all of that gets broken down into
7:59:05percentages. So if your session feels a
7:59:07little bloated, you can actually
7:59:08investigate it, diagnose where the
7:59:10problem is, and then restructure. Hack
7:59:12number six is to compact at 60%. And
7:59:14also clear between tasks. So when your
7:59:17context hits around 60%, then type
7:59:19/compact and cloud code will compress
7:59:21your conversation history so you can
7:59:23keep going without losing important
7:59:24stuff. And something interesting is that
7:59:25you can actually do a /compact, but you
7:59:28can tell it to keep certain things like,
7:59:29hey, / compact, but keep all of the API
7:59:32integration decisions and database
7:59:33schema. So Cloud will automatically
7:59:35shrink everything down and preserve the
7:59:36stuff that you need to keep in there.
7:59:38And if you're actually going to switch
7:59:39to a completely different task and you
7:59:41don't need that conversation history,
7:59:42then use slashclear to just wipe the
7:59:44slate clean and you're starting from a
7:59:45new conversation. But luckily, you still
7:59:47have your cloud. Mmd, you still have all
7:59:48the other files. So it's not like you're
7:59:50actually starting from scratch. So hack
7:59:51number seven is to always start in plan
7:59:53mode. So that means you can hit shift
7:59:54tab to cycle between modes or just
7:59:56choose it manually. And once Claude's in
7:59:58plan mode, it can still read, it can
7:59:59still research, but it won't actually
8:00:01change anything. So Claude will outline
8:00:02the steps. It will ask clarifying
8:00:04questions and it will map out the
8:00:05approach before writing a single line of
8:00:07code which has been shown to improve the
8:00:08quality. Now once you like the plan you
8:00:10switch out of plan mode. Tell it to
8:00:11execute and this alone will dramatically
8:00:13reduce how many times that you have to
8:00:14go back and correct Claude. Hack number
8:00:16eight we have to treat Claude like a
8:00:18junior developer which means don't
8:00:19always give it direct commands like
8:00:21write me a function that does X but try
8:00:23to understand how you can give it
8:00:24problems. So saying, "How should we
8:00:26handle growth tracking?" And let it
8:00:27think through the approach because when
8:00:29it makes its own assumptions and it
8:00:30thinks through decisions, you can ask it
8:00:32to explain those. And this has also been
8:00:33shown to get better outputs when Claude
8:00:35reasons through the problem first. So
8:00:37it's like plan mode, but now you're
8:00:38having it think a little bit deeper.
8:00:40Okay. Hack number nine is to make Claude
8:00:42ask questions. So a lot of times in plan
8:00:43mode, it will do this natively, but you
8:00:45can actually tell it to invoke its ask
8:00:47user question tool. You can tell it
8:00:49continuously ask me questions until
8:00:50you're 95% confident that you understand
8:00:52exactly what I need and exactly what you
8:00:54need to do. And once again, this
8:00:56alignment helps you from having to go
8:00:58back and forth with, you know, three or
8:00:59four rounds of revisions. All right,
8:01:01hack number 10 is build self-checking
8:01:03into the to-do lists. So, you know how
8:01:05Cloud makes a to-do list when it starts
8:01:07building? Well, you can actually build
8:01:08verification steps right into that list.
8:01:10So, let's say one to-do is to build the
8:01:12website. The very next to-do could be
8:01:14take a screenshot of the website and
8:01:16check that everything looks right. And
8:01:17then maybe the next step is to open
8:01:18Chrome DevTools to use the browser and
8:01:21make sure that there are no actual
8:01:22errors in functionality. So you're now
8:01:24baking quality checks directly into the
8:01:25execution plan. So Claude isn't just
8:01:27building stuff and handing it to you for
8:01:28feedback, but now it's building
8:01:30something, checking it, making sure
8:01:31everything's good, and then getting your
8:01:32feedback. And another cool thing that I
8:01:34like to do here is say, don't move on to
8:01:36your next to-do until you're 95%
8:01:38confident that that to-do is good.
8:01:40Because it's AI, it's really hard to
8:01:42oneshot what you're looking for, but
8:01:44you'd rather have it one shot 90% of the
8:01:46way there rather than one shot 60 or
8:01:4865%. Okay, so those were our beginner
8:01:50hacks. Now, let's step it up a little
8:01:51bit. These next ones are for the people
8:01:53who are already kind of using Cloud Code
8:01:54a little bit and want to move faster.
8:01:56All right, so hack number 11 is to
8:01:57deploy sub agents for parallel work. Try
8:02:00telling the main session to use sub
8:02:01aents in your prompt when you're working
8:02:03on complex problems. Cloud will spin up
8:02:05isolated sub aents that each have their
8:02:06own context window. They can each be
8:02:08using their own model and each agent
8:02:09works in parallel which means the main
8:02:11thread stays clean while the sub aents
8:02:13go do research, write tests or explore
8:02:15different approaches. When they're done,
8:02:17they all report back to that main agent
8:02:18with their findings. So it's like having
8:02:20a team of developers instead of just
8:02:21having one. And you can even pair this
8:02:23with the model hack for cheaper tokens,
8:02:25which means you can have all the sub
8:02:26aents running on Haiku for simpler stuff
8:02:28and your main thread can stay on Opus.
8:02:30All right, hack number 12 is to build
8:02:31custom skills. This means you can create
8:02:33reusable prompt files in your
8:02:35do.cloud/skills directory. So for
8:02:37example, you can have one skill called
8:02:38techdebt.md which tells claude exactly
8:02:40how to scan for technical debt. Or you
8:02:42can have one called code review.mmd
8:02:44which knows exactly how to review your
8:02:46codebase and then all you have to do is
8:02:47invoke that skill in natural language or
8:02:49just use the slash command directly and
8:02:50it will run that entire workflow
8:02:52consistently every single time. You can
8:02:54even commit them to GitHub and your
8:02:55whole team can instantly use them as
8:02:57well. So you can automate your actual
8:02:58SOPs. All right, so hack number 13 is
8:03:00something that I alluded to a little bit
8:03:01earlier, but that's basically just using
8:03:03Haiku for sub aents because you can set
8:03:05the model for the sub aents that you
8:03:06spin up. When you have simple tasks or
8:03:08processing a large amount of data, then
8:03:10use Haiku. It's way cheaper and it still
8:03:12gets the job done. Specifically, if you
8:03:13need a sub agent to go scrape a ton of
8:03:15different articles, read hundreds of
8:03:16thousands of tokens, and then just give
8:03:18Opus, give your main agent just a small
8:03:20summary or the key highlights. It just
8:03:22doesn't really make sense to have such a
8:03:23heavy and expensive model reading
8:03:25hundreds of thousands of tokens if it
8:03:26just needs a few bits of information.
8:03:28And if you do this right, it can really
8:03:29keep your cost down without sacrificing
8:03:31quality where it matters. All right,
8:03:32hack number 14 is to constantly be
8:03:34refreshing your claw.md file. Once
8:03:36there's a new discovery about your
8:03:37project, update the claw.md. Once you've
8:03:39made some new skills, update the
8:03:41cloud.mmd. You want Claude to be logging
8:03:43new patterns, new gotchas, and any new
8:03:45conventions that it discovered during
8:03:46your session. So next time that you
8:03:48start it up, it already knows all of
8:03:49this. This will help prevent repeat
8:03:50mistakes, and it will make Claude
8:03:52smarter about you, your business, your
8:03:54project, all that kind of stuff over
8:03:55time. But here's the catch. You don't
8:03:57want to let it bloat because the claw.md
8:03:59file is basically the system prompt and
8:04:01it gets loaded into every single
8:04:02conversation and everything in there is
8:04:04going to eat up at your context window.
8:04:06So I try to keep mine simple and only
8:04:08put the most important information in
8:04:09there. I like to keep it between 150 and
8:04:11200 lines max. If it starts getting
8:04:13longer than that, then it's time to trim
8:04:14down some things. Which leads perfectly
8:04:16into the next hack, number 15, which is
8:04:18to have claw.md route to other files.
8:04:21Because it potentially eats so many
8:04:22tokens, you want to keep it lean, but
8:04:24you do have a lot of information in
8:04:25there. But what's cool is you can route
8:04:26it to different places. So you can have
8:04:28it link out to separate files for stuff
8:04:30like style guides or business context or
8:04:32reference docs. Just point to those
8:04:33files in the cloudmd so claude knows
8:04:35exactly where to look and then you're
8:04:37also not wasting tokens on information
8:04:38that it doesn't always need because in
8:04:40its system prompt it doesn't need to
8:04:41know the exact status of a certain
8:04:42project but it does need to know exactly
8:04:44where to go look to find that
8:04:45information. Hack number 16 is to exit
8:04:48early and reask. So if you notice that
8:04:50cloud starts going down the wrong path
8:04:52don't just wait for it to finish. hit
8:04:53escape, correct course, and then
8:04:55reprompt. Every token that it spends
8:04:56going the wrong direction is just wasted
8:04:58context. So steer tight and steer early.
8:05:01At the end of the day, it is AI. Hack
8:05:02number 17 is to challenge outputs
8:05:04aggressively. So if Claude gives you
8:05:06something that's just okay, push back.
8:05:08Say scrap that, do a more elegant
8:05:10version or this isn't good enough, try
8:05:11again with a completely different
8:05:12approach. Claude will often give you a
8:05:14dramatically better output on the second
8:05:15try when you set a higher bar and now it
8:05:17knows what to not do. The key is once it
8:05:19comes back with something better, tell
8:05:21it to update itself. whether that be the
8:05:22skill or the claw.md so it doesn't make
8:05:24that sort of mistake again. Hack number
8:05:2618 is to use slashre for quick undos. So
8:05:29if you make a wrong turn just try using
8:05:30slre and claude will roll back to a
8:05:32previous point in the conversation
8:05:34without you having to start over. So
8:05:35it's super quick, super clean. Hack
8:05:37number 19 is using hooks for
8:05:39notifications. So if you type hooks, you
8:05:41can set up a notification hook or you
8:05:42can just have Claude code do this for
8:05:43you in completely natural language. So,
8:05:45for example, what I like to do is when I
8:05:47have Claude finish up a session or
8:05:49finish a chat, it sends me an actual
8:05:50noise notification because now I can
8:05:52work on something else on my computer or
8:05:54I can literally have 15 different
8:05:55sessions of Claude code running and if I
8:05:57hear that noise, I know that one of them
8:05:58is done and needs some more input from
8:06:00me. Hack number 20 is using screenshots.
8:06:02Just remember that Claude can actually
8:06:04see and this is a huge unlock which
8:06:05means you can feed it error messages
8:06:07which means you can feed it, you know,
8:06:08inspiration websites. You can also do a
8:06:10really cool selfch check loop where you
8:06:12can say things like take a screenshot of
8:06:13the website and tell me if the layout
8:06:15looks right. And it will literally
8:06:16screenshot it, analyze it visually, and
8:06:18tell you what's off. So, if you remember
8:06:19one of the hacks from earlier where I
8:06:20said to have it check itself, when I'm
8:06:21building websites, I basically have it
8:06:23design the website, screenshot, and then
8:06:25implement new changes, and then do that
8:06:26again. And so, it does like three passes
8:06:28of building and screenshots before it
8:06:30even gives me V1. And in that flow, the
8:06:32V1 that it gives me is so much better
8:06:34than a V1 that it used to give me. Hack
8:06:36number 21 is to use Chrome Dev Tools. So
8:06:38Claude can open a browser, it can
8:06:39interact with an app, it can check the
8:06:41functionality of things. And so it's
8:06:42kind of like the screenshot loop, but
8:06:44instead of for like websites and design,
8:06:46it's for actual functionality of like
8:06:48apps and buttons. This is huge for
8:06:49front-end work. So definitely give it a
8:06:51try. But this also means that it can do
8:06:52things like filling out forms and
8:06:53potentially like recaps and stuff. But
8:06:55this is also huge because if there's not
8:06:57an explicit API somewhere, it can go in
8:06:59and manually do things. I think that it
8:07:00could also solve captas, but it's
8:07:02probably better if you're already signed
8:07:03in somewhere and all it has to do is
8:07:05navigate, click buttons, fill out
8:07:07things. Hack number 22 is to clone
8:07:09inspiration sites. So, you can take
8:07:10screenshots of sites that you really
8:07:12like and feed it to Claude and say,
8:07:14"Make it look like this." Claude will
8:07:15recreate the design patterns without
8:07:16making it look like generic AI slot. And
8:07:18this is huge for front-end quality
8:07:20because you could also use the site as
8:07:22inspiration by taking some of the actual
8:07:23like HTML styling and feed that into
8:07:25Claude, too. So yes, Claude could
8:07:27essentially clone a website, but what
8:07:28you want to do is take that as kind of a
8:07:30template and give it your own touch.
8:07:32Okay, so now we're going to move on to
8:07:33some more advanced stuff. These hacks
8:07:35are for people who really want to push
8:07:36cloud code to its limits. So let's go.
8:07:38All right, hack number 23 is to run
8:07:40parallel sessions with git work trees.
8:07:42Normally when you're working on a
8:07:43project, you've got it in one folder
8:07:44with all your files in it. So if you
8:07:46want to run two different sessions in
8:07:47the same folder at the same time, they
8:07:49might overwrite each other's work, and
8:07:50that's where work trees come in. So
8:07:51think of workree like basically making a
8:07:53parallel copy of your project except
8:07:55it's way more efficient than actually
8:07:57copying the folder. You just type in
8:07:58claude-workree
8:08:00and then that feature name. So claude
8:08:01will then create an isolated workspace
8:08:03on its own branch. You could then open
8:08:04up another terminal and type in the same
8:08:06thing with a different feature name and
8:08:07it will open up a different branch. So
8:08:09now you can be working on the same
8:08:10project at the same time without having
8:08:12those coding agents step on each other.
8:08:14You could have three, four or five of
8:08:16these things going at once. And when
8:08:17you're done, you can have them just
8:08:18merge the branches back together just
8:08:20like you would with any other git
8:08:21branch. So all the work can save back to
8:08:23the main project once again without
8:08:25overwriting each other's files. All
8:08:26right, hack number 24 is to use API
8:08:28endpoints instead of MCP servers
8:08:30depending on the situation. But here's
8:08:32what I mean by that. MCP servers are
8:08:34great because you can look at all the
8:08:35different tools and execute any of them,
8:08:36but they load their entire tool
8:08:38definitions into the context window. So
8:08:40if you're tight on tokens, sometimes
8:08:41it's better to just use direct API
8:08:43endpoints instead. So, for example,
8:08:45let's say you're using Notion and you
8:08:46only actually need to be able to read
8:08:48one database. It makes no sense to show
8:08:50Claude how to do all of the other
8:08:52functions if for this specific project
8:08:53you only have to read one file. So,
8:08:55instead just hardcode in that endpoint
8:08:57and now you're saving tons and tons of
8:08:59tokens. All right, hack number 25 is to
8:09:01use /loop for recurring tasks. So, you
8:09:04can type, hey, every 5 minutes check in
8:09:05on the deployment. And Cloud will rerun
8:09:07that prompt in that same session every
8:09:09single 5 minutes unless you close out of
8:09:11that session. You can set it to monitor
8:09:12a PR, check error logs, pull a build,
8:09:14whatever. It runs in the background and
8:09:16it only interrupts you when something
8:09:17actually needs your attention. You can
8:09:19even set one-time reminders in natural
8:09:20language, like remind me at 3 p.m. to
8:09:22check in with the team on X. Now, the
8:09:24only caveat here is these actual loops
8:09:26will only last for 3 days. So, if you
8:09:28need a scheduled automation that's a
8:09:30little bit longer term, then you're
8:09:31going to want to use the desktop
8:09:32scheduled tasks. Although, the only
8:09:33difference here is every time one of
8:09:35those tasks spin up, it's in an
8:09:36individual session, so it doesn't have
8:09:38that context memory. Number 26 is to
8:09:41host on a VPS for always on sessions. So
8:09:43if you want to run cloud code on a
8:09:45remote server, it'll stay running even
8:09:46when your laptop is closed, which means
8:09:48you can SSH in whenever you need to
8:09:49interact, which means you can talk to it
8:09:51through Telegram anytime. This is
8:09:52perfect for longunning tasks where you
8:09:54don't want to babysit a local terminal.
8:09:56Hack number 27 is you can use remote
8:09:58control from your phone. So this is a
8:09:59pretty new feature, but Cloud Code now
8:10:01lets you control local sessions from
8:10:02your phone or any browser. You start a
8:10:04task at your local desk and then you
8:10:06walk away and you can keep steering it
8:10:07from your phone. Your code never
8:10:08actually leaves your local machine, but
8:10:10it's just the remote connection is on
8:10:12your phone. So, you can start something
8:10:13heavy, go grab a coffee, go on a walk,
8:10:15and you can keep building from your
8:10:16pocket. Hack number 28 is there's no SQL
8:10:19data analytics. So, you can connect CLI
8:10:21tools like Big Query's BQ tool to cloud
8:10:24code. And then you can just ask
8:10:25questions in plain English like, what
8:10:26were our top 10 revenue sources last
8:10:28quarter? And cloud will instantly
8:10:30translate that into the right query, run
8:10:31it, and then give you that answer. No
8:10:33SQL required. And this should work for
8:10:35any CLI based tool. Number 28 is
8:10:37Ultrathink. When you need clawed code to
8:10:39really think through a hard problem like
8:10:40architecture decisions, complex
8:10:42debugging, big refactors, or maybe it's
8:10:44just not giving you the right output
8:10:45after a couple prompts, try using
8:10:47Ultraink. You literally just type the
8:10:48word and it will go all colorful. And
8:10:50this means that it allocates the maximum
8:10:52thinking budget of around 32,000 tokens
8:10:54before Claude actually responds. So,
8:10:56don't always use this for a simple fix,
8:10:57but absolutely use it if you're making
8:10:59decisions that might affect the entire
8:11:00system or like I said, if after the
8:11:02first couple tries it's not giving you
8:11:04what you want. Hack 30 is to edit
8:11:05permissions for safe autonomy. A lot of
8:11:07people, including myself, have shown on
8:11:09videos using dangerously skip
8:11:10permissions to make sure that Claude can
8:11:12just run without asking for approval on
8:11:14every single step. And yeah, it's much
8:11:15faster, but it is called dangerously
8:11:17skip permissions for a reason. So, the
8:11:18smarter way to go about it is to go into
8:11:20your permissions and explicitly allow
8:11:21the commands that you know are safe and
8:11:23then explicitly deny anything that's
8:11:25destructive, like deletes or removes.
8:11:27And now you can actually get to the
8:11:28point where you have the same exact
8:11:30speed and autonomy of dangerously skip
8:11:32permissions without it being very
8:11:34dangerous. And anything in the deny list
8:11:36is going to take priority over anything
8:11:38in the allow list. Hack number 31 is to
8:11:40use agent teams. So remember how we
8:11:42talked about sub agents being able to
8:11:43run agents in parallel that have fresh
8:11:45context but can't talk to each other?
8:11:47Agent teams are like that but all of the
8:11:49agents can talk to each other. So it
8:11:50gets really really cool. They share a
8:11:52task list. They can communicate with
8:11:53each other and they can even assign each
8:11:55other work. and you can actually talk to
8:11:57each of those individual agents instead
8:11:59of just having to go through the main
8:12:00one and then the main one would
8:12:01communicate with sub aents. These are a
8:12:03little bit more expensive and they do
8:12:04run longer, but they will give you a
8:12:06much more cohesive output for a big
8:12:08project. Hack number 32 is context 7
8:12:11MCP. This one's a game changer. You can
8:12:13install the context 7 MCP server and
8:12:15then whenever you need information on
8:12:17current documentation just prompted to
8:12:19use that MCP server. The problem that it
8:12:20solves is that Claude's training data
8:12:22has a cut off, which means sometimes it
8:12:24might suggest functions or APIs that
8:12:26have been renamed or deprecated or just
8:12:28don't even exist anymore. So, Context 7
8:12:30fixes that because it has up to eight
8:12:32version specific technical documentation
8:12:34about live code examples from thousands
8:12:36of popular libraries that you probably
8:12:38need with a coding assistant like
8:12:39Nex.js, React, MongoDB, you name it. So,
8:12:43it's able to pull and read all current
8:12:45documentation and then inject it into
8:12:46the conversation before Claude actually
8:12:48starts writing any code. So, it's
8:12:50basically one command to install and
8:12:51from there all of your coding agents are
8:12:53working with much more up-to-date
8:12:54information and it's a huge quality
8:12:56improvement. All right, so I know that
8:12:57we covered a ton of information in this
8:12:58video. So, what I did is I threw all of
8:13:00this into a PDF resource guide so that
8:13:02you can just come back and reference
8:13:03them whenever you want. That's available
8:13:05completely for free inside of my free
8:13:07school community. The link for that is
8:13:08down in the description. But that's
8:13:09going to do it for this one. So, if you
8:13:10guys enjoyed or you learned something
8:13:11new, please give it a like. It
8:13:12definitely helps me out a ton. And as
8:13:14always, I appreciate you guys making it
8:13:15to the end of the video. I will see you
8:13:16all in the next one. Thanks, everyone.
8:13:20All right. So, here I am in cloud code
8:13:22in the terminal right over here. What
8:13:23you'll notice is that this agent is
8:13:25actually my cloud code agent. So, when I
8:13:27send off this message where I'm asking
8:13:28it to spin up two researcher agents, one
8:13:30to research Google, one to research
8:13:32OpenAI, and then create a visual
8:13:33diagram, you'll see that this agent
8:13:35actually gets that task. And now you can
8:13:37see that we have three agents actually
8:13:38running right now in parallel. This
8:13:40one's researching Google AI strengths
8:13:41and this one is researching OpenAI
8:13:43strengths. And this is the main one that
8:13:45delegated the tasks to those two
8:13:47researcher agents, which you can also
8:13:49see on this right hand side. That's
8:13:50exactly what's being ran right now. And
8:13:52if I wanted to open up another terminal,
8:13:54so I add another agent. It plops in
8:13:56another agent right here in which I can
8:13:58have a different conversation with. Give
8:13:59me a quick analysis of my YouTube
8:14:01comment section from the past couple
8:14:03videos I've made. So now I shoot that
8:14:04message off to this AI agent in Cloud
8:14:06Code, and it's going to pick up the task
8:14:08and start working on other things. As
8:14:10you can see now, this one is searching
8:14:12files and finding out what it needs to
8:14:13do. And you just heard that little
8:14:16notification noise, which means that
8:14:17this spun up a sub aent. And you can see
8:14:20that it was analyzing comments, but now
8:14:21it needs approval. So basically, what we
8:14:23have here is a visual way to see all of
8:14:25the different terminals and sub aents
8:14:27that we have working in cloud code. We
8:14:29get notifications and we get visual
8:14:30things like this that tell us what
8:14:32they're doing. So I can see that those
8:14:33sub aents on this lefth hand side are
8:14:35disappearing. So, I can hop back over to
8:14:37that first terminal and I'll be able to
8:14:38see that this first one is already done.
