YT YouTube2Text

YouTube2Text → Transcripts → CLAUDE CODE MASTERCLASS 4 HOURS: Build & Sell (2026)

CLAUDE CODE MASTERCLASS 4 HOURS: Build & Sell (2026) — Transcript

by Michele Torti · 55,900 words · 7,917 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Welcome to the most comprehensive Claude
0:01Code masterclass, where I take you from
0:03a low beginner in Claude Code to someone
0:05who's actually able to build a gentle
0:07workflows and full systems for yourself
0:09or for other businesses. Now, I use
0:11Claude Code pretty much every single day
0:13to run my six-figure AI business, and
0:15I'm also teaching it to over 10,000
0:17people in my community. And so, I've put
0:19together this masterclass in the exact
0:21order that I wish I had when I had
0:22started out. And as always, no fluff, no
0:25jumping around, just a clear practical
0:26roadmap that takes you from zero to
0:29building real systems. Now, I've divided
0:30this masterclass into five different
0:32modules. So, here's exactly what you're
0:33getting. Module one is foundations. This
0:35is where we install Claude Code, go
0:37through how it actually works,
0:38permissions, setting up your Claude.md,
0:40your Claude directory, so Claude Code
0:41understands your projects before you
0:43even type a word. Module two is
0:45customization. This is where we build
0:46custom skills/commands,
0:48hooks that automate actions for you, and
0:50MCP, which connects Claude Code to every
0:53external tool in your stack. Module
0:54three is prompt engineering. And here
0:56I'm going to be breaking down how to
0:58talk to Claude Code so you get better
0:59outputs every single time. Module four
1:01is agentic workflows. This is where we
1:03talk about sub-agents, agent teams, and
1:05eight hacks I've discovered after
1:07building hundreds of workflows. And at
1:08the end, deployment. So, everything runs
1:1024/7 without you. And module five is the
1:13build. No more theory, you're going to
1:14watch me build real projects from
1:16scratch using everything that we
1:18covered. So, you can see exactly how it
1:20all comes together. Now, I'm also going
1:21to be leaving timestamps right below in
1:23this video, so you can stop the video
1:24and then you can come back to it and
1:26start wherever you left off. And as
1:27always, if you want all the resources
1:29and templates of this video, then you
1:31can check them out in the second link
1:32down below in my free private school
1:33community. You can go to the classroom
1:35section and you can go to the templates
1:36vault. Now, with that said, let's dive
1:38in.
1:43All right, here we have Claude Code.
1:45First thing is pricing. How much does
1:47this actually cost? Now, bear in mind
1:49that we are not going to pay Claude
1:51Code, we're going to pay Claude, and
1:53through paying Claude, we get access to
1:55Claude Code. We get access to Claude
1:56Code Work, and we get access to clock
1:58itself, okay? Now, clock has three
2:00different tiers. It has a free tier, and
2:02by the way, if you have the free tier,
2:03you do not have access to clock code.
2:04So, clock code, you have to pay for it.
2:07Then we have the pro tier, which is good
2:08for everyday productivity, which is 17
2:10bucks a month, and it gives you access
2:12to 45 messages for 5 hours. That is the
2:14limit on usage that you can use clock
2:16code. And this is the best tier that I
2:17can give beginners starting out, unless
2:19you're creating some really complex
2:21applications or you have a lot of
2:22automations running in your business.
2:24And then we have the max plan, so let me
2:26put myself here. And the max plan gives
2:28you 5x of the usage of the pro plan,
2:31which would be about 225 messages for 5
2:34hours, and you will pay about 100 bucks
2:36a month. I personally have the max plan,
2:38so I'm personally paying $100 per month
2:41on clock code, just because I use it
2:43every single day for my business, and I
2:44automate a lot of stuff that I have
2:46day-to-day. Now, you also have the
2:47option to pay 20x, right? So, instead of
2:505x, it'll be 20x of what the pro plan
2:52gives you, and you will also be paying
2:54double the money, which means that if
2:55you're on 100 bucks a month, you'll pay
2:57200 bucks a month. So, as a beginner, I
2:59definitely recommend sticking to the
3:00cheapest plan, which is a pro plan for
3:02now. And if you see yourself kind of
3:04exceeding the limit or exceeding the
3:05token cost or usage, then you might want
3:08to upgrade to the max plan. All right,
3:09so next thing is hosting. What does that
3:11look like? How do we set it up? And
3:12which applications do we use to actually
3:15use clock code? So, first thing is using
3:16VS code. VS code is what we call an IDE,
3:19very fancy way of saying a place where
3:22you can look at all your folders that
3:24you have with the clock code is using.
3:26And you can also install clock code, and
3:27you also have access to the AI of VS
3:30code itself, and it's completely free to
3:32to download, right? Now, similarly, we
3:34have cursor. So, cursor is a very
3:36popular one for coders just because
3:38cursor itself has their own AI. And
3:41their own AI is very, very good when it
3:42comes to coding and doing all the things
3:44that like building applications and so
3:46on. So, if you want the power of cursor
3:47plus clock code, meaning if you're a bit
3:49more on the techy side and you're a bit
3:50more of a developer, then I recommend
3:52using cursor. The only limitation of
3:54cursor is the limit. So, it's going to
3:56get a bit expensive, especially when you
3:58have to pay for cursor's AI, but also
4:00for the Claude AI. And I can tell you
4:02from experience that most times we just
4:04use Claude code inside these platforms.
4:06We never use the AI in these platforms
4:09at all. But I will usually go VS code
4:10cuz I don't use the AI VS code anyways,
4:13so to me it doesn't really matter. Then
4:14we have Anti-Gravity, which is the exact
4:16same as VS code. Anti-Gravity is one of
4:18the popular tools that came out
4:19recently, which is part of Google's
4:21suite. And Anti-Gravity is the exact
4:23same as VS code, and it also has its own
4:25AI that we can use. And I honestly see
4:27no difference between Anti-Gravity and
4:28VS code, unless you want to use the
4:31Gemini models inside of Anti-Gravity to
4:33build your own interfaces and make them
4:35look pretty, right? So, either one works
4:37fine. Then we have the terminal, which
4:38is where we are going to be using Claude
4:40code literally inside of our computer
4:42without having to go to Anti-Gravity or
4:44VS code. The only limitation with the
4:45terminal is that it doesn't look too
4:47pretty. It doesn't look very visual.
4:48It's sort of like Cody and it looks a
4:50bit complex when you're starting out.
4:51So, as a beginner, I do not recommend
4:53this. And then finally, we have Claude.
4:54So, Claude desktop also has the ability
4:56to use Claude code. The only limitation
4:59here is that you will not be able to see
5:01your files, and it won't be as
5:03structured as Anti-Gravity or VS code.
5:06Now, for today's video, we're going to
5:07focus on VS code and Anti-Gravity
5:09because most of the videos in this
5:11course are either using VS code or
5:13either using Anti-Gravity. But they're
5:14very, very similar, so I'll show you
5:15exactly how to download Claude code in
5:18each of these platforms. So, let's start
5:20with VS code. All you have to do is go
5:21to code.visualstudio.com,
5:23and you'll get to this page right here,
5:24the open-source AI code editor. You can
5:26press download for macOS if you're a
5:28macOS, or you can press download here,
5:31and you'll be able to go on this page,
5:32and you can choose your device right
5:34here. I'm just going to press Mac here.
5:36As you can see, now it's downloading the
5:38application itself. We have this little
5:39pop-up come up. It says, "Happy coding."
5:42We can drag this across.