8:14:40And now the OpenAI one is still being
8:14:42worked on, which is this agent right
8:14:44over here. Okay. I don't know where
8:14:45she's off to, but you can see now that
8:14:47both of these agents are done. They're
8:14:49idle. They're done working. So, in the
8:14:51first terminal, we can see here's the
8:14:52visual. It came out clean. We have stuff
8:14:54about Google's edge, Google's gaps,
8:14:56OpenAI's edge, OpenAI's gaps. So, here's
8:14:58that actual visualization that it made.
8:15:00We have strengths, weaknesses for both
8:15:02Google and OpenAI. Sometimes AI
8:15:04generation can be notoriously bad on
8:15:06text, but this one looks pretty decent.
8:15:08Only a few errors I can see. And then
8:15:10coming over to the second terminal, we
8:15:11can see the comment analysis over the
8:15:13last two weeks. So we've got some pain
8:15:14points about cloud API costs, edited
8:15:16into cloud code, getting first client
8:15:18with zero social proof. We've got most
8:15:20requested videos. We've got notable
8:15:21patterns. And then we have recent video
8:15:23performance snapshot. But I'm sure you
8:15:25guys could tell that the point of this
8:15:26quick demo wasn't to actually show you
8:15:28the outputs. It was to show you what we
8:15:29were able to do here and how we could
8:15:31see these two different terminal agents
8:15:32spin up sub agents and we could keep
8:15:34tabs on them visually. So,
8:15:36unfortunately, I'm going to say goodbye
8:15:37to these agents so we can get on with
8:15:38the video and I'm going to kill these
8:15:40terminal sessions right here so I can
8:15:42actually explain to you guys how this
8:15:44works and what I think about it. Okay,
8:15:46so the way that we are doing this is we
8:15:47are using Visual Studio Code and that's
8:15:49where I've been using Claude Code as an
8:15:51extension. If you've been following my
8:15:52channel for a while, we're using an
8:15:54extension in VS Code called Pixel
8:15:56Agents, which basically turns your AI
8:15:58coding agents into animated pixel art
8:16:00characters in a virtual office that you
8:16:01can fully customize. So, I'll have a
8:16:03link to this page right here, and I will
8:16:05have a link to the GitHub repository in
8:16:07the description, but I'll also show you
8:16:09guys like exactly how you set this up.
8:16:10So, before we actually jump into that,
8:16:12real quick, just wanted to go over like
8:16:13what are we doing here, and what is
8:16:14Cloud Code because that's the little
8:16:16agents that we were actually spinning
8:16:18up. Well, Cloud Code is an AI in our
8:16:20terminal. It can write and edit code.
8:16:21code. It can understand our codebase and
8:16:22it can run commands on our behalf. And
8:16:24what's going on when you're using a
8:16:25coding assistant like cloud code is you
8:16:27basically are just looking at terminal
8:16:28or you're looking at what the AI model
8:16:30is thinking and what it's doing. But
8:16:31what's actually happening behind the
8:16:32scenes is files are getting read, plans
8:16:34are being created, code is being
8:16:36executed, and tests are being ran. So
8:16:38there's a lot of stuff going on that we
8:16:40don't really visually get to see,
8:16:41especially when we start to spin up tons
8:16:43of parallel agents and have different
8:16:44terminals running. So basically what
8:16:46happens is cloud code is working. It
8:16:48creates an activity log that already
8:16:49happened by default and pixel agents
8:16:51basically just picks up that activity
8:16:52log and then displays it in some sort of
8:16:55animation. It's honestly a very similar
8:16:56flow if you watch the way I set up my
8:16:58OpenClaw personal assistant and I could
8:17:00see what he was working on and if he
8:17:02spun up any sub aents. And like I said,
8:17:04it's pretty cool to be able to monitor
8:17:06our different agents at a glance rather
8:17:07than having to flip through different
8:17:09terminals, especially because I love to
8:17:11multitask when I'm using Cloud Code. So,
8:17:13one monitor I'll be coding with stuff
8:17:14and the other monitor I'll be working on
8:17:16other things and I can just kind of
8:17:17easily glance back and forth. So, if you
8:17:19want to actually try this out for
8:17:20yourself, it's super easy. You just have
8:17:22to go download Visual Studio Code if you
8:17:23don't have it already. Unfortunately,
8:17:25right now there's only support for
8:17:27Windows. So, if you're on a Mac or a
8:17:28different operating system, I don't
8:17:30think you can use this yet. This
8:17:31literally came out like a few days ago.
8:17:34But anyways, once you are in a fresh
8:17:35instance of VS Code, this is what it
8:17:37will look like. And what you'll do is on
8:17:39this left-hand side, you'll come over to
8:17:40extensions and you'll literally just
8:17:42type in pixel agents. And if you type in
8:17:44pixel agents, you'll be able to see
8:17:45right here pixel art office where your
8:17:47cloud code agents come to life. This is
8:17:49the one that I installed and have just
8:17:50showed you as a demo. It's fairly new,
8:17:52but I am going to go over some security
8:17:53stuff near the end of the video. But you
8:17:55could read through this to understand
8:17:56what exactly this does. You basically
8:17:58have one agent per character. You have
8:17:59live activity tracking. You can design
8:18:01the office with the floors, the
8:18:03furniture, everything like that. You can
8:18:05see speech bubbles. you get sound
8:18:06notifications, sub agent visualization,
8:18:08and there are six different characters
8:18:10that could spawn. It'll go into some
8:18:12requirements and it will go into some
8:18:13things about how to actually use it. But
8:18:15if you want to test it out, then hop in
8:18:17here. And here's where unfortunately it
8:18:19says that it's right now Windows only
8:18:20testing. And on the road map, you can
8:18:22see that there's lots of things they
8:18:23want to add like having cloud code agent
8:18:25teams because right now it's just like
8:18:27parallel agents. So anyways, once you
8:18:29have installed that, what you're going
8:18:30to do is open up a folder. So, let's
8:18:32just say you come to the explorer and
8:18:34you open up a project that you want to
8:18:35be working in. So, here I just opened up
8:18:37my Herk 2, which is basically like my
8:18:39second brain executive assistant type of
8:18:41deal. It's really weird, but if you want
8:18:42to get this project to work, your folder
8:18:44cannot have a space or a period in it.
8:18:47As you can see down here, this used to
8:18:49be called Herkpace 2.0 and then it was
8:18:50Herk 2.0 and now it's HERK-2. So, make
8:18:54sure you don't have those things in
8:18:55there if you want the pixel agents to
8:18:56actually pick up the activity log and
8:18:59actually reflect what you're doing.
8:19:00You're welcome. I had to learn that the
8:19:02hard way. So once you're in here, you're
8:19:03going to open up the terminal. So if I'm
8:19:05on Windows, I'm going to do control and
8:19:07then the little tilda. And this opens up
8:19:09the terminal. And then here is where you
8:19:10can see I have a pixel agents tab, which
8:19:12opens up the little workspace. Now what
8:19:14I want to do is I want to see half and
8:19:16half. So I basically grab this and drag
8:19:18it to the sidebar, which lets me look at
8:19:20it over here. And then all I have to do
8:19:22is click on plus agent. And that
8:19:24basically opens up a cloud code terminal
8:19:25on the right hand side. And it spawns in
8:19:28an agent right here. As you can see now,
8:19:30right now, this agent is idle because
8:19:31it's not working on anything. So, if I
8:19:33go ahead and say, "Please spin up a
8:19:35research agent and just find out what's
8:19:37going on with small businesses and AI
8:19:39adoption." And I go ahead and shoot that
8:19:40off, then this agent should pick up that
8:19:42task, and it should also spin up a sub
8:19:45agent. I'm going to add another one that
8:19:46opens up another terminal, and I can add
8:19:48more and more agents. So, we can see
8:19:50that we could watch all five of these
8:19:52different terminals run at the same
8:19:54time. And then what you can do in here
8:19:56is you can go to your layout and you can
8:19:58change things up. So I could move around
8:20:00this plant for example. I can move that
8:20:02right here. I could go to furniture and
8:20:04I have storage. I have tech. I have
8:20:06decor. I have all this kind of stuff. I
8:20:08could literally just put a monitor in
8:20:09the middle of the kitchen floor if I
8:20:11wanted to. And I could keep adding more
8:20:12and more. So it's kind of funny. You can
8:20:14fully customize the way that this looks
8:20:16and um the colors of like the floors and
8:20:19the walls and things like that. Now, the
8:20:20other setting that you can really play
8:20:21around with would be opening the
8:20:23sessions folder to see what's going on,
8:20:25which opens up like a folder like this
8:20:27to see activity log and stuff. You can
8:20:29also go to the debug view to see what
8:20:31the agents are actually working on. And
8:20:32you can turn on or off the sound
8:20:34notifications, which is pretty cool. You
8:20:36can obviously use something like hooks
8:20:37to do that either way, but it's nice
8:20:39that that is now sort of native with
8:20:41this extension. So, I'm not going to
8:20:42spend too much time here really. I think
8:20:44you guys all understand exactly what's
8:20:46going on. You can see that this agent
8:20:47just finished up. Great timing. I
8:20:49thought I should wrap up this video by
8:20:50actually just talking about like what is
8:20:52the point of this? Is this just like
8:20:55useless? So my first point is about
8:20:57entertainment. I think that there is
8:20:58value to having this as just like kind
8:21:00of something that's fun. And the reason
8:21:02I say that is kind of twofold. The first
8:21:04one is I a lot of times like I said
8:21:06earlier have two monitors up and I'll
8:21:08have cloud code going over here and I'll
8:21:09be working on other things here. And it
8:21:11is kind of nice to be able to look over
8:21:12and just see like how many agents are up
8:21:14or you know do I need to load them up
8:21:16more tasks. But the other piece of it is
8:21:17I think that it's visual and that might
8:21:20lure some people in who aren't super
8:21:22technical or who are intimidated by
8:21:23looking at a terminal. I think it's very
8:21:25similar to how nitn started to blow up
8:21:27is because they visually had some good
8:21:28elements that people felt like they
8:21:30could actually understand just by
8:21:31looking at and that kind of leads into
8:21:33my second point which is about that is
8:21:35exactly what AI coding is lacking is
8:21:37visual elements because kind of like I
8:21:39said earlier what we're looking at is a
8:21:40terminal or you know basically a chat
8:21:42window when in n or something like that
8:21:45you're seeing visual workflows you're
8:21:47seeing the process of the data and
8:21:49you're seeing things be built or you
8:21:51know changed now obviously pixel agents
8:21:53isn't that it's not a representation of
8:21:55what's they're actually doing. It's just
8:21:57like showing you how many are working.
8:21:58But I definitely think it's a step in
8:22:00the right direction cuz I think
8:22:01ultimately where I'd love to go is to a
8:22:03place where claude code builds me, you
8:22:05know, workflows and automations and it's
8:22:07showing me it building them in an NAD
8:22:10style interface where I can see the flow
8:22:12and then we can talk about multi- aent
8:22:14monitoring. There's a lot that can go on
8:22:15with different agents whether you run
8:22:17them in parallel and different terminals
8:22:18or you've got sub agents or you've got
8:22:20agent teams. And it was kind of cool in
8:22:22here that you know each of these people
8:22:24are their own terminal. And so if I see
8:22:26four of this girl, I know that that is
8:22:28one main and then four sub agents.
8:22:30Whereas all of these are just different
8:22:32agents running in terminals because they
8:22:34are different things. Parallel agents,
8:22:36you know, four different terminal
8:22:37agents. Sub agents are a parent
8:22:39delegating tasks to different sub aents.
8:22:41Those sub aents can use different models
8:22:43and they have, you know, their own
8:22:44independent context. And then you have
8:22:46agent teams which pixel agent doesn't
8:22:48yet actually show but agent teams work
8:22:50together a little bit better. They don't
8:22:52share context necessarily but they share
8:22:54a task list. They can talk to each other
8:22:56and they understand when the other
8:22:57person is done or the other agent is
8:22:59done. And even Boris the genius who is
8:23:01kind of like behind claude code said
8:23:03that he's using you know five agents in
8:23:05his terminal at all times and then maybe
8:23:07another 10 on the web or something like
8:23:08that. So he's always working with tons
8:23:10of different agents. So, I am really
8:23:12interested to see what Cloud Code does
8:23:13around sort of like more native
8:23:15visualization. So, real quick, let's hit
8:23:17on the security of this plugin because
8:23:19you should really only be installing
8:23:21extensions that are verified or that you
8:23:22know are safe. So, before I installed
8:23:24it, I just looked into it a little bit.
8:23:26The publisher is named Pablo Deuca. He
8:23:29is a verified publisher on the VS Code
8:23:30marketplace. I looked into his
8:23:32background and he is a real person. He
8:23:34is a co-founder of a company and he has
8:23:35a GitHub repo with 1300 stars and over
8:23:38100 forks. So I think that that means
8:23:40that he's got a little bit of
8:23:41credibility to him. He didn't make some
8:23:43sort of fake profile to put something up
8:23:44there to steal your data. So we looked
8:23:46at some key things like outbound network
8:23:47calls, data xfiltration, shell command
8:23:50injection, suspicious dependencies,
8:23:52credentials, secret handling, file
8:23:53system access, and remote script
8:23:55loading. And we didn't see anything. So
8:23:57it doesn't look like it sends data
8:23:58anywhere. It doesn't collect anything.
8:24:00It doesn't run any commands and
8:24:01everything is staying on your machine.
8:24:02So that is why I felt confident to
8:24:04actually install this and test it out.
8:24:05So, I just wanted to share my experience
8:24:07from a security perspective. So,
8:24:09overall, the way I feel about this is
8:24:11that Pixel Agent just tells me that an
8:24:12agent or agents are working. What I
8:24:15actually want to know is what they're
8:24:16building, what decisions they're making,
8:24:18and whether it's about to do something
8:24:19that I'd agree with or disagree with.
8:24:21That's the real product that I think we
8:24:22need when it comes to visualization.
8:24:24Because us as humans, we're moving into
8:24:27this world where we really just have to
8:24:29be really good managers and keep our
8:24:30agents and our sub agents on the right
8:24:32path. And the best way to do that is be
8:24:34able to just see all the time what's
8:24:36going on. Stop things before they go
8:24:38wrong and proactively keep giving them
8:24:40more work so that they're never just
8:24:41sitting there idle. But anyways, that is
8:24:43going to do it for today. If you guys
8:24:44want to dive deeper into all this kind
8:24:45of stuff and nerd out, then definitely
8:24:47check out my paid group, AI Automation
8:24:49Society Plus. We've got over 3,000
8:24:51members who are building with AI every
8:24:52day and building businesses with AI. So,
8:24:54it's a great place to be if you have
8:24:56similar goals. But if you guys enjoyed
8:24:58the video or you learned something new,
8:24:59please give it a like. It definitely
8:25:00helps me out a ton. And as always, I
8:25:01appreciate you guys making it to the end
8:25:03of the video. I'll see you on the next
8:25:04one. Thanks, everyone.
8:25:09So, we are just absolutely cruising
8:25:10through this course. I'm having a lot of
8:25:12fun. Hope you guys are, too. Now, since
8:25:13we have all these new skills and all
8:25:15this new knowledge, we're going to get
8:25:16into how can we actually like make money
8:25:18off of this or kind of start a business
8:25:20off of this. So, we've got a couple
8:25:22things to talk about. We're going to
8:25:23talk about the mindset of actually being
8:25:24kind of like an AI consultant or an AI
8:25:26service provider. I'm going to talk
8:25:28about how you can think about getting
8:25:29some clients even if you don't have a
8:25:31big YouTube channel or a big audience.
8:25:32And then we're going to talk about
8:25:33actually sort of like deploying and
8:25:35handing over these solutions.
8:25:40If you want to make money with AI, you
8:25:41need to stop selling AI agents and
8:25:43workflows and start selling AI
8:25:45solutions. And I'm not talking about AI
8:25:46operating systems or regular
8:25:48automations. I'm talking about
8:25:49diagnosing problems that businesses face
8:25:51and then using AI to solve those
8:25:53problems. So, in this video, I'm going
8:25:55to break down why selling agents isn't
8:25:56enough. what to sell instead that will
8:25:58actually make you money and how to do it
8:25:59yourself step by step. So, why are you
8:26:01guys all here watching this video?
8:26:03Because AI has exploded. And with that
8:26:05explosion comes hype. Tons of people
8:26:06throwing around buzzwords like agents
8:26:08and workflows. And what's funny is
8:26:10automation has been around forever. It's
8:26:11nothing new. When I was working
8:26:12full-time at Goldman Sachs in business
8:26:14intelligence, [music]
8:26:15automation was my entire role. And that
8:26:16job had existed for years. But the
8:26:18majority of small and mid-size
8:26:20businesses, hardly any automations in
8:26:22place. The difference now is that
8:26:23because you slap the word AI in front of
8:26:25it, suddenly business owners are paying
8:26:26attention. It's like putting a neon sign
8:26:28on an old restaurant, the food hasn't
8:26:30changed. It's nothing new, but suddenly
8:26:31people are noticing it. And here's the
8:26:33thing, most beginners get this part
8:26:34wrong. They get caught up in the tech,
8:26:36the nodes, the HTTP requests, the multi-
8:26:38aent architectures. They're so focused
8:26:40on the shiny parts that it causes them
8:26:41to forget what actually matters to
8:26:43businesses. [music] And I've definitely
8:26:44got carried away with this, too. In
8:26:45fact, my best performing YouTube videos
8:26:47are the ones that are really flashy and
8:26:49have multiple agents [music] because it
8:26:50was trendy and cooler. But if you were
8:26:52to ask me which YouTube builds have
8:26:53actually been the most practical and
8:26:55highest ROI, it would be some of my
8:26:56lowest performing videos. When I sold my
8:26:58very first $1,200 workflow, I didn't
8:27:00walk in saying, "Hey, I got this build
8:27:02with 15 nodes and an API call to blah
8:27:04blah blah." I said, "This will save you
8:27:06hours every single week on content
8:27:08creation." And that's what made it a
8:27:10no-brainer. And here's the truth.
8:27:11Businesses really only care about three
8:27:13things: time, money, focus. That's it.
8:27:16They don't care if you deliver it with
8:27:17AI or a VA or duct tape and chewing gum.
8:27:20Think of it like a taxi ride. If you
8:27:21need to get across town, you don't care
8:27:23if you ride in a Prius, a Tesla, or a
8:27:25horsedrawn carriage. Well, I suppose
8:27:26some of us may, but that's beside the
8:27:28point. You just care that it gets you to
8:27:29your destination fast, cheap, and
8:27:31without stress. Businesses feel the same
8:27:33way. They don't fanboy over AI itself.
8:27:35They fanboy over the outcomes. That's
8:27:37why AIcentric selling doesn't work.
8:27:39Selling agents or workflows as products
8:27:42is already a crowded commoditized
8:27:43market. You've probably seen people
8:27:44offering entire libraries of end
8:27:46templates, claiming that you can resell
8:27:48them all for $5,000 a month. yet they're
8:27:50only charging you $200 to access that
8:27:52bundle of templates. Why? Because it's a
8:27:54race to the bottom on price. Take
8:27:56LinkedIn outreach bots as an example.
8:27:58You could build the flashiest, most
8:27:59complex bot out there, but if you just
8:28:01pitch it as a LinkedIn outreach agent,
8:28:02nobody's going to care that much. But
8:28:04when it is framed as a system that helps
8:28:06you generate qualified leads without
8:28:08spending on ads, suddenly people see the
8:28:09value and they want it. The shift is
8:28:11simple. Stop selling AI tools and start
8:28:13selling the outcomes of those tools.
8:28:15Spotlight the pain. Frame the solution
8:28:17around time, money, and focus. And
8:28:19that's how you win. And that leads us
8:28:20right into the next section. What should
8:28:22you actually be selling? AI solutions.
8:28:24The difference is an AI workflow is just
8:28:26a tool. But an AI solution is a tool
8:28:28that's directly tied to solving an
8:28:30actual business painoint. When you pitch
8:28:31a solution, you're not saying, "I'll
8:28:33build you a chatbot." You're saying,
8:28:34"I'll cut down your customer support
8:28:35workload by 60%. I'll automate your
8:28:38client onboarding so you save 10 hours a
8:28:39week." Or, "I'll help you create AI
8:28:41assisted content so you can slash
8:28:42marketing spend by 30%." [music] And
8:28:44just to get it to stick, I'll throw one
8:28:45more terrible analogy at you guys. Think
8:28:48of it like medicine. If you have a
8:28:49headache, most people don't care whether
8:28:50you prescribe Advil, Tylenol, or an
8:28:52herbal remedy. They just care that their
8:28:54headache goes away. The pill is the
8:28:55tool. The outcome is pain relief. That's
8:28:57exactly how businesses see AI. So, the
8:28:59framework we're going to be talking
8:29:00about today is to diagnose, then solve,
8:29:03then value, then price. And it's a
8:29:05simple process that I use every time.
8:29:07Diagnose means find the pain point.
8:29:08Where is the business leaking time,
8:29:10money, or focus? Solve means build the
8:29:13system that fixes that exact pain.
8:29:15value. Translate that fix into numbers.
8:29:17Hours [music] saved, dollars saved,
8:29:18revenue gained. And finally, price.
8:29:20Anchor your offer around that value. And
8:29:22when you think about that framework and
8:29:23you think about the core problem that
8:29:25you're trying to solve, you can see how
8:29:26reselling a bundle of 20 end templates
8:29:28doesn't really do that. So, let's take a
8:29:30look at an example. Automating client
8:29:32onboarding. First, we diagnose. Right
8:29:34now, your team spends 5 hours a week
8:29:35onboarding new clients. Then, we solve.
8:29:37I'll build a system that can automate
8:29:3880% of that. Then, we turn that into
8:29:40value, which is 200 plus hours saved a
8:29:42year at $50 an hour. That's $10,000 back
8:29:45in your pocket annually. And [music]
8:29:46then you price, I'm saving you 10K a
8:29:48year. I'm only going to charge you 3K
8:29:50right now. Suddenly, the price isn't a
8:29:52cost. It's now just a fraction of the
8:29:54value that they're gaining. Here comes
8:29:55another analogy. Think of it like a home
8:29:57renovation. If you're a contractor and
8:29:59you walk into someone's house and you
8:30:00say, "I've got the best hammer. I've got
8:30:01the strongest nails, the fanciest saws,
8:30:03blah blah blah. Nobody cares." [music]
8:30:05But if you say, "This remodel will
8:30:07increase your home's value by $50,000
8:30:09and we can do it in half the time of
8:30:11other contractors." Now you've got their
8:30:12attention. The hammer doesn't matter,
8:30:14the results do. This is why when I'm on
8:30:16calls, I don't lead with, "Let me show
8:30:18you my workflow." I lead with questions.
8:30:20Where are you losing the most time in
8:30:22your business? What processes do you
8:30:23wish could run themselves? Because the
8:30:25diagnosis always points me to the
8:30:27solution. [music] And the solution
8:30:28always ties back to measurable business
8:30:30value. And when you get good at this,
8:30:32you stop being just another automation
8:30:33freelancer. You become the person who
8:30:35can look at a business, spot
8:30:36inefficiencies, and design solutions
8:30:38that save them time, money, and focus.
8:30:41That's the shift from becoming an agent
8:30:42seller to a solution seller and being
8:30:44seen as an AI expert or an AI partner.
8:30:46So, you know why selling AI doesn't work
8:30:48and that you need to sell AI solutions,
8:30:49but how do you actually do this for
8:30:51yourself? Step by step. That's what
8:30:52we're going to cover now in a few simple
8:30:54steps. And remember, we're going to tie
8:30:55everything back to that same framework
8:30:56that we just talked about. Diagnose,
8:30:58[music] solve, value, price. Okay, so
8:31:00step one, pick a niche and diagnose the
8:31:02landscape. The goal here is simple.
8:31:04Choose one group so that problems repeat
8:31:06and your wins compound. Spend 10 minutes
8:31:08deciding whether you'll serve agencies,
8:31:10real estate, e-commerce, coaches, local
8:31:12services like dental or HVAC, SAS
8:31:15startups, whatever you want it to be.
8:31:16Run your pick through these quick
8:31:18filters. Do they repeat processes
8:31:19weekly? Can they say yes and pay fast?
8:31:22And do you already speak their language
8:31:23through content you've made or past work
8:31:25or interests? [music] When you think
8:31:26about repeatable pains, picture agencies
8:31:28struggling with lead qualification,
8:31:29client onboarding, reporting, and
8:31:31content ops. Real estate teams juggling
8:31:33inbound lead triage, showing
8:31:35coordination and document collection.
8:31:37E-commerce brands needing CX ticket
8:31:39deflection, returns automation, product
8:31:41content, and ops reporting, or coaches
8:31:43begging for application filtering,
8:31:44calendar triage, and content
8:31:46repurposing. The outcome of this step is
8:31:47just to have a short list of niches that
8:31:49you're confident enough to carry into
8:31:51this next step. Step two, talk to five
8:31:53to 10 businesses and diagnose pain.
8:31:55Treat these as informational interviews
8:31:56that could often turn into discovery
8:31:58calls and paid work. I like opening with
8:32:00a simple message. Hey name, I'm mapping
8:32:02the top drains in X niche. In 15
8:32:05minutes, I'll try to quantify your
8:32:06biggest bottleneck and share where AI
8:32:08actually helps and where it doesn't. No
8:32:10pitch unless you ask. I'm genuinely just
8:32:12trying to learn how I can provide value
8:32:13to businesses. On a call, I would then
8:32:15use the LRP framework, the listen,
8:32:18repeat, poke. This basically means that
8:32:19I listen while they describe their week
8:32:21or their processes. I repeat back the
8:32:22pattern to confirm alignment and then I
8:32:24poke to quantify it. So, let's say the
8:32:26business owner is talking to me about
8:32:28some onboarding and reporting flows and
8:32:29I find out that it eats about 6 to 8
8:32:31hours a week. I would then poke back
8:32:32with a question like, "Whose hours are
8:32:34those? What's their hourly value? How
8:32:35often does this process result in an
8:32:37error?" You could also talk about money.
8:32:38Where are you paying people to copy and
8:32:40paste or chase info? You could talk
8:32:41about focus. What interrupts you the
8:32:43most between 9 and noon? You can talk
8:32:45about errors. Where do mistakes cause
8:32:46rework, refunds, or churn? And finally,
8:32:49you could talk about a trigger. If I
8:32:50could remove one weekly fire, which one
8:32:52changes your week? The outcome of this
8:32:54step is to have a ranked list of actual
8:32:56business pain points with rough numbers
8:32:57associated with them. Those numbers
8:32:59being hours, cost, [music]
8:33:00frequency, mistake rate, stuff like
8:33:02that. We can then take that list and
8:33:04move into step three, which is building
8:33:05one simple solution and solve. Once
8:33:08you've diagnosed a clear problem,
8:33:09prototype the fix. And notice how I just
8:33:11said prototype. This doesn't have to be
8:33:12a perfect production ready system. It's
8:33:14a P or a proof of concept that shows
8:33:17that you understand their world and it
8:33:19proves your technical expertise. This
8:33:20doesn't mean you have to dedicate an
8:33:22entire week to it. Here's how you can
8:33:23spin one up in an hour and a half. Spend
8:33:24about 15 minutes drawing the flow. Now
8:33:26that you understand the process, just
8:33:27map out the trigger, the steps, the data
8:33:29sources, the outputs, and a clear
8:33:31definition of done. Spend the next hour
8:33:33doing a rough build on one platform like
8:33:35Niten in so you have something real to
8:33:37click through and demo. And then
8:33:38finally, use the last 15 minutes to
8:33:40record a quick 3minute loom that walks
8:33:42through the before, the actual solution
8:33:45itself, and then the result. And make
8:33:46sure you have your camera on because
8:33:47people want to hire people, not faceless
8:33:49screens. [music] This is exactly why my
8:33:51not so fancy LinkedIn workflow felt like
8:33:53magic to that first client because it
8:33:54removed the pain of ideiation research
8:33:56and writing. And it didn't need to be
8:33:58super fancy to do that. So avoid the
8:34:00classic traps. Don't drift into multi-
8:34:02aent fantasies. Don't stitch together
8:34:04five vendors just to try them. And don't
8:34:06start building a platform when the job
8:34:07is just to fix Tuesday's bottleneck.
8:34:09Just try to take a step back and think
8:34:10about the actual problem you're trying
8:34:12to solve. Because a lot of times you may
8:34:13not even need like a very custom nit
8:34:16agent. Sometimes there's already a
8:34:17platform out there or sometimes [music]
8:34:18you can do it already naturally in
8:34:20someone's CRM. Don't overengineer. Just
8:34:22be resourceful and solve problems. The
8:34:24outcome of this step is that you now
8:34:25have a demo video that you can send to
8:34:27some of your other nurtured leads to
8:34:29show them how you can solve their
8:34:30[music] specific pain points. Next is
8:34:32step four, which can trip up a lot of
8:34:33beginners. This is turning your solution
8:34:35into a price. And the easiest way to do
8:34:37that is to translate what you built into
8:34:39plain math that a business owner can
8:34:40understand and relate to. Here's how I
8:34:42explain it on calls. First, you have to
8:34:44figure out how much time the process
8:34:45takes today. Hopefully, you already did
8:34:47that earlier in step two. Let's say a
8:34:48team spends 10 hours a week on client
8:34:50onboarding. Then ask what those hours
8:34:52are worth. If the average employee is
8:34:53paid $25 an hour, that's $250 a week.
8:34:56Multiply that by 4 weeks to get a month
8:34:58and suddenly the process is costing the
8:35:00business around $1,000 a month or
8:35:02$12,000 annually. Now, let's say you're
8:35:04only able to automate 60% of that
8:35:06workload. That's still $600 saved every
8:35:08single month. Over the course of a year,
8:35:10that's more than $7,000 back in that
8:35:12business owner's pocket. So, when you
8:35:13turn around and say, "Hey, I'm only
8:35:14going to charge you $3,000 for the
8:35:16solution." It's not an expense anymore.
8:35:17It's an investment that has a very
8:35:19clear, measurable return that already
8:35:21pays for itself in 5 months, and it
8:35:22keeps saving more money every single
8:35:24month after that. And that's how you
8:35:25make it a no-brainer. And of course, you
8:35:27have to remember that value doesn't
8:35:28equal time. The client's not paying you
8:35:30for how long it took. When I built my
8:35:31very first workflow for $1,200, it only
8:35:33took me about 2 hours, which is $600 an
8:35:35hour, which is a crazy rate. But once
8:35:37again, the client wasn't paying me for
8:35:38the hours. they were paying for the
8:35:40outcome, which was saving [music] them
8:35:42hours of content creation every week.
8:35:44Finally, make sure to keep your scope
8:35:46clear and simple. Write down the
8:35:47objective, what's included, what's not
8:35:49included, the timeline, [music] what's
8:35:51expected from the client, the payment
8:35:53terms, stuff like that. This avoids
8:35:55confusion, ambiguity. It protects you
8:35:57from scope creep, and it makes you look
8:35:58professional. The number one biggest
8:36:00mistake that I made when I was starting
8:36:01off was underscoping and then having to
8:36:03deal with all of this ambiguity. So if
8:36:05you can make your proposals as specific
8:36:06as possible and saying these 10 things
8:36:08are the exact functionality of the
8:36:10system, this is exactly the definition
8:36:12of done. Then you're really going to be
8:36:13protecting yourself and thanking
8:36:15yourself later. So to recap, the formula
8:36:17is dead simple. Figure out how much time
8:36:18and money the process is costing now.
8:36:20Build a PC and show how your solution
8:36:22cuts down that time and money. And then
8:36:24price your project at a fraction of
8:36:25those yearly savings. Businesses will
8:36:27pay happily when that math makes sense.
8:36:29All right, moving on to step five, which
8:36:31is to stack proof, build confidence, and
8:36:33then scale. I'm going to be honest real
8:36:34quick. When you're starting, imposter
8:36:36syndrome is normal. You're going to
8:36:37wonder if you can deliver and whether
8:36:39the client will be happy. I felt the
8:36:40exact same way. When I started working
8:36:41with clients, I wasn't even calling
8:36:43myself an expert. I was just curious,
8:36:45building for fun, sharing things online,
8:36:47and clients started reaching out. At
8:36:48this stage, I would have been happy to
8:36:50do work for free just to get experience.
8:36:52And the fact that I was getting paid was
8:36:53a big bonus. And if you're stuck with
8:36:54that worry, you have two great options.
8:36:56The first one is to do free or very
8:36:58cheap work in exchange for proof. Say,
8:37:00"Hey, I'll try this build out for you
8:37:02for free. Either way, the system will be
8:37:03yours. I just want to see if I can help
8:37:05create something valuable for you and
8:37:07your business. This removes pressure and
8:37:08lets you optimize for reps, not money,
8:37:10which is exactly what you should be
8:37:11doing when you're starting off because
8:37:13the testimonials and case studies and
8:37:14confidence that you're going to gain are
8:37:16worth far more downstream than a quick
8:37:18$1,200 today. And those first three to
8:37:20five projects really should be just
8:37:21thought of as paid practice. They exist
8:37:22[music] to teach you the process, give
8:37:24you those case studies, and provide
8:37:25proof. Your other option is if you'd
8:37:27rather not do it for free, just offer a
8:37:29money back guarantee. say, "Hey, if
8:37:30you're not happy with it when I deliver
8:37:32it or it doesn't deliver the value that
8:37:34we discussed, you'll get all your money
8:37:35back." That flips the risk, creates a
8:37:37stress-free environment for you to learn
8:37:38and experiment and still gives the
8:37:40client a fair outcome. This matters
8:37:41because in the early days, once again,
8:37:43you're optimizing for experience, proof,
8:37:45reps, not money. [music] And once you
8:37:46have all that confidence, you can
8:37:48actually move from being a freelancer to
8:37:49a consultant and eventually get larger
8:37:51and larger builds and larger and larger
8:37:53contracts. Over [music] the past year,
8:37:54we've been working on projects that have
8:37:56ranged anywhere from $1 to $2,000 all
8:37:58the way up to $30,000 plus. [music] And
8:38:00none of that would have happened without
8:38:01the low stakes project that built up my
8:38:03confidence and portfolio, which let me
8:38:05comfortably charge more and more every
8:38:07time we sculpted out builds. And
8:38:08honestly, I still think a lot of our
8:38:09projects we've undercharged, but that's
8:38:11a conversation for another time.
8:38:12Anyways, once you have those first wins
8:38:14under your belt, make sure you're
8:38:15collecting that baseline and after data.
8:38:18So, your hours saved, your errors
8:38:19reduced, your money saved, [music] and
8:38:20then turn them into simple case studies.
8:38:22Of course, you want to ask for
8:38:23testimonials and referrals and then
8:38:25rerun the framework with new prospects.
8:38:27But now that you have proof in hand,
8:38:29it's going to be a lot easier because at
8:38:30that point, you're no longer saying, "I
8:38:32think I can help you." You're saying,
8:38:33"I've already helped three businesses
8:38:34just like yours. Here's the proof.
8:38:36Here's the results. You want me to do
8:38:38the same for you?" And that's when the
8:38:39imposttor syndrome fades because the
8:38:41market is actually validating that what
8:38:43you're doing is working. And that's
8:38:44pretty much the full loop. Diagnose the
8:38:46problem, solve it with a simple system,
8:38:47tie it back to business value, and then
8:38:49price accordingly. From there, you stack
8:38:51up your proof until you can scale with
8:38:53confidence. If you want to succeed at
8:38:54selling AI, you've got to leave your
8:38:56obsession for technology behind and
8:38:57think purely about the customer, aka the
8:39:00business that you're selling to. So, I
8:39:01know that was a lot of information.
8:39:02Tried to make this one as valuable as
8:39:04possible. What I'm going to do is have
8:39:05all of this condensed into a resource
8:39:07that you guys can access and look at
8:39:08later. I will put that in my free school
8:39:10community. Once again, completely free.
8:39:12The link for that will be down in the
8:39:13description. And if you found this video
8:39:14helpful and you're looking for a full
8:39:16course pretty much just like this, then
8:39:18check out my plus community. The link
8:39:19for that's also down in the description.
8:39:20We just dropped a full course in there
8:39:22called oneperson AI agency. So you'll be
8:39:24able to get a formal course material
8:39:25that goes over starting your oneperson
8:39:27AI agency from scratch. And the great
8:39:29thing about that community is we have
8:39:30over 2,000 other people who are also
8:39:32trying to do that exact same thing. So
8:39:33it's a great space to meet people
8:39:35potentially a business partner and maybe
8:39:36even make a few hires in there or pick
8:39:38up your first couple gigs.
8:39:42People keep telling me they can't land
8:39:43AI clients because they don't have a
8:39:45YouTube channel since I got my first
8:39:46clients through content. But you don't
8:39:48actually need content at all. If I lost
8:39:49my YouTube channel tomorrow and I had to
8:39:51start from zero, there are three simple
8:39:52ways that I'd get clients fast. And I
8:39:54hardly see anyone talking about the
8:39:55third method. So in this video, I'll
8:39:57show you real proof that you can land
8:39:58clients without a YouTube channel and
8:40:00the exact [music] three steps you can
8:40:01follow to sign your first few clients
8:40:03this month. So let's get into it. So
8:40:04just to start off, I'm not going to
8:40:06pretend that YouTube or content doesn't
8:40:08help. It absolutely does because when
8:40:09someone watches a ton of your videos
8:40:10before reaching out, they already have
8:40:12trust and they've heard your voice.
8:40:13They've seen how you think. And by the
8:40:14time they book that call, they're
8:40:15basically sold. And that's the power of
8:40:17content. But what most people don't
8:40:18realize, the agency that I founded and
8:40:20exited, True Horizon, scaled past 100K a
8:40:22month. And even though the majority of
8:40:23our leads were coming from YouTube,
8:40:24that's not where our best clients were
8:40:26coming from. The clients we ended up
8:40:27actually working with were the ones who
8:40:28paid well, stuck around, and were a
8:40:30dream to work with, came from two
8:40:31things: referrals and partnerships. So,
8:40:33we'd make one client happy and then
8:40:34they'd tell their network or we'd
8:40:36partner with tools or other businesses
8:40:38that already had demand that they
8:40:39couldn't service themselves. And that's
8:40:40where the real growth came from. So,
8:40:42yes, content is one path and it's a
8:40:44really great way to build momentum, but
8:40:45it's not the only path. And I get it
8:40:47because I know a lot of you guys don't
8:40:48want to start a YouTube channel or don't
8:40:49want to be consistently posting on
8:40:51LinkedIn every day. You don't want to be
8:40:52on camera. You don't want to wait 12
8:40:53months to see results. And that's
8:40:54totally fair. So, we're going to talk
8:40:55about three ways that you can land
8:40:57clients without posting a single piece
8:40:58of content. Starting with number one,
8:41:00which is cold outreach. Now, before I
8:41:01get into this, I need to be real with
8:41:02you. Cold outreach is brutal when you
8:41:04have zero proof. The first thing that
8:41:06any prospect is going to ask you is,
8:41:07"Who have you done this for?" And if
8:41:08your answer is nobody, you're fighting
8:41:10uphill the entire conversation. So
8:41:11before you go off and send a bunch of
8:41:13cold messages, go get some proof. Even
8:41:15if it's small, even if it's free, even
8:41:16if it's for your cousin's salon, the
8:41:18product doesn't need to match what you
8:41:19eventually want to sell. All that
8:41:21matters is that you can say, "Yeah, I
8:41:22have done this before. I've helped X
8:41:23business get Y result." That alone
8:41:25separates you from everyone who's purely
8:41:27theoretical. So there are four
8:41:28components to cold outreach: platforms,
8:41:30finding leads, reaching out, and then
8:41:32volume. So let's start with platforms.
8:41:33For B2B, here's where you should be
8:41:35looking. LinkedIn, Facebook groups,
8:41:36email, school communities, YouTube
8:41:38channels, Instagram, Reddit. Cold email
8:41:40has an average response rate of 1 to 5%,
8:41:42but personalized emails see up to 17%
8:41:44higher rates. LinkedIn inmail DMs have a
8:41:4710 to 25% response rate compared to
8:41:49email, which is much lower. But the
8:41:51LinkedIn DMs require more time per
8:41:53message. So, I wanted to just throw
8:41:54those stats at you, but pick one or two
8:41:56to start. Don't try to do outreach on
8:41:57all of those platforms. You're going to
8:41:59spread yourself too thin because they
8:42:00all have different methods of what works
8:42:01best. So, once you've picked your
8:42:03platform, here's where to actually find
8:42:04people worth reaching out to.
8:42:05communities in your niche, followers of
8:42:06relevant pages or creators, business
8:42:08owners already complaining about wasting
8:42:10time or making mistakes, job boards and
8:42:12comment sections because if someone's
8:42:13hiring for a role you could automate,
8:42:14that's a signal. You could also look at
8:42:16local businesses in your area. Pro tip
8:42:17though, before you jump into something
8:42:18like Apollo or LinkedIn Sales Navigator
8:42:20to find a lead list, spend 10 minutes
8:42:22asking CatchBT or Perplexity where
8:42:25industry specific directories exist. So
8:42:27like architects have the American
8:42:28Institute of Architects, coaches have
8:42:30coaching directories. Every industry
8:42:31probably has something and these
8:42:32databases have already been filtered for
8:42:34you. Now when it comes to reaching out,
8:42:36keep it simple. Focus on of course the
8:42:37problem, the outcome, not the actual
8:42:39tech solution. Nobody really cares if
8:42:41you built an 18 node automation with an
8:42:43end and Slack and notion MCP. They care
8:42:45about the results, which is, you know,
8:42:46more leads, less manual work, fewer
8:42:48errors, faster turnaround. Short value
8:42:50focused messages work better than long
8:42:52pitches. Your goal is not to close them
8:42:53in one message. That's just not
8:42:54realistic. Your goal is to get them
8:42:56curious, to start a conversation, get
8:42:57them wanting more. That's why the
8:42:59subject line is so important because if
8:43:00they're not curious at all, they're not
8:43:01going to click to open the email. Now,
8:43:03here's something, and I know it sounds a
8:43:04little bit counterintuitive, but just be
8:43:05completely honest about where you're at.
8:43:07That's my approach at least. So, saying
8:43:08something like, "Yeah, I'm just getting
8:43:09started. I'm trying to start up a
8:43:10business." And that doesn't undercut
8:43:11you. It actually creates like a pattern
8:43:13interrupt and it builds trust because
8:43:15the prospect thinks, "Okay, this is
8:43:16actually a real human." Instead of,
8:43:18"Wow, great." You know, another faces
8:43:19pitch of someone trying to sell me. And
8:43:21of course, because you're unknown and
8:43:22unproven, you want to remove all the
8:43:24risk from your offer. So, no payment
8:43:25until results, no contract. The only
8:43:27thing that you ask for is permission to
8:43:29use them as a case study if the solution
8:43:30works. You're not asking them to trust
8:43:32you and give you a ton of money. You're
8:43:34making it a no-brainer for them to say
8:43:35yes. And then one more thing, asking a
8:43:37stranger for 30 minutes on their
8:43:38calendar is a big ask. People are busy.
8:43:40People don't want to hop on calls with
8:43:41random people just to be sold. So,
8:43:42instead, get them curious. Again, ask
8:43:44them if you can send them a 2-minute
8:43:45Loom video explaining the offer. and
8:43:46then you can record a quick video
8:43:48walking through how you'd help their
8:43:49specific business and let them do the
8:43:50heavy lifting before they ever get on a
8:43:52call with you. All right, the numbers
8:43:53game. So, in the wise words of Alex
8:43:55Ramosi, someone I've actually had the
8:43:56pleasure of meeting multiple times,
8:43:57volume negates luck. So, if you want to
8:43:59guarantee clients, you just need to
8:44:01increase the volume. At the very least,
8:44:02you should be sending 100 messages a
8:44:04day. I know that sounds like a lot, but
8:44:05understand this. A small percentage of
8:44:07replies is actually a really good
8:44:08metric. So, we talked about the 1 to 5%
8:44:10average with cold email. If you're doing
8:44:11anything above 5%, you're crushing it.
8:44:13Which means if you send 100 emails and
8:44:14you only get five people to respond, not
8:44:16book in a call or pay you, just respond,
8:44:19that's actually really good. But you do
8:44:20need to track everything. Your reply
8:44:21rate, your positive responses, your
8:44:23meetings booked, your deals closed. Each
8:44:24metric tells you what area you need to
8:44:26improve. And you don't just track the
8:44:28wins, you also have to track all of the
8:44:29negative responses because that tells
8:44:30you what not to do. So when you get
8:44:32positive replies, you write down the
8:44:33company size, the industry, the job
8:44:35title, anything else that's relevant.
8:44:36And then you can start to build your
8:44:38future lead lists around those patterns.
8:44:40And this is how you actually
8:44:41systematically improve instead of just
8:44:42guessing. Now, here are the two things
8:44:44that most people do get wrong with cold
8:44:45outreach. The first one is maybe just
8:44:47don't build before you sell. If you're
8:44:48spending weeks perfecting an automation
8:44:50before you've talked to a single
8:44:51customer, you don't have a business, you
8:44:53just have a hobby, you need to use cold
8:44:54outreach as market research. Create an
8:44:56offer around something that you could
8:44:57build that you know it's feasible and
8:44:59then send messages about it. See if
8:45:00anyone bites. If they do, and you see
8:45:02market validation, then go build it. But
8:45:04don't build before you sell. Once again,
8:45:05if you're spending all your time
8:45:06building something that no one's going
8:45:07to buy, there's just no point. So the
8:45:09second one is to start small and then
8:45:10reinvest. You don't need massive email
8:45:12infrastructure with thousands of
8:45:13domains. You just need a few and each
8:45:15domain can send about 30 messages a day
8:45:16and that's enough to generate real data.
8:45:18Once you close your first deal, you
8:45:19reinvest that money into scaling up your
8:45:22outreach. And look, the thing is people
8:45:23will ignore you. People will tell you to
8:45:25screw off and that's normal. Most people
8:45:26quit because of that emotional weight
8:45:28and how tough it is to do cold outreach.