5:44Now, it's going to download the actual
5:45application in our computer. Now, once
5:47you have it opened up, you'll get on a
5:48page that looks like this. So, this is
5:50Visual Studio Code, VS Code in the name.
5:52You'll have different buttons here, and
5:53you'll have the main interface right
5:54here. So, this is going to be the main
5:56place where we're going to be using
5:57Cloud Code and actually building
5:58different things, whether it's
5:59automations or applications. Now, in
6:01order to download Cloud Code, all we
6:03have to do is go down to the extensions,
6:05and I'll walk you through what each one
6:06means here, so don't worry. Extensions,
6:08you want to look for Cloud Code,
6:11and then you'll see the one from
6:13Anthropic. Make sure it's the one from
6:15Anthropic and not this one here or not
6:17these ones here. This is the verified
6:18one. So, once you press the extension,
6:20you'll get on a page like this. You can
6:21see Cloud Code for VS Code, and you can
6:23simply press install. As of the time of
6:25this recording, it has 10,263,533
6:29people, which is crazy, and it's the
6:30version 2.1.107.
6:32And it was published 10 months ago. So,
6:34very, very new. Now, I can press this
6:36button right here, which is for us to be
6:38able to open Cloud Code. And when we get
6:39on this page,
6:41I can simply say hello.
6:42Now, bear in mind that if you don't see
6:44this page right here to say hello and so
6:46on, it's because it's going to ask you
6:48to log in. Now, as you saw for me, I can
6:50say hello, how are you? because I
6:51already logged in to Cloud um inside
6:53here. But, if you haven't logged in to
6:55Cloud Code yet, then you most likely see
6:57this page that looks like this, which
6:58says that Cloud Code can be used with
7:00your subscription or used with your API
7:02usage, right? So, you can either pay
7:04through the Cloud subscription or you
7:06can use the API usage. Now, bear in mind
7:08that if you use the API usage, which is
7:10the Anthropic console, you will end up
7:12paying a lot more in the long run just
7:13because the usage is going to increase
7:16by a lot, especially when you use Cloud
7:17Code, and you won't even realize it.
7:19That's the thing. So, you can just press
7:20cloud.ai subscription. You can press
7:22open.
7:23It'll take you to this page here. You
7:25can press authorize. It says, "Build
7:26something great." That's all good. Now,
7:28I can go back, and you'll be able to see
7:29that now I can actually chat to Cloud
7:32Code itself. I can press X, and now we
7:34have the full interface. So, this right
7:35here will be the page where we actually
7:37use Cloud Code, and we can run up
7:38different conversations with Cloud Code
7:40itself. On the left-hand side, we have
7:42folders. Now bear in mind that anything
7:44you build in Claude code, and I mean
7:45anything, will be stored inside a
7:47folder. So when I'm building a website,
7:49when I'm building an application, so
7:51where is the HTML stored? Where is the
7:53document stored? Where are all the
7:55images stored, right? They're all stored
7:56inside a folder inside your computer.
7:58And this is great for us because you
8:00always have a place to store your exact
8:03files, right? And so on the left hand
8:04side we can press open folder. Let's say
8:06I add a new folder under desktop. I can
8:08press new folder. I can name it example.
8:12And I can press create.
8:13I can press open. You can press yes, I
8:15trust the authors. Now you won't be able
8:16to see Claude code anymore. So to see it
8:18again, you can press this button right
8:19here, Claude code, and you can press X
8:21again. And now you're back in this
8:23interface. So now we have the folder set
8:25up. We have everything here. We can
8:27start adding files. So file
8:29we can start adding folders. So folder.
8:32And that's how we start to add what we
8:34call the Claude directory. So the
8:35directory of Claude. So where are all
8:37the files stored? I can press delete.
8:40And we're back to zero. Okay, so
8:41everything that we build for this
8:42specific project that we're working on
8:44will be stored under the example folder
8:46right here. Now to show you how that
8:47looks like on your computer, if I add a
8:49file here and I say file.md, say hello.
8:53Okay.
8:54And then I press command S, or you can
8:56go to file save.
8:58Inside this folder called example, we
9:00have the file.md. And so now I can go to
9:02file, I can close the folder. And I can
9:05go back. I can open the folder again. Go
9:07to desktop. I can open the folder again.
9:09Press open. And now we're back exactly
9:11where we started. And so this way you
9:12can have different folders for different
9:13projects, and you can always come back
9:15to Claude code and start from zero. Now
9:16there are a lot of different things in
9:18the interface itself, so let's go
9:19through them step by step. So the first
9:21thing is the conversation history. So if
9:23you want to switch between
9:24conversations, the same way that you
9:25would do inside of Claude, you can go
9:27here to the Claude itself, and you can
9:29see all the past conversations that
9:31you've had. Or you can go to the web tab
9:33right here. And so now I can go to let's
9:34say write a LinkedIn post, which is a
9:36conversation that I had before. And we
9:37can always review the conversation that
9:39we had in case we missed anything or in
9:41case we want to keep going with the
9:42conversation itself. Then we have this
9:43right here, which is just the same thing
9:45as just pressing this button. Right, it
9:47spins up a new Claude instance, which
9:49means a new conversation. And then down
9:51here is where everything's at. So, on
9:53the bottom side we can see plus, in case
9:55you want to add folders or contacts or
9:56browse the web, you can simply do it
9:58here. We typically do this if you want
10:00to add more context to Claude code. Then
10:01we have the filter actions. So, there
10:03are a lot of different things that you
10:04can do inside of Claude code. One of
10:06them is using slash commands. So, you
10:08can put slash something. So, slash
10:10There's a lot of different things you
10:11can do. We can put slash clear
10:13conversation. We can put slash compact.
10:15We can do a slash everything. And we'll
10:17go through exactly what a lot of these
10:18slash commands are uh later in the
10:20video, so don't worry. And then one of
10:21the most important things here is
10:22actually the permission mode. So, if I
10:24press here to edit automatically, this
10:26is the permission mode that we're going
10:28to be using Claude code in. Right? Now,
10:30the permission mode might look different
10:32in 6 months. And it looked different 6
10:34months ago. Right? But, essentially what
10:36it is, it's the way that you're telling
10:38Claude code to act. Right? So, if I put
10:40ask before edits, then anything that it
10:42does, it will ask for approval before
10:44making each edit. If I put edit
10:45automatically, then now Claude will edit
10:47your text uh or the whole file
10:49automatically. If I put here plan mode
10:51as a permission, then what it will do is
10:52that it will create a plan for the
10:53actual thing. Then we have bypass
10:55permissions, which is where Claude will
10:57not ask for approval before running the
10:59potential dangerous commands. Right? And
11:01then right here, we also have the new
11:02feature which is called effort. So, if
11:04you want a very high effort for the task
11:06itself, then you can simply tell Claude
11:07code that you want higher effort. If
11:09it's lower, then you can put low. Now, I
11:10typically have this on effort medium
11:13just because I want to keep it balanced.