8:45:30So just understand that and don't take
8:45:31it personally. Use it as data and then
8:45:33if you can push through, you're already
8:45:34ahead. If you guys are interested in
8:45:35diving deeper specifically on cold
8:45:37email, then I just did a podcast with
8:45:38Savon, who was able to generate over
8:45:40half a million dollars in sales
8:45:41opportunities with cold email in 6
8:45:43months as a beginner. So, I'll go ahead
8:45:44and link that right up here. But
8:45:45anyways, that is cold outreach. So,
8:45:46let's talk about method number two,
8:45:48which is referrals. So, why do referrals
8:45:50work so well? Simple, because business
8:45:51owners trust recommendations way more
8:45:53than strangers. And that's not even
8:45:54about business owners. That's just
8:45:55people in general. Think about how many
8:45:56times your friend has told you about a
8:45:58great restaurant or a TV show. You're
8:45:59way more likely to go check it out than
8:46:01if you saw an ad or a review for it
8:46:03because you trust your friend. In fact,
8:46:0492% of B2B buyers trust referrals from
8:46:07people that they know. And being in the
8:46:08AI space, it's really good to lead on
8:46:11referrals because it's very new right
8:46:13now and a lot of people are thinking,
8:46:14you know, I don't even know where I
8:46:15could go find a vendor to help me out
8:46:17with this problem. Business owners want
8:46:18to know that someone that they trust has
8:46:19already taken the risk and it worked out
8:46:21well for them. When I first started
8:46:22freelancing, I'd make it a priority to
8:46:24overd deliver for the client. So maybe
8:46:25that looks like adding a simple
8:46:26dashboard that they didn't ask for or
8:46:27creating full documentation so that they
8:46:29could understand how everything works
8:46:30and they can maintain it if they choose
8:46:32to. Then, and only then, I knew they
8:46:33were happy. Usually after we'd been
8:46:35working together for a while, I'd whip
8:46:36out the golden question. Hey, do you
8:46:38have any friends or know any other
8:46:40business owners who might need AI
8:46:41automation or might just be interested
8:46:43in talking to me about this kind of
8:46:44stuff? You don't need a fancy script or
8:46:45a formal incentive program. Just a
8:46:47simple question after you've earned the
8:46:49right to ask. Happy clients want to
8:46:50share. Think about how many times that
8:46:51you've wanted to put on a friend to a
8:46:53new restaurant, a new song, or a new
8:46:55product. People typically enjoy telling
8:46:57other people about things that worked
8:46:58for them. So, if you've done great work,
8:47:00they'll probably want to brag about it.
8:47:01But the timing really matters because
8:47:03obviously you can't ask for referrals
8:47:04until you know that they're happy and
8:47:06that they have to truly believe you did
8:47:07provide them value. So typically I'd
8:47:09wait about a month or two after
8:47:10something's been like pushed live. And
8:47:12ideally you'd want to tie that into a
8:47:13moment where you can show them results.
8:47:15So maybe you're doing a month one
8:47:16performance review. So you would
8:47:18obviously say you know like hey when we
8:47:19started here are all the KPIs and the
8:47:20metrics that we had documented and
8:47:22here's what we were going to track and
8:47:23now after 1 month of the solution being
8:47:25in production we've hit all of these
8:47:26KPIs and your business has grown you
8:47:28know x amount. The project has been a
8:47:29success. And then once they actually see
8:47:31that data, you can ask about, you know,
8:47:33the golden question. The conversation
8:47:34then feels a lot more natural because
8:47:35you've just reminded them of the value
8:47:37you delivered. It's not awkward. It's
8:47:38not like you're trying to sell them
8:47:39again. It's earned. Now, you could
8:47:40absolutely set up like a formal referral
8:47:43structure if you want, but honestly, a
8:47:44lot of times you don't need anything
8:47:45formal. Just do great work and then ask.
8:47:47Now, here are the common mistakes I see
8:47:48when people are trying to get into like
8:47:50the referral game. The first one is
8:47:51asking too early before you've actually
8:47:53proven yourself. And then the second one
8:47:54is not asking at all and just leaving
8:47:56that money on the table because you were
8:47:57too scared to ask. According to Dale
8:47:59Carnegie research, only 11% of
8:48:01salespeople are actually asking for
8:48:02referrals. Yet, 91% of customers said
8:48:04that they'd gladly give one if they were
8:48:06asked. So, make sure you are asking for
8:48:08referrals. They will lead to bigger
8:48:09projects. They'll lead to longerterm
8:48:10clients, and they'll be way less
8:48:12friction than doing cold outreach. And
8:48:13once you've got clients that you're
8:48:14working with for longer term, it opens
8:48:16up method number three. And honestly,
8:48:17this is the one that I'd be going allin
8:48:18on if I was starting from zero. I call
8:48:20it the Trojan horse method. The big idea
8:48:22here is that instead of trying to build
8:48:24trust from scratch with cold prospects
8:48:25and build up all of your own authority,
8:48:27you borrow someone else's. You borrow
8:48:28their trust, you borrow their authority,
8:48:30and you borrow their existing client
8:48:31relationships. Partner source deals
8:48:33close 46% faster than other deal types.
8:48:35So, think about it like this. Marketing
8:48:37agencies, consultants, coaches, law
8:48:38firms, they already have clients who
8:48:40trust them and have maybe been working
8:48:42with them for years. They've already
8:48:43built those relationships. And right
8:48:44now, a lot of those clients are
8:48:45thinking, I need AI. And the agency
8:48:47itself is probably thinking, I need to
8:48:49be able to offer AI to my clients. So
8:48:51instead of competing for their
8:48:52attention, you partner with people who
8:48:53already have it. So here's what you
8:48:54could say. Hey, I help businesses
8:48:56implement AI automation. I'd love to
8:48:57offer free AI discovery calls to your
8:48:59clients. No cost to you, no cost to
8:49:01them. You get to look like a hero
8:49:02because you brought in an AI expert and
8:49:04you gave them free value. And if any of
8:49:05them end up wanting to work with me,
8:49:06I'll give you 20% of that project
8:49:08revenue. There's essentially no risk for
8:49:10them. They're not paying you. You're not
8:49:11asking for anything from them upfront as
8:49:13far as money. You're just saying, "Let
8:49:15me make you look good to your clients."
8:49:17So, this works well because when the
8:49:18agency owner says to their client, "Hey,
8:49:19you know, I'm going to bring in this
8:49:20expert to give you a free audit." You
8:49:21instantly have more authority. You're
8:49:23being called in like a consultant.
8:49:24You're the specialist that they are
8:49:25introducing. Compare that to cold
8:49:26outreach when you're a complete stranger
8:49:27in an inbox trying to prove yourself.
8:49:29So, with the Trojan horse method, you
8:49:31borrow credibility from day one and
8:49:33you're getting access into probably a
8:49:35bunch of clients who already trust and
8:49:36already pay and are already warm to
8:49:38spending money on services. So, they're
8:49:39not tire kickers. They're real
8:49:41businesses with hopefully real budgets.
8:49:42Now, when you are able to secure a few
8:49:44yeses, just keep it simple. hop on the
8:49:46discovery call, learn about their
8:49:47business, understand their processes,
8:49:48and tell them where AI could help them.
8:49:49Now, if they hop in the call and say,
8:49:51you know, like, "No, I need to use AI,
8:49:52and I'm glad that we're having the
8:49:53session, but I have no clue where to
8:49:54even start. Tell me what I don't know."
8:49:56What you want to do is help them
8:49:57identify their current constraints. So,
8:49:58a really good question you could ask is,
8:49:59if tomorrow you've got 300 people that
8:50:01wanted your services, what process would
8:50:03break first in [music] that, you know,
8:50:05customer life cycle? And then they'll
8:50:06think about it. They'll run it through
8:50:07in their head, and then they'll say
8:50:08something, and you just dive in on that
8:50:10process. And here's the best part about
8:50:11the Trojan horse method. Even if they
8:50:12don't end up becoming a client, I think
8:50:14it's still a win-win. You got experience
8:50:16doing a discovery call. The business
8:50:17owner got free insights about AI, and
8:50:19the agency looked like a hero for
8:50:20bringing you in. Win-winwin. And some
8:50:22percentages of those calls will turn
8:50:23into paid projects. Remember, it's a
8:50:25numbers game. You need to do 10 times
8:50:27more than you think. Finding partners is
8:50:28the same process as cold outreach.
8:50:30Building a lead list of agencies,
8:50:32consultants, and service providers, and
8:50:33niches that you want to work with reach
8:50:34out with an offer that's hard to say no
8:50:36to. And luckily, that's kind of what
8:50:37this method is. It's an offer with like
8:50:39zero downside for them unless you mess
8:50:41up the discovery call big time. And
8:50:42that's basically the risk that they
8:50:43have. But of course, the upside is
8:50:45they're providing tons of extra value to
8:50:46their clients and they might get 20% of
8:50:48a big project. So if you construct that
8:50:50offer and you hit their inbox or their
8:50:52LinkedIn DMs, most agency owners will
8:50:54actually at least hear you out. So
8:50:56that's the Trojan horse method. Borrow
8:50:57trust, make partners look like heroes,
8:50:59get in front of warm clients without
8:51:00building your own audience. So now that
8:51:02we've talked about those three methods,
8:51:03let me give you a road map. But first, I
8:51:05need to do one thing. I need to flip how
8:51:06you're thinking about this. A lot of
8:51:07people think about this backwards. They
8:51:09spend weeks, maybe even months, building
8:51:10out elaborate systems, perfecting every
8:51:12deal, and then they try to sell it. By
8:51:13that point, they've invested all this
8:51:14time into something that they have no
8:51:16idea anyone actually wants. There's no
8:51:17market validation. So, the smarter thing
8:51:19is to actually flip that order and
8:51:20create an offer around something that
8:51:21you know you could build. Maybe you see
8:51:23other people building it, so you know
8:51:24it's possible. And then start sending
8:51:26outreach as if you have the thing built.
8:51:28See if there's anyone interested. See if
8:51:29people respond. If they do, then build
8:51:31it out. But if nobody bites, then you
8:51:32should pivot to something else and test
8:51:34that offer again before you build it all
8:51:35out. In that case, the only thing you've
8:51:37lost is a little bit of time that it
8:51:38took to write some of those messages.
8:51:40But the lesson is stark. You can have
8:51:41the most technically impressive system
8:51:42in the world. But if you built it in
8:51:44isolation without talking to customers
8:51:45first, you just have an expensive
8:51:46portfolio piece that nobody actually
8:51:48wants. Cold outreach gives you direct
8:51:50primary research from the market before
8:51:52you commit to building anything. People
8:51:53either respond or they don't, and that
8:51:54signal is worth more than any amount of
8:51:56guessing. So, now that we've covered
8:51:57that, let's talk about the four-step
8:51:58framework. Step one is offer first. So,
8:52:00pick a niche, pick a problem, do some
8:52:02research, create an offer around an
8:52:03outcome that you're confident you could
8:52:04deliver. Now, I do think that building
8:52:06simple PC's and demos is fine and
8:52:08probably a good idea just in case on
8:52:09that Loom video you can show them how it
8:52:11works or on a call, but don't pour weeks
8:52:13and weeks into building something once
8:52:14again if you don't know the market wants
8:52:16it. Step two is validate with volume.
8:52:18So, test it with outreach. Aim to get
8:52:20five clients so that you can start
8:52:21getting referrals and case studies. If
8:52:22you're doing manual DMs on LinkedIn or
8:52:24Instagram, maybe aim for at least 100 a
8:52:26day. I know it's aggressive, but that's
8:52:27what it takes to get real data fast. And
8:52:29if you want to take the cold email
8:52:30route, then definitely just invest in
8:52:31some domains. Maybe you could buy
8:52:33domains that are already warmed up and
8:52:34get some inboxes that you can scale up
8:52:36to like 500 emails a day. Cuz keep in
8:52:38mind, even if you're getting a 3% reply
8:52:39rate, that's still pretty good. But
8:52:40here's what separates winners from
8:52:42losers. Because both winners and losers
8:52:43will fail. Everyone hits roadblocks. The
8:52:45difference is that winners understand
8:52:46why they failed and they changed
8:52:48something. Losers keep throwing time and
8:52:50resources into a broken system. So
8:52:52monitor your data every single day.
8:52:53Track your open rate, track your reply
8:52:55rate, track your positive response rate.
8:52:57If something's not working, then figure
8:52:58out why and actually adjust it. Step
8:53:00three is to deliver and then document.
8:53:01So once you do land a client, your job
8:53:03is to overd deliver and then track
8:53:04everything. So before you start, you
8:53:06have to get those baseline metrics. How
8:53:07long does the process take manually?
8:53:08What's the error rate? What are the
8:53:09current KPIs? And then after your system
8:53:11is live, you measure those same numbers
8:53:13again. You show them the before, you
8:53:14show them the after. And that does two
8:53:15things. It first of all proves ROI so
8:53:17they know it was actually a good
8:53:18investment. And the second thing is that
8:53:19it gives you a rockolid case study to
8:53:20use for future clients. And then of
8:53:22course, once you've shown them all that
8:53:23data, ask for the referral. And then
8:53:25step four is to scale with borrowed
8:53:26authority. Now that you've got some
8:53:27results under your belt, start using the
8:53:29Trojan horse method. Reach out to
8:53:30agencies, consultants, and service
8:53:32providers. Offer to do free audits for
8:53:33their clients. Make them look like the
8:53:35hero. Take 20% rev share on any deals
8:53:36that you close. This gets you in front
8:53:38of warm leads without building your own
8:53:39audience. And because you're introduced
8:53:40by someone that they already trust, you
8:53:42have instant credibility. So, I know we
8:53:43covered a ton of information in this
8:53:44video. So, what I did is I threw all of
8:53:46this into a resource guide, which you
8:53:47can access for completely free in my
8:53:49free school community. Link for that is
8:53:50down in the description. And if you want
8:53:51to dive deeper into all this kind of
8:53:53stuff and connect with over 3,000
8:53:54members who are building their own AI
8:53:56automation businesses, then check out my
8:53:57plus community. The link for that is
8:53:59also down in the description.
8:54:02If you can build AI workflows, but you
8:54:04still can't sign a client, then this
8:54:05video is for you because I'm going to
8:54:06show you how to land your first AI
8:54:08workflow client in just 7 days. [music]
8:54:09And this isn't theory. Christian used
8:54:11this exact process and signed his first
8:54:13client in just 5 days. So, in this
8:54:14video, he'll break down what he did and
8:54:16then I'll show you the exact same method
8:54:17step by step so that you can copy it.
8:54:19So, let's get into it. All right, guys.
8:54:21Today I'm here with Christian. He's a
8:54:22member of the community and he's had
8:54:24some huge wins inside the community. He
8:54:26actually got a client in the first 5
8:54:27days of joining and going through the
8:54:29program. So I wanted to bring him in
8:54:30today and just kind of highlight his
8:54:32story a little bit and talk about, you
8:54:33know, some of his big breakthroughs and
8:54:34how he was able to achieve these wins.
8:54:36So Christian, if you want to give
8:54:37everyone a real quick intro, a little
8:54:39bit of your background and, you know,
8:54:40how technical were you before hopping
8:54:41into this AI space?
8:54:43>> Yeah, so uh my name is uh Christian. I'm
8:54:45uh 21 years old out here in Arizona. And
8:54:48for me as a background, I've always been
8:54:50kind of techsavvy, always liked doing
8:54:52like using the tools. But as far as AI
8:54:55itself, I was barely introduced this
8:54:56year. Like it was back in March when I
8:54:58was introduced. So still fairly new.
8:55:00Once I got introduced, I just got
8:55:02obsessed and just started rolling with
8:55:03it. But started this about a couple
8:55:05months ago. Loving the community, loving
8:55:07all the support and all the guidance
8:55:09with it. And it's just it's just been
8:55:10great ever since I joined.
8:55:12>> Awesome, man. Well, yeah. I'm super
8:55:13excited to have you here today and just
8:55:14get to talk about, you know, what got
8:55:15you into this space and how you were
8:55:16able to achieve, you know, these client
8:55:18wins. So, I guess that's the first one
8:55:20to start off with, like what actually
8:55:21was it to you that made you realize
8:55:23like, I want to take advantage of this
8:55:24AI opportunity?
8:55:25>> Once I started getting introduced to it,
8:55:27getting more acquainted with it, I just
8:55:28I love the possibilities and all the
8:55:30different things that you can do with
8:55:31AI. And for me, about it was about 3
8:55:34months ago, I had that light bulb
8:55:35moment. It was like what if I can
8:55:36somehow use this in businesses and help
8:55:39them do that and I can charge them a
8:55:41whatever price but ultimately it'd be
8:55:43solving a rich person problem. It was a
8:55:46really cool moment really when I found
8:55:48it. So
8:55:49>> 100%. Yeah. So once you kind of found
8:55:50out you wanted to go down the AI
8:55:52automation space and I think n was was
8:55:54your big tool as well. What were you
8:55:56doing and like how were you trying to
8:55:58find clients before joining the
8:55:59community? Funny enough, um I actually
8:56:01didn't start using NAN about up till
8:56:04about a couple weeks ago. But before
8:56:06that, my other community that taught me
8:56:08a way to where we can scrape leads off
8:56:10of LinkedIn, find their emails, generate
8:56:13email bodies for them using AI and then
8:56:15loading it into a website called
8:56:16instantly AI where you can just in uh
8:56:19emails like cold emails. It was a great
8:56:21thought to begin with, right? It was a
8:56:23great idea. I was doing a bunch of
8:56:25emails, like 450 a day, but uh it was
8:56:28like a crazy open rate, but nobody was
8:56:30buying. Uh just a lot of shots in the
8:56:32dark. Didn't really know how to
8:56:34structure the emails. I I didn't know
8:56:35how to write copy or anything. So, well,
8:56:37I was basically just trying anything and
8:56:38everything I could. Joining Facebook
8:56:40group, talking to my contacts that I
8:56:42have, other business owners, uh see if
8:56:44they can I can do it for them or if I
8:56:47can just if they have anybody that I can
8:56:49do it for. Just really no proper
8:56:51guidance or no, uh set direction. just
8:56:53kind of shooting off different lanes
8:56:54hoping that one of them was going to
8:56:55hit.
8:56:56>> I think the reason what like cold email
8:56:57is really successful and cold outreach
8:56:58in general I think once you're ready to
8:57:00scale but like to get your first client
8:57:02it's really tough to do that because
8:57:04people are just like I don't know who
8:57:05this guy is and it's like he has nothing
8:57:07really to show me yet. Why would I kind
8:57:09of give him or her my business? So how
8:57:11did then your approach kind of change
8:57:13once you were in the community and
8:57:14started to you know interact with the
8:57:16people in there in the course material?
8:57:17>> Biggest one by far was how to position
8:57:20myself. Right. We're not selling
8:57:22workflows. It's the brand behind it.
8:57:24Scale or get your time back, whatever
8:57:26the goal is for your business.
8:57:27Positioning myself as a proper guide,
8:57:30not a template salesman. That was a
8:57:33really key thing for my for me in my
8:57:35process. And it's it's just it was black
8:57:37and white difference and it helped me
8:57:38out a ton. Within days of me doing that,
8:57:42I posed my first client, which was
8:57:43insane.
8:57:44>> Yeah. Know, I was just looking at our
8:57:45our DM on school right before we hopped
8:57:47on and I remember, you know, you joined
8:57:49the community. We we had a quick chat,
8:57:51got you access to the course, and then 5
8:57:53days later, you messaged me. It's like,
8:57:55I just closed my client for like what
8:57:56was it? Was it 1,500 or something like
8:57:58that?
8:57:58>> Uh yeah, 1500, 1500. And we actually we
8:58:00upped it recently. So, we're doing uh
8:58:032,000 now. And then I think it's going
8:58:04to over time build and build and build
8:58:06from there as I get better, too. Yeah,
8:58:08it was it was 1500 starting out. So, it
8:58:10was it was pretty cool.
8:58:11>> So, you made a post about that win,
8:58:13right? And then I pinned it in the
8:58:14community and then did you say someone
8:58:16reached out to you from that post and
8:58:17you got another client off that? So off
8:58:19that first one, I was getting a lot of
8:58:21praise or a lot of traction. I was like,
8:58:22"Oh, this is cool." And then you know
8:58:24the the gem as well. So that was that
8:58:26was even better. Literally three days
8:58:28later on that was that Friday I posted
8:58:30that Sunday. Another member in the
8:58:31community reached out. He was a newer
8:58:33member and said, "Hey, I'm looking for
8:58:34to hire uh he called it an AI automator.
8:58:37Um and then it went from getting, you
8:58:39know, got the first client, engaged with
8:58:41the community, got the second client,
8:58:43and now I'm able to actually do this
8:58:45full-time and not even have to worry
8:58:47about work." So, which is great.
8:58:49>> Um, and it sounds like your kind of
8:58:51biggest light bulb moment so far from
8:58:52the community has been just the idea of
8:58:54the way you position yourself and um,
8:58:56probably how you conduct yourself on
8:58:58those initial discovery and sales calls.
8:59:00So, is there anything specifically you
8:59:01can think about where you know hopping
8:59:03on discovery calls or trying to talk to
8:59:04business owners before having that kind
8:59:06of like light bulb moment, the way that
8:59:08you would like speak about AI or speak
8:59:09about automations has shifted?
8:59:11>> Yeah, definitely. Uh, before I would I
8:59:13didn't really I didn't have a process. I
8:59:14would walk in there and say, "Hey, this
8:59:16is what I do. I can build template for
8:59:18you if you need and it can solve this.
8:59:20Do you want it? And that was that was
8:59:22kind of it. But going through the uh
8:59:24lessons and actually showing all right
8:59:26this is how you should ask the
8:59:27questions. This is how you find the pain
8:59:30and find ways that we can actually
8:59:31leverage AI to actually help their
8:59:34business and and again I'm an actual
8:59:36guide in their process in a in the
8:59:38journey of AI cuz it's it can be very
8:59:40overwhelming. So once I started
8:59:42positioning myself there it was a
8:59:44complete 180. Yeah. It was just it was
8:59:47yeah complete night and day difference.
8:59:48I'm going to do that with all my
8:59:49processes moving forward. But that one
8:59:51specifically really helped me out a lot.
8:59:53>> So besides the the way that you shifted
8:59:55your perspective as far as like you know
8:59:56let's think about your business and let
8:59:58me help you diagnose problems rather
8:59:59than just here's a template. What else
9:00:01do you think it made them say yes?
9:00:02Because at this point there's there's
9:00:03this mental battle of like how much
9:00:05should I charge or should I do it for
9:00:06free or like what what proof do I
9:00:08actually have? How do I get someone to
9:00:09say yes? So, what else do you think it
9:00:11was about you and about the way that you
9:00:13spoke to this this person that got them
9:00:14to ultimately give you their business?
9:00:17>> Well, for the second client, I mean, he
9:00:19kind of knew my situation a little bit
9:00:20cuz he saw the post, he read it, and
9:00:22then that's when he reached out. I was
9:00:23just very very honest with them and
9:00:25upfront about like, listen, this is um
9:00:27obviously you know the value of the
9:00:28product. So, uh, I went in there, I was
9:00:30very, very open. I was very honest and
9:00:32transparent about me, where I'm at on my
9:00:34journey and my goals for him and was
9:00:37mainly just trying to focus on building
9:00:39that relationship.