11:14Now, the default permission here is
11:15actually edit automatically. And so, the
11:17main ones are really ask before edits,
11:19plan mode, and bypass permissions. So,
11:21let me show you an example. So, let's
11:22say I tell Claude code to make a new
11:24file inside our folder or project and
11:27name it LinkedIn post. And then you can
11:30add a LinkedIn post there about life
11:32with a happy tone targeting people who
11:34are business owners in the US. Okay, so
11:36very strange task, but essentially it
11:38has to edit our folder. It has to edit
11:39something and it has to include the
11:41actual text itself. So, from edit
11:43automatically I can actually go to plan
11:46mode cuz I want it to plan before it
11:48does anything. And here I can just press
11:49go. So, now what happens is that Claude
11:50Code is given a task. It will then think
11:53about what it's doing. It will create a
11:55plan of action. As you can see here it
11:56says, "I'll create a quick plan for this
11:58and then we can execute it." And then
12:00the plan is something that we can review
12:01right here. Context, steps,
12:03verification. So, the plan is to create
12:05a LinkedIn post file. The user wants a
12:07new file in the project folder
12:08containing a LinkedIn post about life
12:10written in a happy tone and targeting US
12:12business owners. So, very strange
12:13request, but we can press yes auto
12:15accept. And now what happens is that it
12:17creates a file on our folder right here
12:19called LinkedIn post that I can open up
12:21and here we have the LinkedIn post. Now,
12:23obviously this is just an example to
12:25show you how plan mode works, right?
12:27Now, let's say I wanted to use bypass
12:29permissions. Bear in mind, this is not
12:31something that you see as default. So,
12:33you will not have access to this as soon
12:34as you get to VS Code. What you have to
12:36do here is go to the settings here and
12:38you can look for allow
12:41dangerously
12:42There we go. You can turn this on right
12:44here, which means that you're now
12:45allowing bypass permissions mode
12:47recommended only for sandboxes with no
12:49internet access. So, that makes no
12:50sense, but just make sure to turn this
12:52on. And all this mode does is that it
12:54will not ask you for approval for
12:56anything that it does. That's why it's
12:57called dangerously skip permissions
12:59because it's dangerous if you don't know
13:01what you're doing. It's dangerous if you
13:02give it access to different softwares
13:04and files and you're not there to
13:06actually see what it does, right? So,
13:08let's say I wanted to use bypass
13:10permissions and let's say I give it the
13:11exact same prompt right here. And let's
13:13say I said Instagram post, right?
13:16What you will see now is that it will
13:18not give us a plan. It will just
13:19straight away make the file itself
13:21without us having to do anything. As you
13:23can see, it made the file here called
13:24Instagram post life as a business owner
13:26that we can use straight away, right?
13:27So, on the one hand, this made a plan,
13:29it asked us for approval. On the other
13:31hand, it didn't ask us for anything and
13:32it made it itself. Now, the way that I
13:34see it is that I usually go to plan mode
13:36whenever I want to create a plan for a
13:38specific project that I'm doing. And
13:40then, once plan mode is done, if it's
13:42something that's a bit more complex that
13:43needs my input, I put it to ask before
13:46edits cuz it will ask me before it does
13:48anything. But, if it's just a very easy
13:49task, then we can put bypass permissions
13:52um cuz I don't want to be there
13:53approving everything all the time,
13:54right? Just understand what it is that
13:55you're building and what Cloud Code has
13:57access to. Now, like we mentioned, the
13:59other way is using a software called
14:01Google Antigravity. Google Antigravity
14:02is known for making websites, for making
14:04nice interfaces. You might have seen it
14:06on YouTube, right? But, it's also a way
14:08for us to use Cloud Code. And so, I can
14:10just go here to antigravity.google. You
14:12can press download and right here you'll
14:14be able to see your own uh computer. I
14:16can press download for Apple Silicon.
14:18You'll then have this page right here,
14:19similar to the VS Code. We can drag this
14:21across. Now, it's downloading
14:22Antigravity to our applications. And
14:24once you get to Antigravity, you'll get
14:26on a page that looks like this. So, very
14:28similar to the one that we had at VS
14:29Code. On the right-hand side, you have
14:31the IDE's AI, so Antigravity's AI, which
14:34is the agent here, which is similar to
14:35the one that we had in VS Code. And on
14:37the left-hand side, you have the
14:38folders, so the same thing as VS Code
14:40right here. And then, we have the
14:41extensions. We can go to Cloud Code. You
14:44can simply install if you haven't
14:45installed it. I already did. And once
14:47you have it installed, you can press
14:48this button right here and you can start
14:50working on it. In the same way that
14:51we're using the example folder in VS
14:53Code, we can go here,
14:56desktop, example, and we can open it up
14:58inside of Antigravity. We open it up, we
15:00can press yes, I trust the authors, and
15:02you'll be able to use the exact same
15:04folder that we're using in VS Code.
15:06We'll use it in Antigravity as well. So,
15:07that's why I say they're pretty much the
15:08same thing. And one more thing when it
15:10comes to setting it up, you can also use
15:11your terminal. So, if I go to terminal,
15:13new terminal here, you can simply press
15:15Cloud
15:17and you can say yes, I trust this
15:18folder, and you'll be able to use Cloud
15:20in terminal as well. So, if you say
15:21hello,
15:23you can see that it says hello, how are
15:24you? Now here we can see the model that
15:25we're using, how many context tokens we
15:27have, what plan you're on, the email
15:29that we're using, and then your
15:31organization, and then the actual folder
15:33example as well. Now as a complete
15:34beginner, I don't recommend using the
15:36terminal for now. Just use the actual
15:38interface here. It'll be much easier,
15:40especially if you're just starting out
15:41uh with the platform itself, cuz you
15:43don't want to overwhelm yourself too
15:44much. All right, so now you have Claude
15:45Code installed inside of Antigravity and
15:48inside of VS Code. If you're just using
15:49Claude Code, I honestly wouldn't see a
15:51benefit of using one over the other. So
15:53just choose one and stick to it and keep
15:55it very simple for yourself. But if
15:56you're someone who's building something
15:57a bit more sophisticated and you want
15:59Antigravity's Gemini model to help you
16:01with this, then I recommend just
16:03choosing Antigravity cuz you have the
16:05upper hand itself that you can use later
16:06on. All right, so last thing to cover
16:08here is built-in {slash} commands. So we
16:10mentioned this before that we have the
16:12little box right here, which gives us
16:14the ability to do actions without us
16:16having to give it a prompt. So here if I
16:17put {slash}, I can see that I can call
16:20any different action if I wanted to. And
16:22you can see we have context, we have
16:23model, we have customize, and we have
16:26{slash} commands. So these {slash}
16:27commands are things, they're actions
16:30that we can do automatically just by
16:31calling them specifically. As you can
16:33see, we have batch, Claude API, compact,
16:36context, cost, debug, extra usage, head
16:38dump, and we have a lot more here that
16:40we can use specifically. I'm going to go
16:42through a few of them just so you get
16:44the idea of how it works. So let's say I
16:45had a conversation with Claude Code. I
16:47say hello.
16:48And I said, how are you?
16:50Now there are cases where the
16:51conversation is very, very long and you
16:53want to start from scratch without
16:55having to go to a new conversation here.
16:57So what you can do here is you can press
16:59{slash} clear conversation. So {slash}
17:01clear conversation, you can go here and
17:03now you start from zero again. So now we
17:04have a new conversation and the next
17:06{slash} command is called context. So if
17:08I put {slash} context,
17:10and I press this one here, what Claude
17:11Code will give me is a breakdown of all
17:14the context and tokens that I'm using
17:15inside my account. So we have a million
17:18of tokens that we're able to use. So
17:20far, we used 17,900
17:22tokens, which is about 2%, and it this
17:25is just by saying, "Hello, how are you?"
17:26right? So, think of every conversation
17:28as a bucket, right? The more words you
17:30add to the conversation, the more water
17:32goes inside the bucket. And it comes to
17:34a point where it can't take any more
17:36water, and so you have to go to a new
17:37conversation, right? And this is what
17:39the percentage is. How much water is
17:41inside the actual bucket? How many
17:43tokens are we using for the conversation
17:44itself? Cuz a token is just words. And
17:46also, it tells us the category of usage
17:49by type. So, we have the system prompt,
17:51which is a prompt that we can't really
17:52see here, but it's taking up 6.2
17:55thousand uh tokens, which is about 0.6%,
17:58and then we have other things like
17:59system tools, we have skills, we have
18:01messages, free space, and buffer, as
18:04well, right? It's fine if you don't
18:05understand these terms right now, cuz
18:07we're going to cover them later down in
18:08the course. Then we also have MCP tools.
18:11So, I downloaded MCPs inside my cloud
18:13code, which is one of the modules that
18:14we'll go through later. And each MCP
18:16takes about X amount of tokens. So, 170
18:19tokens is used by the Gmail
18:21authenticate. Authenticating the Google
18:22Calendar takes 178. And so, the more
18:25MCPs you add, the more tokens. And then
18:27we have skills. So, skills are
18:29essentially instructions that Cloud Code
18:30uses to do a specific task. And each
18:33skill requires different tokens, right?
18:35And so, here you can see a breakdown of
18:38all the tokens that you're using, uh all
18:39the usage. Now, apart from clear
18:41conversation and context, we also have
18:45ones like compact, which is where you
18:47want to compact the conversation, so you
18:49can keep going in the same kind of
18:51context without having to go to a new
18:53conversation. We do that because every
18:55conversation has X amount of usage. So,
18:57let's say right now, we have a
18:58conversation that takes 90% of its
19:01usage. So, if you add {slash} compact,
19:03it will take that 90%, it will break it
19:05down, it will summarize the
19:07conversation, so that we can still have
19:09maybe another 60 or 70% left, right? So,
19:11we can keep working under the same
19:13thread without having to have a new
19:15conversation with Claude code. Then we
19:16have cost as well to see how much it's
19:18costing you, debug to debug, and then we
19:20have others as well, which we will cover
19:22later down the line.
19:25All right, so by this point you
19:26understand how Claude code works and how
19:28our permissions are the things that keep
19:30you in control. But right now, every
19:32single time that you start a new session
19:33with Claude code, it starts from zero
19:35cuz it doesn't remember anything about
19:37your project, which is why in the next
19:38lesson, I'm going to be breaking down
19:40arguably one of the most important files
19:42in Claude code, which is the Claude.md
19:45file. And I'm also going to break down
19:46the Claude directory. So, every single
19:48time that you start a new session,
19:50Claude code understands your project
19:51completely without you having to type a
19:53single word. Let's dive in. All right,
19:55so what is the Claude.md file? I'm sure
19:58you probably heard of it, you've seen it
19:59somewhere, but essentially what it is,
20:01it's the project instruction manual for
20:02Claude code. So, essentially Claude
20:04code, whenever we're working with a
20:06workspace, it needs to have some sort of
20:08memory. And the memory is where all the
20:09files are, what the project workspace
20:11actually is, what is our objective, who
20:13we are, and all that stuff, right? And
20:15so, that's exactly what the Claude.md
20:16file actually is.
20:18Now, the Claude.md file loads
20:19automatically, which means that when we
20:21work with a new workspace, it loads
20:23automatically in a way that before it
20:25starts doing anything, it always looks
20:26at the Claude.md file to get an
20:28understanding of the whole layout of the
20:29workspace before it dives in and does
20:32any single task. The same way that you
20:33would have an employee coming into a
20:35company, you give a memory, and every
20:37single day that he comes back to the
20:38company to work, he will have the memory
20:40pre-trained. So, in this case, whenever
20:42we have a new conversation with Claude
20:44code, we are saying, "Hey, before you do
20:47anything, just look at the Claude.md
20:49file, understand our workspace, and then
20:52execute the task that I want you to do."
20:53Now, the actual Claude.md file is
20:56divided into four different kind of
20:57steps or or pillars, right? The first
21:00one is product description. So, this is
21:02what this app does. It's not necessarily
21:04just an app, it's just saying this is
21:06what this workspace is, Right? For
21:07example, I have a YouTube workspace
21:09where I manage all my YouTube stuff. So,
21:11this workspace is specifically designed
21:13for YouTube ideation, content creation,
21:16XYZ. Right? Then we have the tech stack.
21:18This is looking at what tools and
21:20softwares we use. Then we have the rules
21:22and guardrails, which is what Claude
21:24must and must not do. And rules and
21:26guardrails can be anywhere from the text
21:28it should use, the colors it should use,
21:29the fonts, the step-by-step process it
21:31should go through to then create
21:32something. Right? And these are all the
21:34different rules. It's kind of general,
21:35but you can have like any rules that you
21:37can think of, you can put it there. And
21:38then finally, we have folder structure
21:40where everything lives. Now, this is one
21:41of the most important ones because as
21:43you will see when we make it the actual
21:44file, there is a structure of folders
21:47that we add inside our workspace. The
21:49reason that is is because we want to
21:51make things organized. And Claude knows
21:53exactly where to store all the different
21:54files. So, we'll have a file for the
21:56scripts. We have a file for the output,
21:59which is the things that it gives us. We
22:00have a file for the ideation. Maybe we
22:02have a file for all the passwords and
22:04keys, API keys that we use. Right? And
22:06so, in having things organized, it's
22:08much easier for Claude code to do things
22:10because it's not scattered everywhere.
22:12That's what it is. Now, what Claude code
22:14gains by having this file? Well, the
22:15first thing is that it understands your
22:16project better. And every single time
22:18that you speak to Claude code, you don't
22:19have to repeat all the things. Right? It
22:21already knows. The second thing is that
22:22it follows your conventions. This can be
22:24naming conventions, it can be patterns,
22:26it can be a structure that you use. Then
22:28like I mentioned, it looks at the rules.
22:29So, every single time this happens, do
22:31XYZ. Right? And it won't break what
22:33actually matters to you. And then
22:35finally, it's consistent. That is
22:36probably the most important thing
22:38because it's much better to have
22:39something that's average but consistent
22:42than something that's really, really
22:43high quality but inconsistent. Because
22:45with AI specifically, it's very good at
22:48doing things. Right? Like it's it's it's
22:50nuts. It's crazy. But, you also want to
22:52have strict guardrails, right? Strict
22:54prompts that make it so that it's not
22:56inconsistent every single time. It
22:58actually gives you the output that you
22:59want every single time when you are
23:01doing that specific thing. So, there's
23:03two different types of ways that we can
23:04set it up. The first way is using a
23:07three-layer architectural structure.
23:09Sounds complex, it's very very easy,
23:10actually. And the second way is running
23:13the {slash} in it. Now, the two
23:14different ways represent two different
23:16scenarios. The first scenario is method
23:18one, right? Where we write the
23:19{cloud.md} first, and then we tell
23:22Claude to initialize the project. And
23:24this is only when we have a new project,
23:27when we're starting from zero. Then the
23:28second method is when you already have
23:30an existing project. So, let's say
23:31you're building apps or maybe a landing
23:33page or maybe some agents, right? You
23:35haven't done the {cloud.md} yet. So, now
23:38what we do is we take all the folders,
23:40all the files, Claude will understand
23:42your whole workspace, and from there we
23:44tell it create a {cloud.md} file. So,
23:46one is creating {cloud.md} file from
23:47scratch, and then creating the folders,
23:49and the other one is you already have
23:51the folders, now we create the
23:52{cloud.md}. All right, method one is
23:54writing {cloud.md} from scratch. So,
23:56here we have the Claude Code project. I
23:58made a new project called Claude. And
24:00this is where we will make a file right
24:02here, just pressing this button, and
24:04name it {cloud.md}.