9:00:40>> Yeah. And now, I mean, the rest is
9:00:41history. So,
9:00:42>> yeah, it's awesome because now that your
9:00:43foot's in the door, there's so much
9:00:45opportunity for for more growth and and
9:00:47for bigger projects. And I've always
9:00:48said, you know, you'd rather have one,
9:00:50let's just say 10K a month client rather
9:00:52than having five 2K a month clients. You
9:00:53just have an opportunity to go way
9:00:55deeper. And that just leads into better
9:00:57referrals, better case studies, more
9:00:58opportunities to win more business,
9:01:00things like that. Do you want to walk us
9:01:01through the actual solution that you've
9:01:04provided to this first business or that
9:01:05you're working on and and also how you
9:01:08got to that price point and what sort of
9:01:10like that conversation looks like?
9:01:11>> The one that I built, so I was trying to
9:01:13build a system for updating clients. A
9:01:16huge pain point that we were having
9:01:18simply just communication. Like a lot of
9:01:20clients would call up like, "Hey, uh,
9:01:21what's this? I don't know what they mean
9:01:23by this. when is this going to get
9:01:24scheduled? Or nobody's talked to me.
9:01:26Nobody's reached out. I don't know
9:01:27what's going on with my project or
9:01:29sometimes they're just simple FAQs. They
9:01:31just want to know, hey, what what is
9:01:33shot create? And they can simply Google
9:01:34it, but they're going to call us first.
9:01:36I focus on building a workflow for that.
9:01:38And it actually I got inspiration for it
9:01:41from one of your YouTube videos.
9:01:43Actually, it was your AI assistant. So,
9:01:45I took that kind of blueprint and used
9:01:47it to build out a I called it like a PM
9:01:50assistant. So when the project manager
9:01:52is out on site, he can just simply talk
9:01:54to the bots, it'll, you know, engage
9:01:56with him, do what and execute whatever
9:01:58tasks are needed, whether it's sending
9:01:59clients an update email, logging an
9:02:01update into the CRM, little things like
9:02:03that. He loved it. He completely bought
9:02:05in. Uh I'm still improving it to this
9:02:07day. It's not perfect yet, but it's um
9:02:10it's constant refining. We're meeting
9:02:11every week to get it better and better
9:02:12and better, but that's the one I started
9:02:14with.
9:02:15>> I want to ask you now compared to let's
9:02:17just say 3 weeks ago, how much more
9:02:19confident do you feel? and just your
9:02:20ability to talk to people, show them the
9:02:22value, deliver on solutions, price your
9:02:24work, all that kind of stuff.
9:02:25>> A lot more confident. Before I had no
9:02:27idea how to do any of that. I was just
9:02:29kind of got to go kind of go in and just
9:02:31give like a set base price point for
9:02:33whatever workflow that I build. Um, and
9:02:36again, before I was just trying to sell
9:02:37templates basically. Now I feel a ton
9:02:41more confl confident knowing what my
9:02:43offer is. um having a clear offer for
9:02:46them that they can understand and being
9:02:48able to properly communicate what it is
9:02:50that I do and how it can benefit their
9:02:52business.
9:02:53>> That's awesome to hear, man. I mean,
9:02:54it's a cliche, but it's so true. Just
9:02:56the idea of like surrounding yourself
9:02:58with like-minded people, like what that
9:02:59really does to you, because I know like
9:03:01right before we started recording, we
9:03:02were just talking about how there's just
9:03:04so much noise and there's so much great
9:03:06education everywhere. But when you get
9:03:08stuck in that world where you're just
9:03:09watching, watching, watching, and you're
9:03:11getting like a scattered bit of
9:03:12information, and you don't have like a
9:03:14structured path, it just makes it very
9:03:16overwhelming, of course, but also
9:03:18confusing.
9:03:18>> Yeah, man. I wanted to kind of close off
9:03:20here if you've got any any final
9:03:22thoughts or where could people find you
9:03:23if they want to get in touch with you?
9:03:25>> Um, yeah. Yeah. I mean, it would just be
9:03:27uh my LinkedIn mainly cuz I don't really
9:03:29have a website right now or through the
9:03:31community. I think that's the the best
9:03:32way possible. I'm in both of the
9:03:33communities, regular one and then the
9:03:35plus community as well. I'm mainly in
9:03:37there more than anything [laughter] I
9:03:38would say. Just doing the same thing,
9:03:40just connecting with people, trying to
9:03:42provide value, and then seeing what else
9:03:43that I can learn. So,
9:03:44>> it's really inspiring to hear your
9:03:45story, man. 21 years old, started less
9:03:47than a year ago in this space. And and
9:03:49you guys also just heard him say
9:03:50basically doesn't have a website yet.
9:03:52He's new. He he told clients, "Hey, I'm
9:03:54new. I'm learning, but I want to help
9:03:55you." And he's been able to secure some
9:03:57pretty crazy wins. So, hopefully that's
9:03:59all very motivating to all of you guys
9:04:00that are watching this that are trying
9:04:02to get to a spot where where Christian
9:04:04now is. So, thanks so much for coming on
9:04:06today, Christian. I really enjoyed
9:04:07getting to know you and and getting to
9:04:08hear a little bit about your journey.
9:04:10Yeah, just keep crushing it. Best of
9:04:11luck, man.
9:04:11>> Thank you, man. I appreciate it. I'm
9:04:12going to try my best.
9:04:13>> So, now let's look at the exact method
9:04:15that I gave Christian that landed him
9:04:16his first client in just 5 days. By the
9:04:18end, you'll have everything that you
9:04:19need to go sign your first AI client.
9:04:21All right. So, before I walk through
9:04:23that actual 7-day plan, I just want to
9:04:24slow this down and ground it in what
9:04:26actually happened with Christian because
9:04:28this is where most beginners go wrong.
9:04:29If you think back to what Christian was
9:04:31doing before he joined the community,
9:04:32it's exactly what I see almost everyone
9:04:34doing when they get started, which is
9:04:35high volume, tons of cold outreach with
9:04:37no soul and no positioning, using tools
9:04:39or automations to scrape LinkedIn,
9:04:41pulling emails, using AI to write
9:04:43generic copy, loading it all into
9:04:44instantly, and then blasting out like 4
9:04:46or 500 emails a day. Now, on paper, this
9:04:48feels really productive because you're
9:04:49sending out so many messages. But in
9:04:51reality, you're just another random
9:04:52person who's selling AI templates with
9:04:54no proof, no trust, and no real
9:04:56connection. And at that point, you're a
9:04:57commodity. And when you're a commodity,
9:04:59it turns into a race to the bottom on
9:05:00price, and you'll almost always lose.
9:05:02Because there's always going to be
9:05:03someone out there who will offer your
9:05:04service for a cheaper price. So, what
9:05:06changed everything for Christian was not
9:05:08better tools or more automation. It was
9:05:09the mindset shift. He stopped thinking
9:05:11like someone who was trying to sell
9:05:12workflows and started thinking like
9:05:14someone who was leading with outcomes.
9:05:15He positioned himself as a partner, not
9:05:17a template salesman. So, instead of
9:05:18saying things like, "Here's what I
9:05:19built. Here's this automation. Here's
9:05:20this template I can, you know, customize
9:05:22for you," he was saying, "Here's how I
9:05:23can help your business do X." And he was
9:05:25targeting a specific painoint. So, that
9:05:27one shift is the entire reason that this
9:05:28worked in 5 days instead of never
9:05:30working at all. Now, before we get into
9:05:31the 7-day plan, I want to talk about
9:05:33three psychological hurdles that
9:05:34everyone goes through. I went through
9:05:35them. Christian went through them.
9:05:37You're probably going through them right
9:05:38now. And the first one is imposttor
9:05:39syndrome. Everyone will feel this at the
9:05:41beginning. And honestly, I still feel
9:05:42this sometimes right now. Thoughts like,
9:05:44"Do I even have the right to charge for
9:05:46this?" Or, "I don't want to feel
9:05:47scammy." Or, "What if I'm not ready
9:05:48yet?" And that feeling will never fully
9:05:50go away. But you do learn how to work
9:05:51through it. So, the key is just to not
9:05:53overpromise because that's when you run
9:05:54into trouble. Both Christian and I did
9:05:56the same thing early on when we were
9:05:57getting started. Literally saying, "Hey,
9:05:59I'm new to this. I'm obsessed with it.
9:06:00I've been learning a lot. I've been
9:06:02building a lot and I want to help you
9:06:03with this problem." And that type of
9:06:04honesty makes you sound like a real
9:06:05human. Like I said, especially early on
9:06:07when you're doing your first one or two
9:06:09clients for free or very cheap. That's
9:06:10not a weakness. It's low risk for them
9:06:12and it gets you experience, which is the
9:06:14only thing that should actually matter
9:06:15to you right now. Your goal should be to
9:06:16overd deliver. Prove to yourself, and
9:06:18also prove to the client that you can
9:06:19actually do this. And that imposter
9:06:20syndrome starts to fade because now you
9:06:22have actual experience to talk about and
9:06:23you have actual results that you've
9:06:25driven. Now the second hurdle is about
9:06:26pricing. People get stuck here way too
9:06:28early. They obsess over retainers and
9:06:29monthly pricing and they want to hit
9:06:31that 10K per month figure that
9:06:32everyone's going for. But if you're
9:06:33trying to maximize revenue before you
9:06:35even ever delivered value or got
9:06:37experience, then you're going to fail.
9:06:38So the thing that I want you guys to
9:06:39lock into your brain is that trust comes
9:06:41before retainers. Asking someone for a
9:06:43retainer right away would be like asking
9:06:44someone for a referral right away before
9:06:46you even have any sort of partnership.
9:06:48It's very hard to get someone on a
9:06:49monthly agreement if they've never
9:06:50worked with you before. So, what should
9:06:52matter to you first is just delivering
9:06:53value and earning trust and then you can
9:06:55win a retainer later. That's why I
9:06:56always talk about valuebased pricing.
9:06:57And I made a full video on this and I'll
9:06:59link it up there if you want to check it
9:07:00out. But anyways, you want to get your
9:07:01foot in the door, solve a real problem,
9:07:02and clearly show the value that you
9:07:04delivered. And once that trust is built,
9:07:05then talking about a retainer becomes
9:07:07super easy and much more natural. Like I
9:07:09said earlier, the same rule applies for
9:07:10testimonials and referrals. You can't
9:07:12ask for those until you've actually done
9:07:14something. So, don't think about a
9:07:14retainer as step one. Think of it like
9:07:16something that you've earned once you've
9:07:18built a relationship. And the third
9:07:19hurdle is just rejection. This one
9:07:20scares people more than it should. When
9:07:22you're new, you are going to get
9:07:23ignored. People will say no. Some won't
9:07:25respond at all. That's completely
9:07:26normal. The important part is what you
9:07:27actually do with that rejection. Every
9:07:29rejection is feedback. It's more data
9:07:30for you to learn from and win from
9:07:32because the reality is everyone gets
9:07:34rejected. So the difference between
9:07:35winners and losers is that the winners
9:07:37pay attention to why they were rejected
9:07:38and they actually fix something. They
9:07:40change something about it. And losers
9:07:41just get rejected over and over and they
9:07:42don't change anything. So if someone
9:07:44ignores you, ask why. Was my message
9:07:45unclear? Was it too long? Was it vague?
9:07:47Did they even have a reason to want to
9:07:49care to respond to me? And then you fix
9:07:50that. And then your next attempt gets
9:07:52better. Okay, so those are the three
9:07:53hurdles. Now, let's talk about actually
9:07:55breaking down the 7-day framework and
9:07:57connecting it back to what Christian did
9:07:58because this is where people tend to
9:07:59over complicate things. So, the first
9:08:01thing I want to make very clear is this.
9:08:02Christian did not start by picking some
9:08:04super ultra specific niche and having
9:08:06every problem figured out and then
9:08:07blasting out cold emails. He started
9:08:08warm and that matters way more than
9:08:10people think. warm outreach, referrals,
9:08:11and known contacts are just
9:08:13statistically shown to convert way
9:08:14better than cold outreach, because
9:08:15there's already trust there, even if
9:08:17you're borrowing trust from a friend of
9:08:18a friend. That's why it works so much
9:08:20faster, and it feels way less awkward,
9:08:22especially as a beginner. So, big
9:08:24picture, the flow looks like this. You
9:08:25start warm, you stay fairly broad, you
9:08:27talk to people, you notice patterns, and
9:08:28then you niche down later. Cold outreach
9:08:30comes last when you're ready to scale,
9:08:32not first when you want to get your
9:08:33first client. So day one is all about
9:08:34setting a loose direction. Not locking
9:08:36yourself into one super hypersp specific
9:08:38niche, but just think of it as like a
9:08:40working hypothesis, not a lifelong
9:08:42decision. So instead of saying, "I'm the
9:08:43AI automation guy for dental clinics in
9:08:45Georgia, losing time with onboarding,"
9:08:47you're just going to say something much
9:08:48simpler like, "I help small businesses
9:08:50automate boring, repetitive tasks with
9:08:52AI." That's enough to get conversations
9:08:53started. At this stage, you should also
9:08:55have a short menu of example problems in
9:08:57your head, example use cases in your
9:08:59head, so that you're not just speaking
9:09:00in theory. So things like, you know, I
9:09:01know how to automate lead follow-up or
9:09:03intake forms or data syncing between
9:09:05CRM. And like I said, you're not
9:09:06committing to any of these yet, but
9:09:07you're just testing to see what
9:09:08resonates. On day one, you also want to
9:09:10build what I would call a trust map. So
9:09:12open up a simple Google sheet and list
9:09:13out people where trust already exists.
9:09:15Your goal for day one would be to write
9:09:17down 20 people that you can reach out
9:09:18to. Friends or family who run
9:09:20businesses, past co-workers, managers,
9:09:22clients, people who you know from
9:09:23communities, Slack groups, Discords, or
9:09:25online spaces, and the second degree
9:09:26connections. So, friends of friends, for
9:09:28each of these people, just note what
9:09:30kind of business they're in or what kind
9:09:31of industry they're in, how well you
9:09:32know them, whether they could be a
9:09:34potential client or an intro to one, or
9:09:36just someone who can give you some
9:09:37insight. And when you actually start to
9:09:38write this stuff down and think through,
9:09:40you know, your list of contacts, it
9:09:42makes everything feel a lot less random
9:09:43and a lot less overwhelming. Now, days
9:09:45two and three are about having 5 to 10
9:09:47warm, low pressure conversations. These
9:09:49are not sales calls. You're not
9:09:51pitching. You genuinely just need to
9:09:52understand where repetitive work eats
9:09:54time in their business or in their daily
9:09:55workflow. So, when you're reaching out
9:09:57to them, you should just sound like a
9:09:58curious entrepreneur, not a salesperson.
9:10:00So, something as simple as saying, "Hey,
9:10:02I'm trying to start a business where I
9:10:03help other businesses automate
9:10:04repetitive work with AI. I'm not trying
9:10:06to sell you anything. Could I just ask
9:10:07you a few questions about where things
9:10:09feel manual or annoying in your
9:10:10day-to-day?" And once you start having
9:10:11these conversations, write down all the
9:10:13key insights from that call because
9:10:14you're going to need to look at these
9:10:15later. And if you truly cannot think of
9:10:17anyone in your network who might own a
9:10:18business or be an executive at a
9:10:20business or in a position to buy or
9:10:22anything like that, then you still don't
9:10:23have to jump straight to cold outreach.
9:10:24you just reach out and ask something
9:10:25like, "Is there anyone you know who this
9:10:27might be helpful for?" or "Is there
9:10:28anyone you know who this might benefit?"
9:10:30And so that little like wording shift is
9:10:31huge because no one really wants to sell
9:10:33to their friends and you're not selling
9:10:34to your friends. You're just asking if
9:10:35they know anyone who might benefit from,
9:10:37you know, the kind of stuff that you're
9:10:38talking about. It feels way more
9:10:39natural, less uncomfortable, and it
9:10:41still lets you kind of borrow some
9:10:42trust. You also could jump on a site
9:10:44like Upwork, but once again, that's
9:10:46pretty much going to be similar to cold
9:10:47outreach because you haven't yet built
9:10:49trust with those leads and you're just a
9:10:50name and a profile picture. So, after
9:10:52you've had those conversations, day four
9:10:53and five are where you turn those
9:10:55conversations into a tiny pilot. So,
9:10:57look back at your notes, pick the person
9:10:58with the clearest, most painful,
9:11:00repetitive task. You're looking for the
9:11:01clearest pain point here. And then you
9:11:03propose something very small and very
9:11:05low risk for that prospect, which is
9:11:07typically a free workflow. You build it.
9:11:08You measure one simple outcome. And then
9:11:10you decide together if it's worth
9:11:11expanding. So, you can reach back out to
9:11:13them after that initial conversation and
9:11:14say something like, "Hey, I'd love to
9:11:16build you a small automation that
9:11:17tackles X painoint as a free pilot. My
9:11:19goal here is just to prove that these
9:11:20workflows actually can save you time.
9:11:22And in return, I'd just ask for some
9:11:23honest feedback because that's going to
9:11:25help me learn and grow. Now, days five
9:11:26and six are about actually building that
9:11:27tiny MVP. So, this doesn't have to be
9:11:29complex. In fact, the simpler the
9:11:30better. Your goal isn't to impress them
9:11:32with all of these API calls and all this
9:11:34advanced tech. Your goal is just to show
9:11:35them I saved you time. I saved you, you
9:11:38know, confusion here. And while you do
9:11:39this, just pay attention to the language
9:11:40they use, how they describe the problem,
9:11:42how they describe that win. The language
9:11:44is gold later. Okay. Now, day seven is
9:11:46where you see if your hard work paid
9:11:48off. The goal of this day is super
9:11:49simple. You're not trying to hard close
9:11:51anyone. You're just trying to figure out
9:11:52with the potential client what the next
9:11:54logical step is based on how the pilot
9:11:56went. So, the first priority would be to
9:11:57see if you can keep helping that same
9:11:58client, either by maintaining what
9:12:00you've built, or by expanding on it
9:12:02functionality-wise. So, you can lay out
9:12:03clear, low pressure options. You explain
9:12:05that there are a couple ways that you
9:12:06could keep helping. The first option is
9:12:07maintenance. This is you sticking around
9:12:09to make sure that the automation keeps
9:12:10working, handling fixes if something
9:12:11breaks, and making small tweaks as their
9:12:13tools or processes change. And the other
9:12:15option is expansion. So, this is where
9:12:16you build on top of what's already
9:12:18working. Usually, by adding one or two
9:12:20concrete features that you notice during
9:12:21the build that would make the outcome
9:12:22even stronger. The key here though is
9:12:24just to communicate these two actual
9:12:26offers very, very simply. So, a
9:12:27maintenance offer could be like, hey,
9:12:28I'll maintain this for the next few
9:12:30months, so if anything breaks or if you
9:12:31want small changes, I can handle it. And
9:12:32an expansion offer might be, hey, we
9:12:34could turn this into a slightly bigger
9:12:35system by adding this and this so that
9:12:38the result is x more reliable. And then
9:12:39you can ask if they want to scope out
9:12:41that and then you can send over a basic
9:12:42proposal, nothing fancy. During this
9:12:44conversation, you could also naturally
9:12:45surface more work by asking thoughtful
9:12:47questions. For example, you could say
9:12:48something like, "Hey, while while I was
9:12:50building out this workflow, I noticed
9:12:51this other process that's related. Would
9:12:53have helped if that was automated, too.
9:12:54Or now that you've seen what this
9:12:56workflow can do, what's the next thing
9:12:57that you would like to get off your
9:12:58plate?" These types of questions show
9:12:59that you are a partner and that you're
9:13:01keeping things collaborative rather than
9:13:03being salesy. And then, if they're happy
9:13:04with the pilot, but they don't
9:13:05necessarily want to continue working
9:13:06with you right now, that's completely
9:13:08fine. At that point, your goal is to
9:13:09just lock in some proof, not push more
9:13:11work on them. So, you can ask if they'd
9:13:12be open to a short video testimonial
9:13:14explaining how the workflow helped. Only
9:13:15after that, and only if it feels
9:13:16natural, you can then mention if they
9:13:18ever think of someone that might also
9:13:20benefit from AI workflows like this,
9:13:22then an intro would be super
9:13:23appreciated. But, of course, if the
9:13:24pilot did not go well, you don't try to
9:13:26sell anything. You don't ask for a
9:13:27testimonial. You just take that
9:13:28feedback, you learn from it, and you
9:13:29improve next time, and you start the
9:13:31process again of warm conversations and
9:13:33offering a free pilot because of course
9:13:35ultimately what you're optimizing for
9:13:36right now is experience, proof, not
9:13:38money. So, that's day seven. The entire
9:13:40focus is on deepening the relationship
9:13:41first. Maintenance or expansion comes
9:13:43before testimonials. Testimonials come
9:13:45before referrals. And if it doesn't
9:13:46work, you learn and you move on without
9:13:48forcing anything. Now, here's the most
9:13:49important part. This 7-day cycle is not
9:13:51a onetime thing. [music] It's a loop
9:13:52that you're going to have to run through
9:13:54multiple times. But each loop gives you
9:13:55more data, more confidence, better
9:13:57language, clearer positioning, and only
9:13:58after you've done a few cycles of this
9:14:00does it make sense to then build lead
9:14:01lists and apply the same framework to
9:14:03colder outreach because you actually
9:14:04have some proof behind your name now. So
9:14:06then when you do it, it's no longer just
9:14:08feeling like guessing and feeling random
9:14:09and overwhelming because you have a
9:14:11system that works. And that's the exact
9:14:12path that Christian followed. Start
9:14:13warm, lead with outcomes, earn trust,
9:14:15deliver value, and then just repeat. And
9:14:17if you do this two or three times, you
9:14:18were no longer a beginner guessing in
9:14:19the dark. You can see that Christian
9:14:21actually got his second client basically
9:14:22just by posting his win in the community
9:14:24and then someone reached out because
9:14:25they needed a developer that could work
9:14:26with clients. So I know that we covered
9:14:28a lot of information in today's video.
9:14:29So what I did is I threw all of this
9:14:31into a practical resource guide that you
9:14:33can access for completely free in my
9:14:34free school community. The link for that
9:14:35is down in the description. Now, if you
9:14:37did find this video valuable, signing
9:14:38your first client is only one part of
9:14:40making money with AI. There's actually
9:14:42several components involved, and I cover
9:14:43all of them in detail in this video. So,
9:14:45if you want to make money with AI in
9:14:462026 and click there to check it out.
9:14:48Otherwise, if you enjoyed the video or
9:14:49you learned something new, then please
9:14:50give it a like. It definitely helps me
9:14:52out a ton. And as always, I appreciate
9:14:53you guys making it to the end of the
9:14:54video. I'll see you on the next one.
9:14:56Thanks, everyone.
9:15:00Everyone's been asking how to actually
9:15:01price AI workflows. What's too cheap?
9:15:03What's too expensive? and how do you
9:15:04charge enough without losing clients?