24:06And inside here, and by the way, the
24:07reason why it's MD is because it's
24:09markdown. Markdown is just a way that we
24:11format text, right? Cuz this is title
24:13one.
24:15This is title two, two hashtags.
24:17Then we have boldings, which is hello,
24:20right? And so on. So, with that said, we
24:22have the {cloud.md} file that we made.
24:23Now we have to actually write the
24:25instructions, and then initialize the
24:27project, okay? Now, one thing that I
24:29want to show you as well is if I go to a
24:31new chat, I'm looking here cuz it does
24:33mention "Make a {cloud.md} file for
24:35instructions, Claude will read this
24:37every single time." So, even Claude Code
24:39incentivizes you, it tells you, "Hey,
24:41make this file because it reads this
24:43every single time." Okay, so here we
24:44have the prompt that I'm going to give
24:46Claude Code in the {cloud.md} file. I'm
24:49going to copy this.
24:51I'll bring it back, I'll go here, and
24:53then paste it. So, what does this prompt
24:55tell Claude Code? "So, you operate
24:57within a three-layer architecture that
24:59separates responsibilities to maximize
25:01reliability. That's some fancy names.
25:03LLMs are probabilistic, while most
25:06business logic is deterministic and
25:07requires consistency. This system solves
25:09that problem. This means that LLMs are
25:12inconsistent, but business logic is
25:14consistent. And so, we want to solve for
25:16that inconsistency with this exact file,
25:18right? And then we give it the three
25:20different layers. So, layer one is
25:21directives. We're basically saying,
25:23"Anything that has to deal with what to
25:25do goes in the directives folder."
25:28Anything that deals with decisions goes
25:31and follows this framework right here.
25:33And then everything to do with doing the
25:34work, so the execution, is this folder
25:37right here. And we're also making a .env
25:39file where we store API tokens,
25:42passwords, and so on, okay? And then we
25:43have other principles as well. I can
25:46save it first of all, so
25:48it doesn't run. And I can say, "Can you
25:50initialize this project based on the
25:51claude.md file?" claude.md
25:55I'm going to press go. I'm going to make
25:57this bypass permissions. It's reading
25:59claude.md. Now it's looking at all the
26:00different things that are inside. And as
26:02you can see, it's starting to make the
26:03folders. It's making the folders, which
26:05are .env, executions, directives, .tmp,
26:08which is what we call the Claude code
26:11directory. The Claude code directory is
26:12essentially where all the different
26:14files are stored. And the Claude code
26:15directory usually looks like this,
26:17right? It's with these lines. So, we
26:19have Claude's folder, which is this main
26:21folder. And inside the Claude code
26:22folder, we have claude.md. We have .env,
26:26which is this one here. So, add your API
26:27keys, OpenAI, Google API, passwords that
26:30you're going to use. Execution,
26:32we have directives,
26:34and we have .tmp, which is empty. So,
26:36let's say I go to a new tab, and I can
26:38write this down. I can say, "Build me a
26:40landing page for Jim Solutions, an AI
26:42automation agency that helps businesses
26:44to automate their workflows using AI
26:45agents, uh which will have a hero
26:47section, what we do, testimonial
26:48section, and footer." I'm going to press
26:50go. I'm going to make this
26:52plan mode, so it plans the landing page
26:54before it does anything. As you can see,
26:56it's looking at the cloud.md file. It
26:59didn't find any brand guidelines, which
27:00is fine. Now it's creating a plan. All
27:02right, it just made the plan. We have
27:03the context, the tech stack, which is
27:05essentially what we're going to use, the
27:06structure, and then we have decision
27:08design, hero section. I can just press
27:11yes auto accept. I can make this bypass
27:13permissions
27:15because that means that it doesn't ask
27:16me to approve things all the time. And
27:18the whole point of showing you this
27:19landing page is so that you see how the
27:21different files of the landing page are
27:24stored in the different folders here.
27:25All right, it just made the landing
27:26page. I can open it up.
27:31As expected, it's pretty decent, to be
27:33honest, very simple. And you can see
27:35here that it made another folder called
27:36front end, which has all the different
27:38components of the landing page, but it
27:39also made a MD file for the directives,
27:42which is what to do, right? Directive,
27:44build landing page, objective, tech
27:46stack, page structure, service section,
27:48testimonials, footer, file management,
27:50and so on, right? And so this is showing
27:52you that all the different things, like
27:54we saw here, of what to do, right? SOPs,
27:57the objectives, the inputs, scripts,
27:59outputs, edge cases, are all going to be
28:01stored under the directives. All right,
28:04so now you got to see exactly how we
28:06made the folders from scratch by adding
28:08the cloud.md file first. But there's
28:10also method two, which is using {slash}
28:12init. So let's say I have a workspace
28:14called YouTube, and YouTube has a lot of
28:16folders, it has context, it has
28:17executions, outputs, all that stuff. It
28:19has all the ENV, which are all the
28:21different passwords and so on. But the
28:23one thing that is missing is the
28:24cloud.md file. And so in this case, we
28:27wouldn't use the three-layer
28:28architecture because the folders are
28:30already made. We already have the
28:32folders, right? Which could have been
28:33done manually by me or someone else. And
28:35so in this case, to create the cloud.md
28:37file, all I have to do is go to {slash}
28:40init.
28:42And now what it's doing is that it's
28:43going to look at the whole organization.
28:45It's looking at different files that it
28:46has.
28:47And it says, "Now I have a good
28:48understanding of the workspace. Let me
28:49create the cloud.md file." So, now what
28:51we've done is that we have the cloud.md
28:54file, which has the memory, it has the
28:56context of the whole workspace. So,
28:58whenever we're making an automation,
28:59whenever we're making something else,
29:01when we start cloud from scratch, like
29:02this, it basically has the
29:04understanding. And to show you another
29:05example, if let's say I go back to the
29:07one that I had before, let me delete the
29:09cloud.md. Let's say you had these done
29:11manually. I would go here. I would save
29:13this, first of all.
29:15And as you can see here, the first thing
29:16it's telling me is to tell cloud code to
29:18remember what you've told it using
29:19cloud.md. Well, in this case, I can
29:21press {slash} init again.
29:24Bypass permissions. What it will do is
29:26that it will look at all the different
29:27files that are here, which in this case
29:29is to build the landing page. As you can
29:30see, now it made the cloud.md file,
29:32which is the project overview, which is
29:35landing page for JM Solutions, an AI
29:36automation agency. It provides guidance
29:38to cloud code. So, the architecture,
29:40what we actually use to build things,
29:42the front-end structure, the commands,
29:44the design system, and the important
29:45notes, right? So, that's a snippet of
29:47memory that cloud code will now use
29:49whenever it makes something new. Now,
29:50the last thing I wanted to show you is a
29:52more advanced version of a cloud.md
29:54file. Because under my YouTube
29:55workspace, I have commands, I have
29:57skills, right? All these things, which
29:59allow me to do much more with cloud code
30:02than you would just by having the
30:03default features. And so, in this case,
30:05we can see that the cloud.md file, same
30:07thing here, provides guidance. Uh what
30:09this is is a cloud workspace template, a
30:11structured environment for working with
30:12cloud code as an agent assistant across
30:13sessions. Uh run {slash} prime at each
30:16start of the session to load essential
30:17context. So, if you're running from
30:19scratch, all you have to do is go here
30:21and put {slash} prime.