9:15:06So, in this video, I'll break down
9:15:07exactly how to price your workflows step
9:15:09by step. And at the end of the video,
9:15:11I'll give you guys a practical breakdown
9:15:12of a real example of how I priced one of
9:15:14my workflows. So, let's get [music] into
9:15:15it. Before we get into numbers or
9:15:17frameworks, you need the right mindset
9:15:18around pricing. Most beginners price
9:15:20their workflows based on the time it
9:15:21takes or the amount of effort that they
9:15:22put in. But businesses don't pay for
9:15:24your time. They pay for outcomes. When
9:15:25you build an AI workflow, automation
9:15:27system, infrastructure, whatever you
9:15:29want to call it, you're usually doing at
9:15:30least one of three things for a
9:15:32business. saving them money, saving them
9:15:33time, or reducing human error. That's
9:15:35where the real value is, and that's what
9:15:37your price needs to be tied to. This is
9:15:38called valuebased pricing. You're not
9:15:40charging for the inputs. You're charging
9:15:41for the return on their investment. And
9:15:43when you start thinking this way,
9:15:44everything changes. Instead of asking
9:15:46yourself, how long will this build take
9:15:47me? You start asking, "How much time or
9:15:49money will this save the business every
9:15:51week? How much more can they generate
9:15:52because of this system?" Now, I know it
9:15:54can feel awkward when you're new.
9:15:55There's always that weird dance of like,
9:15:57who's going to throw out a number first?
9:15:59And realistically, a lot of clients
9:16:00right now genuinely have no idea what
9:16:02this kind of stuff should cost. But
9:16:04here's the part that you need to nail.
9:16:05Whatever number that you say, you should
9:16:06be able to show them exactly how you got
9:16:08there. You should almost assume that
9:16:10after you tell them your price that
9:16:11they're going to say, "Can you tell me
9:16:12exactly how you got to that number, you
9:16:14need to be able to confidently walk them
9:16:15through that math." Because when you can
9:16:17explain the ROI, not only do you build
9:16:19trust, but you come across as a
9:16:20long-term AI thought partner, which is
9:16:22what businesses desperately need right
9:16:24now. And when you explain that clearly
9:16:25and you anchor that price on real
9:16:27metrics, suddenly the price doesn't feel
9:16:29like an expense anymore. It feels like
9:16:30an investment that will pay for itself
9:16:32because it will. So once you've got that
9:16:33mindset locked in, the next step is
9:16:35learning the actual pricing models that
9:16:37you can use. So before we dive deeper, I
9:16:39want to be clear about something. I've
9:16:40tried almost every pricing model that
9:16:42you can think of. I've charged a fixed
9:16:43fee and delivered a JSON file. I've
9:16:45built hourly. I've sold bundles of
9:16:46hours. I've done monthly retainers with
9:16:47no set scope. I've done retainers with
9:16:49strict scopes. I've literally
9:16:50experimented with tons of different
9:16:51models. And the reason I'm telling you
9:16:53this is because I don't believe that
9:16:54there's currently an industry standard
9:16:56way to price AI automation and
9:16:58implementation services yet. Nobody has
9:16:59agreed on the right way to price
9:17:01workflows. Everyone's still trying to
9:17:02figure it out. But after trying all of
9:17:04these different models, there are
9:17:05definitely two that I think work the
9:17:06best and they work together. So let me
9:17:08explain what I mean by that. They're
9:17:09simple, they build trust, and they're
9:17:10the easiest to scale. So the first model
9:17:12is valuebased pricing. If I were
9:17:14starting from scratch today, this is the
9:17:15model that I would start with.
9:17:16Valuebased pricing is where you charge
9:17:18based on the business impact like we
9:17:20talked about, not the hours that you
9:17:21spend building. Anchoring your price
9:17:22around time save, money save, or
9:17:24efficiency gain. This is the easiest way
9:17:26to get your foot in the door with a new
9:17:27client because the math is simple and it
9:17:29builds trust. What's interesting about
9:17:30valuebased pricing is that you can build
9:17:32the exact same system for two different
9:17:34businesses and you can charge them
9:17:35completely different prices because
9:17:36value is not universal. If you tried to
9:17:38sell me a bottle of water right after
9:17:40I've run 5 miles in the desert, I would
9:17:42pay you a lot more than someone who's
9:17:43been sitting inside air conditioned
9:17:45office all day. It's the same product,
9:17:47but it's different value. And that's
9:17:48kind of the idea behind this value based
9:17:50pricing. So, you price based on what the
9:17:51solution is worth to that business. Now,
9:17:53once you deliver one or two projects,
9:17:55you build some trust, and you show that
9:17:56your systems work. This is where you
9:17:58start moving into longerterm
9:17:59engagements, which is the next model,
9:18:01monthly retainers. A monthly retainer is
9:18:02when a client pays you a predictable
9:18:04monthly fee in exchange for ongoing
9:18:06access to your expertise [music] and a
9:18:07defined level of service. In the
9:18:09consulting world, these can range
9:18:10anywhere from 1,500 to 15,000 or more
9:18:13per month. And it's usually within a 3,
9:18:156, or 12 month term. And here's why
9:18:17retainers matter. Clients get
9:18:18predictable costs and priority access
9:18:20[music] and consistent improvement. You
9:18:21get stable income and a long-term
9:18:23partnership instead of chasing one-off
9:18:24builds. And honestly, you'd much rather
9:18:26work with one $15,000 per month client
9:18:28than five $3,000 per month clients. Even
9:18:30though the total revenue is the same,
9:18:32you'll have less chaos, less stress, and
9:18:33you get to build a real relationship.
9:18:35You get to put all of your time and
9:18:36energy towards that one business. You
9:18:38get to overd deliver, drive real
9:18:39results, which is going to increase the
9:18:40chances that you're able to win more
9:18:42business in that organization. So, that
9:18:44will lead to higher CLV or LTV, customer
9:18:46lifetime value, and potentially more
9:18:47referrals, which is more business that
9:18:49is also warm and there's more trust,
9:18:51[music] and they're easier to close.
9:18:52Now, retainers do come in different
9:18:54styles. They could be based on hours,
9:18:55based on deliverables, or hybrid.
9:18:57[music] Personally, I would recommend
9:18:58sort of a milestone-based retainer.
9:19:00Hours make you look like a freelancer.
9:19:01Milestones position you as a consultant,
9:19:03an AI partner. But a big question I get
9:19:05when I kind of talk about this is, how
9:19:07do [music] I keep using valuebased
9:19:08pricing inside a retainer? At this
9:19:10stage, your goal kind of shifts because
9:19:12you're not trying to squeeze maximum
9:19:13value and money out of each individual
9:19:14[music] project. Your priority becomes
9:19:16building trust, stacking wins, and
9:19:18increasing the retainer over time as
9:19:19your impact [music] scales, almost like
9:19:21you're trying to work your way into sort
9:19:22of a chief AI officer type of role.
9:19:24Plus, as these systems get complex, it's
9:19:26harder to deliver, test, and refine
9:19:28everything in a single month. In fact,
9:19:30that's actually just [music] really
9:19:31unrealistic, especially if you want the
9:19:33system to be fully QA before production
9:19:35and then in production, taking time to
9:19:36iterate and constantly refine. [music]
9:19:37So, what you'll probably find is that
9:19:39you're naturally going to be making more
9:19:40total revenue on this model anyways. And
9:19:42it's a lot more predictable. So, the
9:19:44less time and energy that you spend on
9:19:45worrying about where your next dollar is
9:19:47coming from, the more time you spend
9:19:48making a real impact. And that opens the
9:19:49door for higher retainers or even
9:19:51performance bonuses tied to key business
9:19:53metrics. [music] And don't forget,
9:19:55you'll also be maintaining the previous
9:19:57systems, updating things when new models
9:19:59come out, fixing bugs, making sure that
9:20:00you're, you know, staying on top of any
9:20:02updates, and just keeping everything
9:20:04aligned with the original scope. That
9:20:06alone justifies ongoing fees. So, the
9:20:08last thing I'll say is this. We're still
9:20:09early. Businesses need enablement,
9:20:11audits, education more than ever. They
9:20:13don't know what's possible, and they
9:20:14don't know where AI fits into their
9:20:16operations [music] even though they know
9:20:17they need to start using it. So, if
9:20:18you're already that person who's been
9:20:20wrapped up in their ecosystem, delivered
9:20:21real value for them, they'll happily
9:20:23trust you to guide them through that
9:20:24process, too. And if you guys want, I
9:20:25can make a full breakdown video on
9:20:27audits, enablement, and actually
9:20:28structuring sort of like educational
9:20:30offers. Just let me know in the comments
9:20:31down below. But anyways, after you've
9:20:32picked the right model, the question
9:20:34becomes how you actually figure out what
9:20:36exact number to charge. So, this part
9:20:38always goes back to your discovery phase
9:20:40because you can't really give a real
9:20:42price if you don't understand the
9:20:43current manual process in its entirety.
9:20:46That's your job during discovery. Before
9:20:47you think about pricing anything, you
9:20:49need to map out the process from start
9:20:50to finish. So that means you need to
9:20:52understand things like how often does
9:20:53the process happen, what triggers it,
9:20:54who the key stakeholders are, how long
9:20:56it takes each time, what each hour of
9:20:58that work is worth in dollars based on
9:21:00salary, number of employees involved, or
9:21:02even software costs, and other things of
9:21:04that nature. And then once you know all
9:21:05of that, you can start comparing the
9:21:06manual version to what the automated
9:21:08version of that system would look like.
9:21:10But the key here is you just can't
9:21:11overpromise results. You have to be
9:21:13realistic about it. So maybe you're
9:21:14building a system that only automates
9:21:15half of the process. And make sure
9:21:17you're factoring in that half, not
9:21:18automating that full process into your
9:21:20ROI. But the idea is the second the
9:21:22workflow goes live, the business
9:21:24instantly starts getting back time,
9:21:25money, and focus. And it's your job to
9:21:27translate that into numbers that the
9:21:28client can understand. So here's a
9:21:30simple example. Let's say the process
9:21:31you're automating is a customer support
9:21:33process, and it takes the rep one hour
9:21:34every single day. Let's say 1 hour of
9:21:36their time is worth $50. Over the course
9:21:39of a year, that's roughly $12,000 worth
9:21:41of time that they're spending on that
9:21:42task. So that means that the workflow
9:21:43you're building is saving the business
9:21:45$12,000 annually. And this is where
9:21:47valuebased pricing kicks in because as a
9:21:48good rule of thumb, you want the client
9:21:50to be able to see a 10 times return on
9:21:51what they invest in that first year. So
9:21:53if the workflow is going to save them
9:21:54$12,000, 10% of that would be $1,200.
9:21:57[music] And that's your starting point
9:21:58for pricing. You see how easy it was for
9:22:00me to explain that to you? That's
9:22:01exactly how you would just want to
9:22:02explain it to a client. When someone
9:22:04asks, "How did you come up with that
9:22:05number?" You walk them through the math
9:22:06step by step. It's going to build
9:22:07instant trust. And this doesn't even
9:22:09include opportunity cost. Because
9:22:10freeing up an hour of that employees
9:22:12time every single day doesn't just save
9:22:13the business money, it's also enabling
9:22:15that employee to shift their time into
9:22:16higher value work. Real revenue
9:22:18generating opportunities. Answering
9:22:20fewer repetitive customer support
9:22:21tickets might mean that they can support
9:22:22new hires, talk to high-v value
9:22:24customers, or help improve systems,
9:22:25things that actually push the business
9:22:27forward. That's why the ROI often
9:22:28compounds. [music] Maybe month one they
9:22:30save $1,000, but maybe month two it's
9:22:321,500. Month three it could be $2,000.
9:22:34So over a full year, it could end up
9:22:35being far more than just that $12,000
9:22:37that you originally projected. And when
9:22:39the value grows, so does your pricing
9:22:40power. Now, once you do a couple of
9:22:41these projects and you've built some
9:22:42trust, that's when you can start to talk
9:22:44about retainers. And this is where the
9:22:45math does shift a little bit because now
9:22:47you're not pricing for one workflow.
9:22:49You're pricing your time, your team, if
9:22:50you have one, and your ongoing support.
9:22:52So with a retainer, you want to protect
9:22:53your margins. Think about what does it
9:22:55cost for you to operate and for you to
9:22:57deliver. What is your time worth? What
9:22:58does it cost if you bring on a
9:22:59developer? What does it cost if you
9:23:00maintain the systems that you've already
9:23:02built? Even if you're not selling hours,
9:23:04you still need to understand how much
9:23:05time things are taking because this
9:23:07helps you think about staffing for
9:23:08milestones. Here are the [music]
9:23:09questions I would ask myself. Is this
9:23:10just me working full-time inside the
9:23:12client's business? Does this require a
9:23:13part-time engineer? Does it require a
9:23:15full-time engineer? Does this also
9:23:16require an account manager or a project
9:23:18manager? Once you know those things, you
9:23:19can start to estimate what it costs you
9:23:21as a business to run every month. So,
9:23:23let's say you figure out that servicing
9:23:25this particular client properly would
9:23:26cost you $5,000 per month. In the
9:23:28consulting and agency world, the general
9:23:29target margin for this type of work is
9:23:3150 to 70%. 50% is a good baseline to
9:23:34shoot for. 70% is great and it gives you
9:23:36room to scale. But just remember that
9:23:37we're in a service-based business where
9:23:39the work is very custom and bespoke. So
9:23:40you have to make sure you're protecting
9:23:42at least 50% margin. Otherwise, you may
9:23:44struggle to pay your employees and it
9:23:46could just get messy. So in this case,
9:23:47if it cost you $5,000 per month to
9:23:49deliver, you might want to price the
9:23:51retainer at $10,000 per month. That
9:23:52gives you a 50% margin and a lot of
9:23:54breathing room to bring on help or
9:23:56expand the scope later. And this is why
9:23:57I always tell people to start solo or
9:23:59with one developer because the second
9:24:00that you start juggling multiple clients
9:24:01and multiple engineers, your margins can
9:24:03shrink fast if you're not careful and
9:24:05doing projections like this. If you've
9:24:07got a developer on salary, you never
9:24:08want them sitting idle because you
9:24:10overestimated [music] workload or
9:24:11promised unrealistic timelines because
9:24:13that's how agencies burn cash. If you've
9:24:15got some developers kind of on the bench
9:24:16as hourly contractors, it's not as
9:24:18damaging to the business, but still not
9:24:20the ideal situation. So when you think
9:24:21it through and you protect your margins
9:24:23and you [music] build a repeatable way
9:24:25of delivering, that's when you get
9:24:26predictable profit. And predictable
9:24:27profit is what actually lets you scale,
9:24:29make more hires, bring on new team
9:24:30members without tons of stress. Now,
9:24:33even if your price is perfect and
9:24:34[music] you have all the ROI
9:24:35calculations to back it up, how you
9:24:36present it makes all the difference in
9:24:38whether a client says yes or no. And
9:24:40here's the key. You never start with the
9:24:41price. You want to start with that
9:24:42transformation before you ever say a
9:24:44number. You want the client imagining
9:24:45what the business looks like once that
9:24:47system is live, what their team's
9:24:48day-to-day looks like, what problems
9:24:49disappear, what becomes easier, and once
9:24:51they're bought into that outcome, the
9:24:52price starts to make more sense. When
9:24:54you walk them through the proposed
9:24:55solution, you want to explain exactly
9:24:56everything that's actually included in
9:24:58that total cost. So that usually looks
9:25:00like setup, hosting, testing and quality
9:25:02checks, optimization, their involvement
9:25:03in what you need from them,
9:25:04documentation, enablement, and training,
9:25:06and if the conversation goes there,
9:25:08maintenance. And when the client
9:25:09understands everything that's included,
9:25:11it will also be much easier for them to
9:25:12see your price as an investment instead
9:25:14of just some random expense. This is
9:25:15also where visuals help a lot. So simple
9:25:17screenshots, a small wireframe, or even
9:25:19a rough diagram of the workflow can make
9:25:21the process feel real and tangible. If
9:25:23you've got a working demo or past case
9:25:24studies, this is the perfect moment to
9:25:26use them as well. Even in discovery, I
9:25:27talk about how important it is to be
9:25:29writing down all the steps while you're
9:25:31sharing your screen so the client can
9:25:32see exactly, you know, what you think of
9:25:34the process and make sure you guys are
9:25:35aligned. Same thing with a wireframe.
9:25:37and then you guys are fully aligned
9:25:38before you ever talk about money.
9:25:39Because one of the biggest mistakes that
9:25:40I made early on was having scopes that
9:25:42were way too vague, way too ambiguous. I
9:25:44would deliver something and the client
9:25:46thought that it needed more. And we
9:25:47ended up going back and forth debating
9:25:48what done actually meant. When you
9:25:50present your proposal, be very clear
9:25:51about what the final outcome actually
9:25:53looks like. Spell it out, what's
9:25:54included, what isn't, what counts as
9:25:56completion. You should have a very
9:25:57clear, bulleted list of the exact
9:25:59functionality requirements of the system
9:26:01in this scope of work. This [music] will
9:26:02save you frustration. It protects your
9:26:03time. It sets the relationship off on
9:26:05the right foot and it prevents scope
9:26:06creep because naturally as you start
9:26:08building and the client starts to get
9:26:09updates and give you feedback, you guys
9:26:11are both going to realize that there's
9:26:12different features you'd want to add
9:26:13into the system. And in that case, you
9:26:15just have to say, "Hey, I'll add these
9:26:16to the backlog, all of these feature
9:26:17requests, and when we do a V2, we can
9:26:19add them all in. And also, what you're
9:26:21doing there is you're planting a seed
9:26:22that there will be future work between
9:26:23you two. Something else that you need to
9:26:25make clear is the QA process, the
9:26:26quality [music] assurance process. Even
9:26:28the best workflows need a real testing
9:26:30period before they go into full
9:26:31production. And even then, you want to
9:26:33consider constantly monitoring them and
9:26:34evaluating them. That means you first
9:26:36have internal QA, internal testing. Then
9:26:38you have the client do testing and
9:26:39feedback and you fix bugs and you add
9:26:41small adjustments and then you do
9:26:42another round of internal QA, client QA.
9:26:44Finally, you're able to evaluate it with
9:26:46live data and then consider pushing it
9:26:48into full production. And you need to
9:26:49make sure that they understand their
9:26:51role in the process because they're
9:26:52going to have to give you access to
9:26:53tools, provide test data, answer
9:26:55questions quickly, give you feedback,
9:26:56and if they're slow and unresponsive,
9:26:58then everything kind of slows down and
9:26:59that is not on you. So once again,
9:27:01setting these expectations up front
9:27:03makes the project smoother for both
9:27:04sides. Now, of course, not every client
9:27:06will instantly agree. So you do need to
9:27:08know how to handle objections without
9:27:09lowering your value. If they push back,
9:27:11adjust the scope, not the price. Reduce
9:27:13complexity, remove a feature, break it
9:27:14into phases, but [music] don't discount
9:27:16the value that it's actually providing.
9:27:17And as you talk to more clients who
9:27:18start to notice these patterns in their
9:27:20objections, and once you see those
9:27:21patterns, you can start to address those
9:27:23concerns before the client even ever
9:27:24brings them up. This is how you remove
9:27:26those hidden costs and it builds a lot
9:27:27more trust. When someone is budget
9:27:29sensitive, bring the conversation back
9:27:30to the long-term value instead of the
9:27:32short-term expense. You can remind them
9:27:33that the goal is to save time, money
9:27:35every single month, not just to have
9:27:37something built. It's [music] also super
9:27:38important to understand when to walk
9:27:39away. Just because someone is willing to
9:27:41pay you $10,000 upfront, that does not
9:27:43mean you want them as a client. If they
9:27:45undervalue your work, question every
9:27:46step, or they show early signs of being
9:27:48a difficult partner, walking away is
9:27:50often the best decision. Protects your
9:27:51energy and your reputation. A lot of
9:27:53times, it just might be a feeling that
9:27:54you get in your gut. One of the most
9:27:55common questions I get is about
9:27:57intellectual property though, so I
9:27:58thought I'd address that here for you
9:27:59guys real quick. Clients want to know if
9:28:01they will own what you build. Now, the
9:28:02simple answer is yes. It's just the best
9:28:04way to handle it. All IP that we build
9:28:06for them is [music] theirs. I usually
9:28:08handle this by saying, "Yeah, we build
9:28:09on open infrastructure. We use internal
9:28:11components or templates to speed up
9:28:13development. But the real IP in the
9:28:14system is the prompts, your data, your
9:28:16workflows, and how everything is
9:28:18combined in this ecosystem." So that's
9:28:19[music] very specific to your business
9:28:20and provides no value to anyone else.
9:28:22You're not losing anything by giving
9:28:23them ownership. Just make sure you're
9:28:25protecting your own reusable components,
9:28:26though. So, that's only one example of
9:28:28an objection. In the free resource guide
9:28:29that will come with this video, I'll add
9:28:31like six more common objections and
9:28:32exactly how you can handle them. You can
9:28:34download it for free inside my school
9:28:35community. The link for that will be
9:28:36down in the description. So, once you've
9:28:37closed a few clients, the next step is
9:28:39turning those one-time projects into
9:28:40consistent recurring income. Even if
9:28:42you're not on a full monthly retainer,
9:28:44you can still introduce simple recurring
9:28:45services that can benefit both sides.
9:28:47These are small add-ons that create
9:28:48stability for you [music] and peace of
9:28:49mind for the client. and they're easy
9:28:51for clients to say yes to because
9:28:52they're tied directly to maintaining the
9:28:54system that you already built. So, here
9:28:55are a few examples. The first one you
9:28:57could do is a maintenance fee. This is
9:28:58one of the simplest add-ons. You can
9:29:00charge a small monthly fee to make sure
9:29:01that their workflow stays healthy if an
9:29:03API changes, if a model updates, if
9:29:05something breaks, or if the client needs
9:29:06minor feature enhancements. You're there
9:29:08to keep everything running smoothly.
9:29:09It's a light retainer that covers
9:29:11reliability and support. Now, this could
9:29:12be a flat monthly fee, something like
9:29:14200 to 1,500 bucks per month per system.