30:23And what this will do is that it will
30:24look at the whole structure, it will
30:26look at the cloud.md file, it will look
30:28at everything to have an understanding,
30:30the context that it needs to be able to
30:33tell me who I am, the workspace
30:34structure, and everything that it needs
30:36to know about me specifically. We have
30:37the workspace structure. Again, this is
30:38the cloud code directory. When you hear
30:40cloud code directory, that's all this
30:42means. So, we have Claude MD, we have
30:43shellalize.md, we have .env here, we
30:46have ncp.json, which is this file right
30:48here, the Claude commands and skills.
30:50So, Claude commands and skills. Then, we
30:52have references execution outputs,
30:54references execution output. Then, we
30:56have commands. It's telling me all the
30:57commands that I have. Then, we have the
30:59YouTube content pipeline, which are all
31:00the different skills, the thumbnails,
31:02analytics and monitoring, skill system,
31:04additional skills, critical
31:05instructions, and notes at the end as
31:07well. And that's the exact Claude MD
31:09file that I use every single time and
31:11the one that I always have sitting there
31:13as a way for Claude to understand who I
31:15am and to have context before I ask it
31:18anything. Quick one before we get back
31:19to the video. If you're looking to work
31:21with me one-to-one to be able to start
31:22and scale your own AI automation agency
31:24in a 90-day one-on-one mentorship
31:26program, then check out the first thing
31:27down below. With that being said, let's
31:28get into it.
31:34Hey, congrats on finishing module one.
31:36You've now got the foundations locked
31:38in. You have Claude running, you
31:40understand how it works, and most
31:42importantly, it knows your project
31:43inside and out. But as always, you've
31:45pretty much tapped into 5% of what
31:48Claude could can actually do, which is
31:49why in the next module, we're going to
31:51be taking Claude from a smart assistant
31:53all the way to building a fully
31:55customized system that's built around
31:58the way that you work. And it all starts
32:00with something that most people
32:01overlook, which is skills and {slash}
32:03commands. Let's dive in. All right, so
32:05we have Claude's brain, we have skills,
32:08which makes super Claude because Claude
32:11on skills is like Claude on steroids cuz
32:12it can do so much more than just the
32:14normal interface. What are skills? So,
32:16skills are instruction manuals that
32:18Claude reads before doing a task. So,
32:21the same way they use a recipe card,
32:23which are instructions to make
32:24something, in this case a chocolate cake
32:26or it can be a burger, it can be a
32:27pizza, whatever it is, which is a list
32:28of instructions that show you how you
32:30can do that thing, Claude has Claude's
32:31skills, which are instructions to handle
32:34PDFs, to create websites, to create
32:36presentations, and to do all those
32:37things. So, that when we have a task
32:39that you give to Claude code, let's say
32:41you say, "Let's make a website." And
32:42Claude, which is the agent, goes to the
32:44skills, right, and says, "Okay, which
32:46skill do I have to use?" [music] In this
32:47case, it will use the website skill,
32:49right? And it will use the website skill
32:51to give us a perfect result. Now, why is
32:53this important? Because without skills,
32:55Claude guesses, [music] and the result
32:57is inconsistent. The reason why the
32:58result is inconsistent is because it
33:00doesn't have a proper way of doing
33:01things if it doesn't have skills, right?
33:03Because skills are step-by-step
33:04instructions. And so,
33:05>> [music]
33:05>> if it doesn't have step-by-step
33:06instructions, it does its own thing,
33:08however it wants to do it in its own
33:10way. [music] And the way that AI does
33:12things can change over time, right? And
33:13so, what it does now might be different
33:15than what it does in 30 minutes, right?
33:16Now, with skills, Claude follows
33:18instructions, step-by-step instructions.
33:20And so, with that said, because it has a
33:22clear prompt that it has to go through,
33:24and that has to look at, the margin of
33:25error will decrease massively, which
33:26means that it will give us a much better
33:28output. Now that you know exactly what
33:29Claude could skills are on a general
33:31level, and why they're important, where
33:33do you actually store them? So, I'm here
33:34in my folder called the YouTube
33:35workspace, which is where I manage most
33:37of my YouTube stuff. Strategy,
33:38thumbnails, and ideation, titles, hooks,
33:41bodies, all that stuff. No, this is not
33:42scripted, so don't worry. Uh but
33:44essentially, if you see here,
33:46inside this folder, there are series of
33:48folders. So, dot Claude, brand assets,
33:50contacts, execution, outputs,
33:51references, scripts. By the way, if this
33:53is the first time that you've ever been
33:54introduced to Claude code, check out
33:55this video up here, cuz it will give you
33:56the overview of what most of these are,
33:59and why they're, you know, that way. But
34:00if you already know that, and if you
34:01already have the fundamentals, then this
34:02will make more sense. We will have a
34:04folder called the dot Claude folder.
34:06Now, the dot Claude folder, you'll see a
34:07dot behind it because you actually won't
34:09be able to see it in your folders.
34:10[music] I know that doesn't make sense
34:11yet, but I'll show you. But essentially,
34:13we have this folder called dot Claude,
34:14and inside the folder of dot Claude, we
34:16have another folder called skills.
34:18>> [music]
34:18>> So, that's where we store skills. So,
34:19now, if I go to a new page, and I name
34:21this general, I can have the folder, I
34:23can press this right here, and I can put
34:25dot Claude.
34:27Boom. And now, inside the dot Claude
34:28folder, I can then go here, and put new
34:30folder, and put [music] skills.
34:33Okay? So, now when we start adding
34:35skills, we add it to the folder itself.
34:36Now, why do we do this? We do this
34:38because we want to be organized, right?
34:40And so, Claude knows that every single
34:42time that it needs to look at a skill,
34:44it always goes to the skills itself. And
34:46it goes to all of them step by step.
34:47Now, if you noticed, the doc Claude
34:48folder also has commands. And so, you
34:50can also
34:52go here
34:53and you can put new folder and put
34:55commands.
34:56So, as you can see now, we have inside
34:58of doc Claude folder, we have commands
34:59and we have skills. Obviously, nothing
35:01inside, but we'll add them later. For
35:02the sake of this video, I'm only going
35:03to be talking about mainly skills, but
35:05that's essentially where we store Claude
35:06code skills. All right. So, right here,
35:08if you've ever seen something that looks
35:09like this, which we call the folder tree
35:11inside of Claude code, this is how the
35:13folders are stored. So, doc Claude is a
35:14folder, commands is another folder,
35:16skills is another folder. Inside each
35:18folder, there's obviously uh documents
35:20here. So, this contains only two
35:21documents, which we'll go through. And
35:23skills [music] contains different
35:24folders. And each folder is a skill.
35:26There's a lot of folders and skills that
35:27I've set so far, but essentially every
35:28skill that we add under skills folder is
35:30a folder in itself because a folder
35:32consists of documents. [music] If that
35:33didn't make too much sense, I hope this
35:35will. So, here we have doc Claude
35:36skills. Inside of skills, let's say we
35:37have the LinkedIn writer. So, let's say
35:38we want to make a skill that writes
35:40LinkedIn posts. That's it. Inside each
35:42skill, we have [music] three different
35:44things. Okay? We first have a skill.md,
35:47which is a process step-by-step
35:48instructions that we tell Claude code,
35:50like, "Hey, here's how the skill was
35:51done." And then we have two more
35:53folders. [music] Okay. So, each skill,
35:54which is a folder, has two more folders.
35:56The first folder is called references.
35:57Now, just so you know, these two,
36:00they're not always going to be there.
36:01Like, sometimes you need reference docs
36:03or examples or templates. Sometimes you
36:05don't, right? So, this is like if or or
36:07you need it. But in the case that you do
36:08need them, we'll have references, and
36:10this includes documents of examples or
36:12templates or reference docs, anything
36:13that you need there. On this side, we
36:15have scripts. Scripts is a Python codes,
36:17which is again optional in case you need
36:18them. But essentially, if you're running
36:19some sort of automation, the code
36:21represents the automation itself. Okay.
36:22So, this is the instructions, "Hey,
36:24here's how you do the task." References
36:25is like, "Here's supporting documents
36:26that you would need to make the task
36:28happen. And the scripts will be like the
36:29Python codes, which you can think of
36:30like the different N 8 10 nodes, the
36:32different steps of the automation, which
36:34are code, right? So, that's exactly how
36:35it's stored. Then going back to my
36:36previous example here of my YouTube
36:38workspace, which is where I store
36:39everything for my kind of YouTube
36:41projects. If I go to Doc Cloud, this is
36:42the folder that contains the skills
36:45folder as well. It also contains this
36:47right here, agent memory, agents and
36:48commands, which I'll walk you through
36:49later on. But, skills is the one that we
36:51have to worry about. And if I press
36:52skills, you can see that it has these
36:53many folders, which are all skills
36:55themselves. We have Cloud MD generator,
36:57we have comment analysis report, we have
36:58competitor analysis, Excalidraw diagram.
37:00We have a set of different skills that
37:01do different things. And if I press on
37:03one of them, let's say comment analysis
37:05report,
37:06I'll be able to see the folder of
37:08references, the folder of scripts, and
37:10then [music] the skill.md, which is a
37:11document that outlines everything. Okay?
37:13And so, skill.md, references, and
37:15scripts are all folders inside the main
37:17folder of LinkedIn writer inside the Doc
37:20Cloud skills folder. So, they're all
37:21folders, right? One folder here, one
37:23folder here, one folder here, one folder
37:24here, one folder here, and then we have
37:26the skill.md. Okay, but that's the
37:28structure of different folders inside
37:30Doc Cloud. And this is pretty much the
37:31same kind of structure as the competitor
37:32analysis. If I go here, I also have
37:33references and scripts. This also has
37:35the JSON file, which is just something
37:37that we use to store Google tokens. But,
37:39some of them only have the skill.md. And
37:41so, that's why I say that this is
37:44mandatory. You always need to have the
37:46skill.md. The references and scripts,
37:48you don't always need to have them.
37:49Because this right here, for example, we
37:50don't need any reference files, right?
37:51We just need a skill.md, which has a
37:53bunch of instructions as to how to
37:55actually make the Excalidraw diagram.
37:56So, here I have five different what we
37:58call marketplaces. So, the same way that
38:00you go to Amazon to get a product, we go
38:03to this marketplace to get a skill.
38:04[music]
38:05The only difference is that these are
38:06free. Which means you don't have to pay
38:08for the skill itself. At least as far as
38:09I know. I'm sure there are skills out
38:10there that you have to pay for, but as
38:12far as these ones, you don't have to pay
38:13for them. Now, the sites here, I'm going
38:14to leave the links all down below so you
38:16can check them out. But, I have AITMPL,
38:17which is a Cloud Code templates, which
38:19is the one that I personally use pretty
38:20much most of the time. Then we have this
38:22right here, which is also pretty decent
38:23one. Then we have this
38:25bunch of skills as well. Then we have
38:28this here. We have this as well that
38:31contains different skills if I go here.
38:33There we go. Now, it might seem a bit
38:36intimidating at first when you look at
38:38this just because it does look very,
38:39very techy. I mean, all of this didn't
38:41really make much sense to me when I
38:42started, but it's actually very, very
38:44simple once you break this down. So,
38:45let's say I go to the CloudCode
38:46templates, okay? And like I mentioned,
38:48skills are just pre-built instructions.
38:50It's a prompt, right? It does a specific
38:51thing. So, let's say [music] I wanted to
38:53create a website. Okay, if I wanted to
38:55create a website, I would definitely use
38:56the front-end design. Why? It's because
38:58the front-end design creates distinctive
39:00production-grade front-end interfaces
39:01with high design quality. Use a skill
39:03when the user asks to build web
39:05components, pages, artifacts, posters,
39:06or applications, dashboards, React
39:08components, HTML, CSS layouts, or when
39:10styling, beautifying any web UI. [music]
39:12It generates creative, polished code,
39:14and UI design that avoids generic AI
39:16aesthetics, okay? So, this is great when
39:18you're building website, when you're
39:19building a UI, which is a front-end, the
39:21thing that you see inside of software.
39:23Um and so, we use this here. Now, this
39:24skill has different components. It has
39:25the name right here. It shows you the
39:27amount of downloads that a skill has,
39:29and typically the higher the downloads,
39:30the better the skill is. And front-end
39:32design is one of the most used skill
39:33there is out there, which is great. We
39:35have the install command. And what this
39:37is is a command where you can copy this
39:39and you can simply just paste it into
39:40CloudCode, and I'll show you how to do
39:42that, and it will automatically install
39:43the skill without you having to give it
39:45any instructions, okay? Then, we have
39:47the description of the actual skill. We
39:48have the license, which is the just a
39:50document to I'm pretty sure it's to not
39:53use copyright, like to not copyright the
39:55the actual skill so you don't copy it
39:56and then put it on your own marketplace,
39:58right? It's it's owned by this person.
39:59And then we have the preview, which is a
40:01preview of the actual skill and what it
40:03is and what it does. And then, most
40:05importantly, we have the code. And the
40:07code really is a skill.md document. So,
40:09when we spoke about the skill.md here,
40:11all it is it's this. It's this prompt
40:13right here. And so, every skill is just
40:15a prompt, right? Don't make it more
40:16complex than it is, because that's all
40:18it is. Just a prompt that tells Cloud
40:19Code what to do. Now, if I go here, just
40:21to show you how it looks like on a
40:22different website, content ideas. Let's
40:24do content content creation. Content
40:25creation, let's do a random one here.
40:27And here, we can see that we have the
40:28name of the skill, we have the
40:30description, we have the command, which
40:32is what we talked about before, and then
40:33we have the file. So, in comparison to
40:35here, where we can we are able to
40:37actually copy the prompt and paste it to
40:40to Cloud Code or do whatever. I mean,
40:41usually you just use a command, is the
40:43easiest thing to do. Just paste the
40:44command and it will do everything for
40:46you. Here, you also have the command,
40:47but you have an option to download the
40:49documents. Okay, because this right here
40:51will be a document [music] for the
40:52skill.
40:53Okay, with the skill.md document. And
40:54that's it. Instead of downloads, you
40:56have stars. So, 57,000 stars tells me
40:58that it's pretty good and that a lot of
41:00people have downloaded it, which tells
41:01me positive things about the skill. If I
41:03go to here, I can see that I have Let's
41:06do a file search.
41:08Let's see what this shows us. One-click
41:09install,
41:11one-click copy. So, we can copy this.
41:13File explorer with the skill.md. Here's
41:15the full prompt.
41:16So, they're all like I mean, the
41:17websites look a bit different, but they
41:19all have the exact same thing, right?