9:29:17Second option is kind of optimization
9:29:19and monitoring. This is another option
9:29:20that is ongoing optimization. You can
9:29:22check the systems weekly or monthly,
9:29:24review logs, look at outputs, and
9:29:26tighten things up so the workflow keeps
9:29:27improving over time. This is really good
9:29:28for AI heavy systems because the prompt
9:29:31tuning, the new models, the quality
9:29:32checks, and the retrieval steps often
9:29:34get better the more that you iterate and
9:29:35the more feedback you get from the real
9:29:37world. And then the third one is
9:29:38expansion projects. This one's big. Like
9:29:40we talked about earlier, whenever you
9:29:41sell a valuebased project, you and the
9:29:44client almost always uncover a backlog
9:29:46of improvements, new ideas, or version
9:29:48two upgrades. and that's completely
9:29:49normal. It's also a perfect opportunity
9:29:51to turn a one-off project into a
9:29:52longerterm partnership. Now, the beauty
9:29:54of these different retainers is that you
9:29:55could sell just one of them or all three
9:29:57of them as a package. You can think
9:29:58about pricing all of these different
9:30:00retainers as a monthly fee, maybe
9:30:01somewhere from 200 to upwards of 1,500
9:30:03per month, but you could also do it as
9:30:05an hourly maintenance package where you
9:30:06would give them maybe 5 to 20 hours per
9:30:08month and assign a dollar amount to
9:30:10those hours. Or the last way, which is
9:30:12probably the best way to do it, would be
9:30:13a percentage of the original project
9:30:14cost, anywhere from 10 to 25% of the
9:30:17price that they paid. So, recurring
9:30:19revenue makes your income more
9:30:20predictable. It makes your
9:30:21month-to-month more stable. But the real
9:30:23purpose of these add-ons is relationship
9:30:25building. Because every time that you
9:30:26handle maintenance, optimize something,
9:30:28or help them launch a V2 or a V3, you're
9:30:30reinforcing [music] that you understand
9:30:32their business, their workflows, their
9:30:34processes, and their goals better than
9:30:35anyone else does. And that's how you
9:30:37become the partner that they trust, the
9:30:39person that they can call when something
9:30:40important needs to get built. And the
9:30:41deeper that they become integrated with
9:30:43you, the harder it is for them to switch
9:30:44vendors. You're embedded. You're the one
9:30:46who knows how everything works under the
9:30:47hood. And that's leverage. The more
9:30:48value you provide, the more expansion
9:30:50work you will likely unlock. And that's
9:30:52how you build a business that grows
9:30:53consistently instead of constantly
9:30:54trying to find new clients. So to make
9:30:56all of this easier to repeat, it helps
9:30:58to have a simple internal framework that
9:30:59you can follow every single time that
9:31:00you price a workflow. That's why I came
9:31:02up with the price framework. [music]
9:31:03It's a five-step process that takes all
9:31:05of the guesswork out of pricing and
9:31:06makes sure that you can stay strategic,
9:31:08consistent, and grounded in value. So
9:31:10the P is for prepare. We started off
9:31:11this video talking about the mindset. So
9:31:13before you even think about numbers,
9:31:14make sure you're grounding everything in
9:31:16value based pricing. Just remember,
9:31:17you're not charging for hours. You're
9:31:18charging for outcomes, and you're
9:31:19building systems that [music] save time,
9:31:21money, and reduce human error. The R is
9:31:23for research. So, this is your discovery
9:31:25phase. You want to fully map out the
9:31:26manual workflow from start to finish.
9:31:28How does it happen? What triggers it?
9:31:29Who does it? How long does it take? What
9:31:31tools are involved? What does the time
9:31:33cost the business? So many things that
9:31:35you need to figure out. And you have to
9:31:36use this to actually be able to price it
9:31:38properly, which leads directly into the
9:31:39eye, which is for identify. Identify the
9:31:41ROI. This is where you turn your
9:31:43research into actual numbers. What are
9:31:44the monthly savings, annual savings, the
9:31:46opportunity cost, the efficiency gains,
9:31:48and remember the 10 times investment
9:31:50kind of golden rule. They should roughly
9:31:52see 10 times the return on whatever they
9:31:53pay you within the first year. The C is
9:31:55for communicate. Now that you have that
9:31:57price, you still have to present it in a
9:31:58way that makes sense. You want to paint
9:32:00the picture of what the business will
9:32:01look like once the system is live.
9:32:02Explain what's included. Explain how QA
9:32:04works. Explain [music] what you need
9:32:05from them. If you can't communicate
9:32:06clearly, you'll probably lose them long
9:32:08before you even get to the price
9:32:10conversation. [music] And then the E is
9:32:11for expand. Once the project is
9:32:12complete, your job still is not over.
9:32:14You want to look for opportunities to
9:32:16continue building on top of that
9:32:17relationship. That could be maintenance,
9:32:18optimization, monitoring, additional
9:32:20workflows, V2 upgrades, retainers, or
9:32:22performance bonuses. Your goal isn't
9:32:24just one sale. Your goal is to become
9:32:26their long-term partner. So, this
9:32:27framework gives you a repeatable,
9:32:28reliable way to price any AI workflow,
9:32:30whether it's your first project or your
9:32:31hundth. So, to wrap this up, I wanted to
9:32:33show you guys a practical example of how
9:32:35this framework works in real life
9:32:36because once you see the numbers laid
9:32:37out, pricing becomes a lot less scary
9:32:39and a lot more logical. So, we worked
9:32:40with a client who wanted help with their
9:32:42inbound sales process. They were getting
9:32:43about 20 leads every single week from
9:32:45their form on their website. And every
9:32:47one of those leads took about an hour of
9:32:48an employees time to reach out, qualify
9:32:50them, nurture them, and ultimately have
9:32:52them book in a call with the sales team.
9:32:54These employees who set those meetings
9:32:55were valued at about $40 per hour. So,
9:32:57this meant that this one process was
9:32:59costing the business $800 per week,
9:33:01which is $3,200 per month. And when you
9:33:03zoom out and you annualize it, it ends
9:33:04up being about $38,000 per year. So,
9:33:07let's walk through that example with the
9:33:08price framework. So P prepare. Right
9:33:10away I grounded myself in value based
9:33:12pricing. My job here was not to think
9:33:14about how long the workflow would take
9:33:15me to build. It was to think about how
9:33:16much value it will return to the
9:33:18business. The R is for research. Then
9:33:20came the discovery phase. We mapped out
9:33:21the entire manual process. When leads
9:33:23came in, who handled them, exactly how
9:33:25long it took, what tools they were
9:33:26using, and where the bottlenecks were.
9:33:27That's how we arrived at the $38,400
9:33:30per year in labor costs opportunity. I
9:33:33identified the ROI. Once we had these
9:33:35numbers, the investment became obvious.
9:33:36I knew automating that process was going
9:33:38to save them about $38,000 in a year
9:33:40alone, not even counting the opportunity
9:33:42cost. So, I priced the product at 15% of
9:33:45those annualized savings, and that came
9:33:46out to $5,500. Because the ROI was so
9:33:49clear, it made complete sense to them,
9:33:51and they could see exactly how the
9:33:52system would pay for itself. See,
9:33:54communicate. Before I ever said the
9:33:55price, I walked them through the
9:33:57solution, how the sales agents would
9:33:58work, what steps it would automate, what
9:33:59testing would look like, what I would
9:34:01need to know from them, and how it would
9:34:02improve both speed to lead and lead
9:34:04quality. I also showed them how we would
9:34:05track success. By the time I anchored
9:34:07the price, they already believed in this
9:34:09outcome and they were bought in. [music]
9:34:10And finally, the E expand. Once we moved
9:34:12into QA and prepared to push the
9:34:14workflow into production, that's when we
9:34:15started to talk about ongoing support.
9:34:17We agreed on a simple maintenance and
9:34:18optimization plan at $550 per month,
9:34:21which is about 10% of the original
9:34:23project's fee. This covered bug fixes,
9:34:24keeping everything aligned with the
9:34:26scope, making adjustments as models
9:34:27updated, and running monthly health
9:34:30reports. Now, this is where we actually
9:34:32did mess up. I should have been proving
9:34:33to the business how valuable the system
9:34:35really is once it was pushed into
9:34:36production and show them how it was
9:34:38actually growing the business. So in
9:34:39hindsight, I should have been tracking
9:34:41things like how fast leads were being
9:34:42contacted, which was basically [music]
9:34:43instant now, how many leads per week
9:34:45were coming in, how much total time was
9:34:46being saved and how the sales team felt
9:34:48using the system. And if we were
9:34:50actually tracking all of these metrics
9:34:51kind of, you know, month after month and
9:34:52we're able to show them that, it would
9:34:54be very clear how the system is
9:34:55explicitly growing the business. And
9:34:57this would have given the client more
9:34:58visibility into the actual impact and a
9:35:00natural entry point for us to make
9:35:02version two improvements and the
9:35:03potential of getting them on a retainer.
9:35:05So, I know that we covered a lot of
9:35:07stuff today and I want to make sure that
9:35:08it all sticks. So, I've thrown all of
9:35:10this into a full resource guide that you
9:35:11can get for free by joining my free
9:35:13school community. The link for that will
9:35:14be down in the description. And if you
9:35:15want to dive a little bit deeper, I've
9:35:16got a full course on all of this kind of
9:35:18stuff. I talk about everything that I've
9:35:20been talking about on YouTube, but in
9:35:21much more depth. So, if that interests
9:35:22you, then definitely check out my plus
9:35:24community. The link for that is also
9:35:25down in the description. Anyways, that's
9:35:26going to do it for today. If you guys
9:35:28enjoyed, you learned something new,
9:35:29please give it a like. It definitely
9:35:30helps me out a ton. And as always, I
9:35:32appreciate you guys making it to the end
9:35:33of the video. I'll see you on the next
9:35:34one. Thanks everyone.
9:35:38Okay, so before we get into this next
9:35:39one, which is talking about delivering
9:35:40and handing over workflows or
9:35:42automations, I just wanted to clear
9:35:44something up because what you'll notice
9:35:45is in the video I do make a few specific
9:35:47references to nen and obviously this
9:35:49course is about cloud code. But the
9:35:51reason why I wanted to still include
9:35:52that is because the foundational
9:35:55concepts are still the exact same. The
9:35:56conversation just moves from okay,
9:35:58instead of an NN JSON workflow, it's now
9:36:01just like a Python script or a
9:36:03TypeScript or whatever coding language
9:36:04you use. The key confusion with NN was
9:36:06like, okay, who owns the account and
9:36:08where is the account hosted? Now the
9:36:10question just becomes whose GitHub or is
9:36:12this in? Whose trigger.dev or you know
9:36:14similar hosting platform does this
9:36:15actually live in? Whose domain or cloud
9:36:17project is the automation running under?
9:36:19You still have the exact same three
9:36:21layers about account ownership, about
9:36:22infrastructure ownership, about access
9:36:25control. The only thing you want to be
9:36:26careful about is if there's a specific
9:36:28provider that has some sort of licensing
9:36:30agreement or terms of service that you
9:36:31have to make sure you're not violating.
9:36:33But essentially, it's the same
9:36:34conversation being upfront with your
9:36:35client and saying, "Hey, where do you
9:36:37guys feel comfortable with this living?"
9:36:38Okay, I need to either have you send me
9:36:40over the API keys in an encrypted way or
9:36:43you need to sign in over here and put in
9:36:45your environment variables, put in the
9:36:46API keys so that this automation can
9:36:48run. Whether that's on a trigger.dev
9:36:49account that I own and I charge you for,
9:36:51or whether that's one that you spin up
9:36:53and then I just basically put my
9:36:54automations in there. Just remember, at
9:36:56its core, an automation is code. So, you
9:36:58just have to figure out where does that
9:36:59code live and what does that code need
9:37:01in order to run every day. So, like I
9:37:03said, as you're watching this next
9:37:04section, just keep that in mind. So, you
9:37:07guys are always asking me, Nate, how do
9:37:08I host workflows? What does the handover
9:37:09process look like? Should I host a
9:37:11workflow to the client? What about
9:37:12security? These types of questions come
9:37:14up all the time. And since there's
9:37:15literally no good video covering all of
9:37:16it, I figured I'd make one. So, in this
9:37:18video, I'm going to walk you guys
9:37:19through the entire process of fulfilling
9:37:20an AI workflow or agent after a client
9:37:22pays you step by step. And at the end,
9:37:24I'll show you a real example from one of
9:37:25the first AI workflows that I ever sold.
9:37:27Let's get into it. So before [music] you
9:37:28build anything, the first question that
9:37:30you have to answer is simple. Who is
9:37:31going to host the workflow? I get this
9:37:32question constantly. And since most of
9:37:34you guys are using NDN, I'm going to
9:37:35frame this answer around how NN actually
9:37:37works because its license is actually
9:37:39kind of what decides what you can and
9:37:41cannot do. I've also worked with tons of
9:37:42different clients and I've delivered
9:37:43workflows to them in tons of different
9:37:45ways. So I basically just charge for the
9:37:46JSON file and then I would give that to
9:37:48them as well as a Loom setup guide of
9:37:50how it works. I've had them invite me to
9:37:52their own end account and I would just
9:37:53develop right there in their own
9:37:54environment. I've developed the workflow
9:37:56in my own environment and then they
9:37:57would give me their login credentials
9:37:58and I would log in and port everything
9:38:00and then I would just log out and I've
9:38:01also helped people spin up their own
9:38:03instance of NadN and then they would
9:38:04invite me once they have the account
9:38:05[music] as a team member. So I say all
9:38:07that to say that there's a lot of
9:38:09different ways to do it. So let me tell
9:38:10you about the main options and what I
9:38:12recommend. So the key rule that you need
9:38:13to know is that you can only use Nitn
9:38:15for your own business internally unless
9:38:17you have a paid commercial or enterprise
9:38:19license if you're hosting it. You are
9:38:20allowed to sell services around nitn but
9:38:22you are not allowed to turn nitn into
9:38:24your own product. So everything about
9:38:25hosting comes down to one thing. Is this
9:38:27nen instance being used by one business
9:38:29internally or are you exposing it as a
9:38:31platform to other people? So now that
9:38:32we've gone over that rule, we have three
9:38:34options for hosting workflows and
9:38:35honestly I would pretty much always go
9:38:37with the first option and that's what
9:38:38I'm recommending to you all, but I'll
9:38:39still touch on the other two because
9:38:40it's still important to understand the
9:38:41difference. All right, so option number
9:38:43one is where the client hosts NAD. This
9:38:45is the safest and cleanest model for
9:38:46almost everyone. This means that each
9:38:48client that you're working with would
9:38:49have their own nitin instance and you
9:38:51simply work inside of it. So what this
9:38:53means is that the client either buys
9:38:54their own anoden cloud and invites you
9:38:56as a user or they sign up on a
9:38:57self-hosted instance or locally hosted
9:38:59instance and then they give you access
9:39:00to it. Now I know that as a service
9:39:02provider you want to make their
9:39:03experience as frictionless as possible.
9:39:05So what you can do is you can help them
9:39:06set it up. You can help them configure
9:39:08their server or even provision an
9:39:09account for them, but then they have to
9:39:11own it and they have to pay for it
9:39:12because you can't mark up the hosting or
9:39:14essentially charge for access to NIDAN.
9:39:16This [music] helps you stay compliant
9:39:17because Naden is being used for that
9:39:19company's internal business processes.
9:39:21You are just providing consulting and
9:39:22workflow development which is completely
9:39:23allowed and you are not giving multiple
9:39:25clients access to the same instance.
9:39:27This is basically the Zapier model where
9:39:28every client has their own seat and
9:39:29you're the builder. So if you're doing
9:39:31client work, this is what I recommend
9:39:32pretty much all the time. Now option two
9:39:33is you host NM but only for your own
9:39:36agency. So in this setup, you run nitn
9:39:38on your own server, but nobody else sees
9:39:39it except for your business. And this is
9:39:41for your own internal automations like
9:39:43lead routing, content workflows,
9:39:44internal AI agents, or anything else
9:39:46where the client never touches NADN
9:39:48directly or ever needs to. Now, you can
9:39:50also deliver a service using Eniden
9:39:51running on your own infrastructure. As
9:39:53long as clients don't log in, don't
9:39:54connect their own API keys, things like
9:39:56that. This is compliance because NIDAN
9:39:58is only powering your own operations.
9:40:00You're not exposing it as a hosted
9:40:01platform and you're not giving clients
9:40:03access to it. So just think of it as
9:40:04your own internal engine. And an example
9:40:05of this could be if you were sending a
9:40:07client something like an automated
9:40:09report or research and as a service to
9:40:11them that's what you're basically
9:40:12delivering even if your own internal
9:40:14automation is powering that on your own
9:40:15hosted NN that's fine because they are
9:40:17just paying for that deliverable. Then
9:40:19option three is when you host N&N as the
9:40:21product and this is where you would need
9:40:22a commercial or enterprise type of
9:40:24license. This is the line that you
9:40:26cannot cross on the free or the
9:40:27sustainable use license. So essentially
9:40:29you cannot build a SAS product software
9:40:31as a service product where NN is the
9:40:33value. Even if the client never sees the
9:40:35NN UI if your offer is basically give me
9:40:37your credentials and I'll run your
9:40:38automations on my end server. That is
9:40:40not allowed without a commercial
9:40:42agreement. And this model only really
9:40:43makes sense if you're building out a SAS
9:40:45or you're selling automation as a
9:40:46subscription where ended is clearly the
9:40:48engine. If that is the case you
9:40:50definitely want to talk to end sales and
9:40:51get a commercial or enterprise license.
9:40:53And I will say those things are not
9:40:54cheap. So the short version is simple.
9:40:56If you're building workflows for clients
9:40:57let them host it. If you're running
9:40:58automations for your own company, host
9:41:00it yourself. And if you want to build a
9:41:01SAS or platform, you need a commercial
9:41:03license. And once you've decided where
9:41:04the automation will actually live, then
9:41:06you can move into planning out the
9:41:07build, structuring the data, and
9:41:09preparing for the handover. So next,
9:41:11let's look at security and data
9:41:12protection. Once you've built the
9:41:13workflow, your job is to make sure the
9:41:14data moving through it stays secure.
9:41:16That means no leaking sensitive
9:41:18information, no exposing personal data,
9:41:19and no breaking privacy laws like GDPR.
9:41:22So let's go over how Aniden handles
9:41:23security and how you can explain this
9:41:25clearly to your clients. And in a few
9:41:26minutes, I'll talk more specifically
9:41:27about API key management and billing
9:41:29since that also deserves its own
9:41:30section. The first thing to understand
9:41:31is how Naden protects sensitive fields
9:41:34inside of your automations. Credentials
9:41:35inside Naden are encrypted at rest and
9:41:38they're decrypted in memory at the
9:41:39moment the workflow runs. So these nodes
9:41:41with your credentials in there simply
9:41:43reference credentials by name. And if a
9:41:44teammate or client doesn't have
9:41:46permission to that workspace that has
9:41:47those credentials, they won't be able to
9:41:49see the raw values, the API keys, the
9:41:51passwords, whatever you want to call
9:41:52them. This is why handing off a workflow
9:41:53is safe when it's done correctly because
9:41:55you're not exposing sensitive fields
9:41:56inside nodes and you're not storing
9:41:58anything in plain text. The platform is
9:41:59built to keep secrets locked away so
9:42:01only the automation engine can access
9:42:03them when it needs to. Another big part
9:42:04of security is web hook hardening. So a
9:42:06web hook is basically like a public door
9:42:08into a workflow which means you need to
9:42:09treat it with the same seriousness as
9:42:11you would any inbound request to an
9:42:13application. So maybe this means using
9:42:14HTTPS so that data is encrypted in
9:42:17transit or signing secrets or
9:42:18verification tokens for services like
9:42:20Stripe, GitHub or any provider that
9:42:22supports signature validation. What you
9:42:23can do is have EndN verify that
9:42:25signature before it actually trusts the
9:42:27payload and lets it into the workflow
9:42:28and never put sensitive data inside the
9:42:30URL of a web hook. And if you really
9:42:31want to get advanced and a little more
9:42:32technical here, if the use case called
9:42:34for it, you could implement things like
9:42:35rate limits or additional authentication
9:42:37checks to prevent spam, brute forcing or
9:42:39automated abuse. So the way I would
9:42:41explain this to clients is simple. Every
9:42:42external trigger hitting your automation
9:42:44must be authenticated so that only
9:42:45approved systems can talk to this
9:42:47workflow, which means random people
9:42:48cannot guess the URL and start hitting
9:42:50your CRM or your internal systems for
9:42:52data. And while we're on this topic,
9:42:53it's also worth thinking about prompting
9:42:54in some guard rails and building a
9:42:56system so that people can't jailbreak or
9:42:58prompt inject your AI agents. Now,
9:42:59another big responsibility is handling
9:43:01CRM, payment, or other personal data
9:43:03because this information is often
9:43:05regulated under GDPR and similar laws.
9:43:07Anything that could identify a person is
9:43:09protected. So, this is not legal advice.
9:43:11However, there are a few basic best
9:43:13practices to follow, but you should
9:43:14always be consulting with professionals
9:43:16based on your industry and your laws
9:43:17that govern you and your clients. But
9:43:19you could do stuff like using data
9:43:20minimization, which means only bringing
9:43:22in the fields that you actually need.
9:43:23You could limit who can see the workflow
9:43:24runs, so only the right people have
9:43:26access to the logs and the payloads. You
9:43:28should also understand that a client
9:43:29must have a legal basis for collecting
9:43:31and processing the data that they're
9:43:32passing into the workflow. And if you're
9:43:34processing data on the client's behalf,
9:43:35you're usually acting as a data
9:43:36processor, which means you may need a
9:43:38data processing agreement in place. You
9:43:40also need to make sure that your
9:43:41automations don't make it impossible for
9:43:42the client to honor requests like data
9:43:44deletion, data corrections, or access
9:43:46requests. Basically, what that means is
9:43:48you should know exactly where the data
9:43:49flows so that it can be removed or
9:43:51updated if needed. And Naden helps with
9:43:52this because you can prune executions,
9:43:54trim logs, and limit the amount of data
9:43:56that the system stores over time. And
9:43:57one of the biggest advantages of NADN is
9:43:59that it's source available, which is
9:44:01essentially better known as open source,
9:44:02which means that it can be fully
9:44:04self-hosted, which gives the client the
9:44:05option to keep all of their workflow
9:44:07data in their own environment. So you
9:44:08can run noden on their infrastructure,
9:44:10connect it to only tools that they
9:44:11approve and trust and even use local or
9:44:13self-hosted AI models instead of sending
9:44:15data to OpenAI or other closed source
9:44:17proprietary models. This gives true data
9:44:19sovereignty. The client chooses where
9:44:20the server lives, how it's secured, and
9:44:22who has access. So for privacy sensitive
9:44:24clients, this is a huge selling point.
9:44:25Instead of pushing sensitive data
9:44:27through random cloud services, you can
9:44:28run the entire automation engine inside
9:44:30their own locked room, whether that's on
9:44:31prem or inside a private VPS. And that's
9:44:33the foundation of security and data
9:44:35privacy. Your job is to build workflows
9:44:36that move data safely, keep sensitive
9:44:38information protected, and give clients
9:44:40confidence that nothing is leaking. And
9:44:41now that you understand the security
9:44:42side, the next section I'm going to walk
9:44:44you guys through is about API key
9:44:45management and billing because that's
9:44:46another area where I get a ton of
9:44:48questions. So, the main ones I get are
9:44:49who owns the API keys and who pays for
9:44:51the usage. And the cleanest answer I can
9:44:53give you is that the client should
9:44:54always pay for their own API keys and
9:44:56usage. This keeps everything
9:44:57transparent, predictable, and it avoids
9:44:59a lot of headaches later. So the ideal
9:45:00setup is having the client sign up for
9:45:02the tool themselves. Enter their billing
9:45:04information, generate the API key, and
9:45:05paste it directly into Ended. When you
9:45:07do it this way, the key never gets sent
9:45:08over the internet to you, and the client
9:45:10keeps full control over their account.
9:45:11They can see the usage, they can see the
9:45:13charges, they can turn it off if they
9:45:14want, and nothing about the automation
9:45:15is hidden from them. It's just a much
9:45:17cleaner working relationship. Now, the
9:45:18best way to do this, because like I
9:45:20said, you want to remove as much
9:45:21friction as possible, is to send them a
9:45:22Loom video walking them through exactly
9:45:24where to click and how to create that
9:45:25key and where to paste it into N. Just
9:45:27keep it dead simple for them. You could
9:45:29even walk them through it on a Zoom call
9:45:30if they prefer. Now, could you set up
9:45:32the API accounts on your side and then
9:45:34just build them later? Yes, you could.
9:45:35There's nothing about that is
9:45:36non-compliance, but it could create all
9:45:38kinds of problems because they don't see
9:45:39the usage. They don't understand where
9:45:40their money is going. And if you're
9:45:42marking up usage or you're charging a
9:45:43fixed rate, it can get confusing fast.
9:45:45As the automation scales or if something
9:45:47breaks or spikes, even if it does feel
9:45:49easier right in the moment because the
9:45:50client doesn't actually have to go do
9:45:51anything, it definitely in the long run
9:45:53can create more questions than answers.
9:45:54So, letting them own it and pay for it
9:45:56keeps everything clean. Now, if a client
9:45:57is intimidated by that process and wants
9:45:59you to handle the key directly, you
9:46:01still want to make sure you're
9:46:02transferring it securely. So, just send
9:46:03them over Slack, ClickUp, text, or
9:46:05email. Have them drop it into some sort
9:46:07of secure vault like a one password or
9:46:09any encrypted secret sharing tool where
9:46:11they can generate a onetime link. Then,
9:46:12you can copy that key into NADN yourself
9:46:14and that link can expire or something
9:46:16like that so that no one else could ever
9:46:17access that vault. As a small bonus
9:46:19feature, you could even think about
9:46:20offering them a dashboard that shows all
9:46:22their API keys and all the billing in
9:46:23one spot. That gives them visibility and
9:46:25it gives you credibility because you're
9:46:27helping them manage their system like a
9:46:28real piece of infrastructure. So the
9:46:30simple rule is clients own their API
9:46:31keys, clients pay for their usage and
9:46:33you make the process painless for them.
9:46:35This keeps everything clear, secure, and
9:46:36scalable as the automation stack grows
9:46:38and as your professional relationship
9:46:40matures. Now once all that's been
9:46:41decided and you're actually starting to
9:46:42think about handing over the project,
9:46:44you have to make sure that it's been
9:46:45fully tested in the right way before you
9:46:47actually send it over. If you're not
9:46:48careful, the workflow may have bugs that
9:46:49you didn't spot before and this can hurt
9:46:51your reputation and relationship with
9:46:53the client. So, the first step is
9:46:54planning your test data with the client.
9:46:56You don't want to test with madeup
9:46:57examples that have nothing to do with
9:46:58their business. So, early on, ask them
9:47:00for a small sample set that looks like
9:47:02real usage. I typically do this before
9:47:03we sign the contract so they know what
9:47:05is expected of them because if they
9:47:07delay getting you all that sample data,
9:47:09then it's going to delay your process as
9:47:10well. So, that could be emails, support
9:47:12tickets, transcripts, CRM records,
9:47:14whatever actually fits the workflow when
9:47:15it's in production. And of course, if
9:47:17needed, they can anonymize it. Then, you
9:47:18can agree on what success looks like and
9:47:20what is a good output and what must
9:47:21never happen. things like wrong tags,
9:47:23broken links, leaking info, or sending
9:47:25the wrong person the wrong message. So,
9:47:26you can explain it very simply. Before
9:47:28we go live, we're going to run your real
9:47:30examples through the system so that you
9:47:31can see exactly how it behaves. And then
9:47:33inside your own testing, you want to
9:47:34think less like a developer clicking in
9:47:36every node and looking at the
9:47:37configuration and think more like an
9:47:39engineer who's planning for failure. So,
9:47:40with automations, especially when they
9:47:42have AI, you have to accept that you
9:47:44don't know what you don't know. And once
9:47:45the system goes into production, real
9:47:47users and real data will always reveal
9:47:48edge cases that you didn't think about.
9:47:50So during testing, you should
9:47:51intentionally look for worst case
9:47:52scenarios and ask yourself what happens
9:47:54if this, what happens if that, bad data,
9:47:56no data, duplicate data, or even
9:47:58something completely unexpected. Instead
9:48:00of assuming the workflow will run
9:48:01smoothly forever, you want to build in
9:48:02guardrails. Maybe the workflow can time
9:48:04out gracefully so nothing happens. Maybe
9:48:06you set up an error workflow that alerts
9:48:07the team. Maybe you log all failures
9:48:09into a Google sheet so you can track
9:48:10patterns over time. The idea is not to
9:48:12eliminate every possible issue because
9:48:14you can't, but to make sure that when
9:48:16something does break, it breaks safely
9:48:17and quietly and in a way that gives you
9:48:19enough information to go fix it fast. So
9:48:21once that works, you step back and you
9:48:22treat the whole workflow like a black
9:48:24box. You feed in a lot of examples, not
9:48:25just one or two, ideally dozens or even
9:48:27hundreds of sample inputs if you can get
9:48:29them. And then for each one, you log
9:48:30what came in, what happened in the
9:48:31middle, and what the final output was.
9:48:33Then you compare those outputs to the
9:48:34success [music] criteria that you agreed
9:48:36on with the client. You flag the
9:48:37failures, the weird edge cases, and
9:48:38borderline results that you want to talk
9:48:40about with them. And this is your
9:48:41internal QA or quality assurance pass.
9:48:44The goal is to catch as much as possible
9:48:45before the client ever tries it. And I
9:48:47would do internal QA for at least a few
9:48:49days before having them get in there and
9:48:51provide any feedback. Now AI adds
9:48:52another layer on top of that because
9:48:53you're not just checking that the AI
9:48:55node runs. You're checking the quality
9:48:56of what it says. So for beginners, focus
9:48:58on a few simple checks like relevance
9:48:59and correctness. Does the answer
9:49:01actually respond to the request with
9:49:02accurate information? Another thing is
9:49:04tone and safety. It shouldn't be toxic,
9:49:05off-brand, or leaking hidden system
9:49:07prompts and private info. And then
9:49:08you've got the element of consistency.
9:49:10So, if you send in the same 10 inputs,
9:49:11are you getting roughly the same 10
9:49:13answers every time? And behind the
9:49:14scenes, you can run simple AB tests and
9:49:16evaluations where you try different
9:49:17prompts and different models on the same
9:49:19data set and track which ones give you
9:49:20the best results. And to the client, you
9:49:22can phrase that like, "We tested several
9:49:23prompts and models on your real examples
9:49:25and kept the one that hit the highest
9:49:27quality and the most consistency. Here's
9:49:29the evaluation data we ran." And you
9:49:31could even use Eniden's built-in
9:49:32evaluation feature. Of course, you have
9:49:34to be able to actually show all that.
9:49:35So, that's where logging comes in. And
9:49:36logging is what makes all of this not
9:49:38feel like magic because I like to have
9:49:39my workflow store execution history in a
9:49:41Google sheet where it tracks all the
9:49:42inputs, all the outputs, the tool calls,
9:49:44the errors and tokens so that you can
9:49:46actually look through this log and you
9:49:47can identify patterns, common failure
9:49:49types, recurring bad inputs or weak
9:49:51spots in your prompt or model choice.
9:49:53That same log becomes your evidence when
9:49:54you talk to the client and you can show
9:49:56them what you tested and why you made
9:49:57certain decisions or improvements. And
9:49:59after you're happy internally with the
9:50:01system, you can move to the client-f
9:50:02facing QA where you give them a clear
9:50:04way to test the system. That might be a
9:50:05chat box or a form or a simple UI. Just
9:50:08make it simple. You don't want them to
9:50:09actually have to get into Nen and look
9:50:11at all of that mess. Then you ask them
9:50:13to tell you what they think about the
9:50:14system, about the outputs, the tone, the
9:50:16formatting, anything like that. And a
9:50:18lot of times at this point, if you did
9:50:19everything right, you will just be doing
9:50:21little prompt tuning and model tuning
9:50:23tweaks. And then to wrap up, you can
9:50:24record a short update video, show one or
9:50:26two full runs from the input to workflow
9:50:28to output, point to the logs, and
9:50:29explain how the system handled those
9:50:31real examples. And that is the kind of
9:50:32QA or quality assurance that builds
9:50:34trust and makes clients want to work
9:50:36with you again. Now, once you've tested
9:50:37the workflow successfully, it's time to
9:50:39start the handover process. This is the
9:50:41delivery phase and it looks a little
9:50:42different depending on your current
9:50:43situation. So, the first thing that
9:50:45could affect handover is whether or not
9:50:46you built the workflow directly in their
9:50:48environment because if you did, the
9:50:49handover is a lot easier because
9:50:51everything's already set up and their
9:50:52credentials and everything like that.
9:50:53But if you didn't, then you'll have a
9:50:54bigger transfer process where you
9:50:55actually help them move credentials and
9:50:57you have to recreate some of those
9:50:58connections. The second thing that can
9:50:59affect your handover is whether or not
9:51:01this is the end of the project or if
9:51:02you've already scoped out more work or
9:51:04some sort of ongoing maintenance
9:51:05retainer. If you're sticking around, you
9:51:06might keep some testing infrastructure
9:51:08in place, but if you're not, then your
9:51:09handover needs to be more final and
9:51:11complete. There are a few key steps you
9:51:13want to follow either way. First, you
9:51:14want to duplicate the workflow. So, keep
9:51:15one version somewhere as a backup or a
9:51:17testing version and then push the clean
9:51:19one into production. This is exactly how
9:51:21software teams work. You kind of have a
9:51:23test environment where you experiment
9:51:24and iterate and then you have a
9:51:25production environment where only a
9:51:27stable version lives. So anytime that
9:51:28you need to make changes or updates, you
9:51:30do that on the test version first.
9:51:31Confirm that everything still works and
9:51:33then you can move that version into
9:51:34production. And that same idea can apply
9:51:36to when Naden itself releases updates or
9:51:38when the tools or integrations change.
9:51:40You never want to just update your
9:51:41production environment blindly. You want
9:51:43to update the test setup first, load the
9:51:44workflow, make sure all the
9:51:45functionality is still intact. And then
9:51:47once you know it still behaves as it
9:51:48should, then you can update the version
9:51:50that the client would actually rely on.
9:51:51This avoids outages, broken automations,
9:51:53and a lot of unnecessary stress. On top
9:51:54of that, you always want to back up your
9:51:56workflows. So you can store the exported
9:51:57JSON on GitHub, Google Drive, or even a
9:51:59simple Google sheet. So you always have
9:52:01previous versions that you can revert to
9:52:03if needed. And if you want to go a step
9:52:04further, you can build an automated
9:52:05backup process using an itself so that
9:52:08it periodically exports and saves the
9:52:09workflows somewhere else. The next thing
9:52:11you want to do is about workflow
9:52:12hygiene. So you want to make sure that
9:52:13it's clean and easy to understand. Use
9:52:15clear naming in each step, label each
9:52:17step, have sticky notes around the
9:52:18workflow explaining what it's doing and
9:52:19why you built it that way. The goal is
9:52:21that anyone from their team or your team
9:52:23later on could open up that workflow or
9:52:25have a PDF of that workflow and see and
9:52:28understand the logic right away. You
9:52:29also want to double check that there are
9:52:30no sensitive keys or tokens anywhere in
9:52:32the workflow before you hand it over
9:52:33because you want a clean handoff where
9:52:35the client knows exactly where their API
9:52:36keys go and how to set them up on their
9:52:38own account. So in this set of
9:52:39deliverables, it could also have a Loom
9:52:41walkthrough where it's a quick 1 or 2
9:52:43minute video where you show how the
9:52:44system works, how to configure it, and
9:52:45what to do if certain elements need
9:52:47updating because there's never one exact
9:52:49right solution for a process as far as
9:52:51like how to automate it. And everyone's
9:52:52brains work a little differently. So if
9:52:54you can explain to them what you were
9:52:55thinking when you built it, it's going
9:52:56to be really helpful. And whether this
9:52:57is the only project you're doing for
9:52:58them or if you're staying on retainer,
9:53:00good documentation is always valuable
9:53:01because like I said, if someone on their
9:53:03team later takes it over or if you bring
9:53:04in a developer later to help maintain
9:53:06the account, everything is super clear
9:53:07and no one has to guess what was built
9:53:09or why. And it really protects both
9:53:10sides because it helps the client feel
9:53:12supported and confident. And it helps
9:53:13you avoid being the bottleneck whenever
9:53:15something needs to change. And this is
9:53:16how you deliver workflows professionally
9:53:18and set yourself up for a long-term
9:53:19relationship instead of one-off
9:53:21projects. Now, this is the part that
9:53:22people never really discuss online, and
9:53:24that's the legal and financial side of
9:53:25things. After handing over the workflow,
9:53:27there are a few things you want to make
9:53:28sure are agreed on in advance. And the
9:53:30first one is billing. You want to be
9:53:31crystal clear on this. First, you close
9:53:33out the current project and get paid for
9:53:34what you just built. Then, you decide if
9:53:36there will be an ongoing paid
9:53:37relationship to keep everything healthy
9:53:38over time through a maintenance
9:53:40retainer. Before anything else, you want
9:53:41to revisit the scope of work that both
9:53:43of you guys agreed on at the start. Your
9:53:44contract or scope of work should already
9:53:46say what finished means, which workflows
9:53:48you promised to build, which systems
9:53:49they would connect to, what success
9:53:51looks like, essentially what the
9:53:52definition of done is. Because at
9:53:54handover, you want to walk through each
9:53:55item and confirm that everything works
9:53:57the way that you agreed. And once the
9:53:59client confirms that the project is
9:54:00complete, you send the final invoice.
9:54:01You can frame this as simple as the
9:54:03project ends when the agreed workflows
9:54:05are live, tested, documented, and
9:54:06accepted. And then the project invoice
9:54:08is due. The next piece is deciding if
9:54:09you want to offer a maintenance
9:54:10retainer. A retainer is a separate
9:54:12ongoing agreement where the client pays
9:54:14you to keep the system up and running.
9:54:15This usually covers things like bug
9:54:16fixes, small tweaks, updates, dependency
9:54:18changes, monitoring, and basic security
9:54:20checks. It does not cover new features,
9:54:23new workflows, or major scope changes.
9:54:25Those should be a separate project. You
9:54:26can also set basic service levels so
9:54:28clients know what to expect. For
9:54:29example, a critical outage might get a
9:54:31response within a few hours. Minor
9:54:32requests might be handled within a few
9:54:34days. These don't have to be
9:54:35complicated, but the expectations there
9:54:36should be clear. Now, you also want to
9:54:38get clarity on ownership and IP. Many
9:54:40consulting agreements say that the
9:54:41client owns the work product once paid,
9:54:43but you should still protect yourself so
9:54:44that you can keep the right to reuse
9:54:46generic patterns or components that are
9:54:48not specific to their business, such as,
9:54:49like I said, reusable tools or
9:54:51subworkflows or basic templates. It also
9:54:53helps to define the exit process. So, if
9:54:55the client ever wants to move away from
9:54:56your services, you should outline what
9:54:58you will hand over. This could include
9:54:59exported workflows, documentation, and
9:55:01handover call along with what is
9:55:03included and what is billable. So, a
9:55:04simple explanation of this is once the
9:55:06product is paid for, the client has the
9:55:08right to use and run these workflows in
9:55:10their business. If they later want to
9:55:11move providers, you will help handle
9:55:13everything off in a structured way. For
9:55:14beginners, the main thing is simple.
9:55:15Stop doing all of this informally. Put
9:55:17scope, definition of done, payment,
9:55:19maintenance, service levels, bugs versus
9:55:21changes, ownership, and exit terms into
9:55:23a clear written agreement. When both
9:55:24sides know what they are buying and what
9:55:26happens after going live, projects run
9:55:28smoother and you avoid miscommunication.
9:55:29So now that you know what to do in
9:55:31theory, let's take a look at a real life
9:55:32example of a workflow that I sold and
9:55:34the process I went through when handing
9:55:35it over. So this one was a personal
9:55:36assistant workflow and one of the first
9:55:38ones that I ever delivered and the
9:55:39client had never actually heard of N& he
9:55:41just watched my YouTube video of the
9:55:42ultimate assistant and then reached out
9:55:43and said that he wanted something like
9:55:44that. So after discovery and after
9:55:46signing the contract, we got on a
9:55:48kickoff call and on that call I had all
9:55:49of these things that I needed from him
9:55:51which were listed out in sort of like
9:55:52our client expectations portion of the
9:55:54contract. So, I walked him through
9:55:56exactly what we needed to get and I
9:55:57helped him go get those API keys, sign
9:55:59up for an end account, things like that.
9:56:01And then I showed him how to invite me
9:56:02to that ended instance. Right there on
9:56:04the call, we connected his CRM, his
9:56:06calendar, his email, and the data
9:56:07sources that he wanted this assistant to
9:56:09be able to use. From there, I was able
9:56:10to just plug in a few of my own
9:56:11credentials for testing purposes [music]
9:56:12and then I could just hit the ground
9:56:14running. And the best part about this
9:56:15was at the end, a handover was almost
9:56:17instant because all I had to do was swap
9:56:18out a few of his credentials for mine
9:56:20and then he [music] could just start
9:56:21using it right away and giving me
9:56:22feedback. Now, with something like a
9:56:23personal assistant that is super
9:56:25autonomous, QA can be intense because
9:56:27it's super conversational and there's
9:56:28[music] lots of different tools that it
9:56:30could call. It's also client facing. It
9:56:31has memory. It has tone. It has lots of
9:56:34things that you need to actually make
9:56:35sure that the client is happy with. So,
9:56:37there was a lot of back and forth. There
9:56:38was a lot of tweaks, refinements, and
9:56:40specifically with the system prompt, and
9:56:41[music] that's completely normal. But
9:56:42throughout this process, he began asking
9:56:44for bigger features, new integrations.
9:56:46And at that point, I had to protect the
9:56:47scope because that's super important to
9:56:49make sure that we're not adding in all
9:56:51the stuff that you're not getting paid
9:56:52for. So I told him which of those
9:56:53requests would fit inside version one
9:56:55and which ones would need to be added to
9:56:57the backlog for a future phase. Then
9:56:59after version one was complete and
9:57:00accepted, we would scope out a new
9:57:01project around those extra features. So
9:57:03that alone saved me from doing a ton of
9:57:05unpaid work. And the final lesson here
9:57:06ties back to API keys. Early on, I used
9:57:09to try to make things easy for clients
9:57:11by running everything under my own
9:57:12billing and just sending them an invoice
9:57:14at the end of the month. Now it sounds
9:57:15nice in theory, but in reality, it gets
9:57:16messy fast. Like I mentioned earlier,
9:57:18clients want predictable costs and token
9:57:20usage is impossible to estimate
9:57:21perfectly. You also could deal with late
9:57:23invoices or confusion about what they're
9:57:25actually paying for. And it all just
9:57:26comes back to the same rule. It's so
9:57:28much cleaner and simpler and way more
9:57:29scalable if they own those accounts and
9:57:31their keys from the start. It makes the
9:57:32handover easier. It makes maintenance
9:57:34easier. And it keeps you out of the
9:57:36billing babysitter role. So hopefully
9:57:37seeing a real example gives you a better
9:57:39sense of how all the pieces come
9:57:41together in a live client project. This
9:57:42is how you build, host, test, handover,
9:57:44and maintain AI workflows without
9:57:46creating headaches for you or the
9:57:47client. So that's the full process.
9:57:49First, you decide where the workflow is
9:57:50going to live. I always recommend that
9:57:52clients self-host everything whether
9:57:53that's ended in cloud, a VPS or
9:57:55something local and you just help them
9:57:57configure it. Before you deliver
9:57:58anything, you make sure security and
9:57:59data privacy are handled the right way.
9:58:01Then you figure out who owns what with
9:58:02API keys and how those keys get into the
9:58:05system without creating a mess. After
9:58:06that, you run your testing and QA so you
9:58:08know that the workflow is reliable,
9:58:10safe, and producing the right outputs.
9:58:11Then you move into the actual handover,
9:58:13which is how you deliver the system, set
9:58:14expectations, and give them
9:58:16documentation. And finally, you close
9:58:17out the project on the legal and billing
9:58:19side and decide if there will be ongoing
9:58:21maintenance after go live. That's the
9:58:22full life cycle of building and
9:58:24delivering AI workflows the right way.
9:58:25So, I know that we covered a lot of
9:58:27information in this video. So, what I've
9:58:28done is I've thrown all of this into a
9:58:30full resource guide that you guys can
9:58:31access for completely free. All you have
9:58:33to do is join my free school community.
9:58:34The link for that is down in the
9:58:35description. If you enjoyed this one and
9:58:36you want to dive even deeper, then
9:58:38definitely check out my plus community.
9:58:39We've got over 3,000 members in there
9:58:41who are building businesses with NN
9:58:42every single day. So, it's a great
9:58:43environment to surround yourself with
9:58:45like-minded people. So, that's going to
9:58:46do it. If you learned something new,
9:58:47please give it a like and subscribe. It
9:58:49definitely helps me out a ton. And let
9:58:50me know what else you guys want to see
9:58:51in the comments. As always, I appreciate
9:58:52you guys making it to the end of the
9:58:53video. I'll see you in the next one.
9:58:55Thanks, everyone.
9:58:58If you guys are watching this, it means
9:58:59you made it to the end of the course.
9:59:00And I just want to say a huge thank you,
9:59:02but also congratulations. You just sat
9:59:04through so many hours of me talking. I'm
9:59:06sure you guys are sick of my voice. But
9:59:08like I said, I just wanted to say thank
9:59:09you. I put a lot of time and energy into
9:59:11this course, so I really hope that it
9:59:12was beneficial for you. So, if it was, I
9:59:14would really enjoy if you could leave a
9:59:15like, drop subscribe. That would mean a
9:59:17lot to me. And if you want to continue
9:59:19to support me, then definitely check out
9:59:20other resources. One of my biggest
9:59:22passions right now is AI Automation
9:59:24Society, as you can see by this
9:59:26sweatshirt that I've been wearing pretty
9:59:27much this entire course. You can check
9:59:28it out. The link for that is down in the
9:59:29description. It is my free community
9:59:31called AI Automation Society. And we've
9:59:32got hundreds of thousands of members.
9:59:34And my goal is just to make it a
9:59:35resource hub for people that want to
9:59:37learn AI, but also just like a really
9:59:38cool place for people to meet each
9:59:40other. and we're going to be doing some
9:59:42virtual events every quarter. We're
9:59:43going to be doing inerson events. So,
9:59:45that would be a great way to support me
9:59:46would be by becoming a member of AI
9:59:48Automation Society. And building on top
9:59:49of AI Automation Society, we also have a
9:59:51plus group. We're also going to have
9:59:53higher ticket tokate coaching coming
9:59:54out. So much more stuff is going to be
9:59:56coming in the AI Automation Society
9:59:57ecosystem, which I'm really pumped
9:59:59about. But that's my pitch. I appreciate
10:00:00you guys as always making it to the end
10:00:01of the video, and I'll see you on the
10:00:03next one. Thanks so much everyone.

About this transcript

This page contains the full transcript of Build & Sell with Claude Code (10+ Hour Course) by Nate Herk | AI Automation, generated from the public captions YouTube serves with the video. The transcript has 151,204 words across 20,888 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.