41:20They have the description, they have the
41:22the skill itself. The most important
41:23thing is you understanding what skill do
41:26I need to be able to do what I want to
41:28do. And then, going into these
41:29marketplaces and then looking at the
41:30skill that you need to then be able to
41:32copy and then paste it. I'm going to
41:33copy this and I'm going to go to Cloud
41:35Code in a new version here and I'm going
41:37to paste this here.
41:38Okay, and I'm going to go and run. Now,
41:40it's asking me if I can do this. I'm
41:42going to press yes and actually turn
41:43this
41:44to
41:45bypass permissions. As you can see now,
41:47it made a folder called dot cloud. If I
41:48go inside, I can see that it made a
41:50folder called skills. And inside skills,
41:52it made a folder called front-end
41:53design. And inside front-end design, it
41:54had the skill.md document, which is a
41:57full prompt that we saw here.
42:00The exact same thing. And so, like I
42:02showed you before, I made the folder
42:04manually, but you can also have Cloud
42:05make the exact same folder automatically
42:07for you. Now, let's say I want to use a
42:09different skill. So, let's say I went
42:10here and I said, "Okay, let me use the
42:12senior architect." Okay, I go here, I
42:15copy this, and then I simply just paste
42:17it here.
42:18Actually, let me go to a new window.
42:20And then I can go and press done. I can
42:23go to bypass permissions, so it doesn't
42:24ask me every time to approve. And now
42:26you'll be able to see a new skill being
42:28added here. There we go, senior
42:29architect. It's as easy as this. But,
42:31senior architect in comparison to
42:32front-end design has references, it has
42:35scripts, and then it has a skill.md.
42:36[music] So, sometimes it needs
42:37references, sometimes it needs scripts
42:39to be able to fulfill the actual skill,
42:40but sometimes it doesn't, right?
42:41Sometimes you just simply have a prompt,
42:43which is just instructions [music] of
42:44what it needs to do. Now, bear in mind
42:45that the skill.md document is just a
42:48list of instructions, and it's a
42:50standard [music] across all the
42:51different LLMs. And so, whether you're
42:53using Claude, Cursor, Codex, Gemini CLI,
42:57or Copilot, or any other of these 20
42:59different tools, [music]
43:00you will be able to apply this exact
43:02document to all of them. And that's why
43:03it's a very, very important document,
43:05and it also makes skills way more
43:06valuable than people think. [music] I'm
43:08going to go here and delete the skill
43:09right here. Cuz I'm not going to need
43:11it. But, next, I want to show you how to
43:12build a website and how Claude builds a
43:14website before skills and then after
43:16skills. Okay, so I'm also going to
43:17delete this, front-end design. Actually,
43:19I'm going to delete the whole thing, so
43:20Claude
43:21folder, delete it. I'm going to go here,
43:22and I'm going to say, "Can you make a
43:24website for my agency, JM Solutions?
43:27It's an AI automation agency that helps
43:29marketing agencies to build delivery
43:30systems and helps them to double their
43:33client delivery in 30 days." I'm not
43:35going to put too much effort into this.
43:36I want to keep everything equal and just
43:37show you the output, right? I'm also
43:39going to say, "The color scheme is going
43:40to be dark blue, uh and yeah, give me
43:42the landing page in a link that I can
43:44access on my web." I'm going to go here
43:46on plan mode, cuz I want it to kind of
43:48plan first before it gives me the actual
43:51website, and then we can go bypass
43:52permissions to to see the full thing.
43:54All right, so after a few minutes, we
43:55got the plan here, which is a plan on
43:56the landing page, dark blue, contacts,
43:58approach, color, sections. So, we press
44:01yes and auto accept, and then we go to
44:03bypass permissions, so I don't have to
44:04do this every single time.
44:07Bypass permissions, [music] and that's
44:08it. All right, I just made the landing
44:09page. I'm going to copy this.
44:11I'm going to paste it here.
44:12And here we have [music] the landing
44:13page itself. Get your client delivery in
44:1530 days. So, as you can see, this is the
44:17design style that it made, okay, with
44:19the actual cal.com link and everything,
44:21okay? Which is not bad, which is not
44:22bad.
44:23>> [music]
44:23>> Now, let's say I go back and I delete
44:25this.
44:27And
44:29Okay, and I give it the exact same
44:30prompt.
44:31But now, I'm going to use the skill
44:33front-end design, copy the command, and
44:35then paste it here. And now we're going
44:37to download the skill and then do
44:38everything again and show you the end
44:39result. Okay, so it just made the skill
44:41docload folder skills front-end design,
44:43and we have the skill.md, which is a
44:44prompt. Cool. Now, I can go here. I'm
44:47going to paste the exact same prompt.
44:48The only thing I am going to say is, can
44:50you use a different localhost uh number
44:52to give me the link? Just because it
44:54gave me the link of localhost 8080, and
44:56I want to show you the difference
44:57between this and that. So, I can press
44:58this, I can go to plan mode, it'll make
45:01the plan, and then we go directly to
45:02build the website. Okay, it just made a
45:03plan. I can press yes, auto accept, and
45:06then go to bypass permissions. All
45:07right, now it made us a new link, so I
45:08can copy this as well. I can go here.
45:10You can see this is the old website.
45:12I can paste the new link. And we'll have
45:14the new website here.
45:18Which is much nicer.
45:19It has numbers as well. It has
45:20testimonials in case we need any. And we
45:22also have the option to book a call as
45:24well. It says no obligation, no pushy
45:25sales, just a clear road map for your
45:27agency. So, as you can see, with the
45:29skill right here, it has made the
45:30website a lot better because of the
45:32prompt that we gave it here. And you can
45:34apply this thing to pretty much anything
45:35that you do with these pre-built kind of
45:37skills that you can see on the
45:38marketplace. And now that you know
45:39exactly how to go to a marketplace and
45:40use someone else's skills, in the next
45:42section we're going to look at how you
45:43can build your own skills step-by-step.
45:45[music]
45:45I'm going to be building my own, show
45:47you the full process, so you eventually
45:49you can do the exact same. Now, to be
45:50able for us to make the skill, you need
45:52to understand the foundation of how a
45:53skill is actually made, right? Now, you
45:55know that a skill here has different
45:58folders. So, for example, comment
46:00analysis report has this folder, this
46:02folder, and this document, okay? Now,
46:03this document is by far the most
46:04important document because it has the
46:06instructions. Because not all the skills
46:07have the reference and scripts, but all
46:09of them do have an MD document, okay?
46:11Now, let's go here to this one, for
46:12example, and let me actually pull up
46:14another one as well. Let me actually go
46:16here. Put this in the right. So, you can
46:17see the the same kind of structure here.
46:19So, first thing that we have with the
46:20skill that I'm the document is the name
46:22and description. So, the name and
46:24description are used because when you
46:26tell Claude, "Hey, I want to make an
46:28Excalidraw diagram or I want to fetch my
46:30competitors' uh comments, right?" What
46:32it does is that the first [music] thing,
46:34it looks through all the skills
46:36that we have here, which are going to be
46:37like 12. Only then does it look at the
46:39skill that I'm the It looks like this
46:41first part, name and description,
46:42[music] and then it chooses which skill
46:44it's meant to use. The reason why it
46:46does that is because if you don't have
46:47this small snippet of name description,
46:49it will go through all the files of all
46:51the skills, and it will consume so many
46:53tokens, and it will cost you a lot more.
46:55And so, here we have a preview saying,
46:57"Hey, Claude, if you want to use the
46:59skill, here's the description, and this
47:01is the name. If you mention something
47:02similar to this, let's use that." It
47:03also increases the accuracy of which
47:05[music] skill it uses. So, first things
47:07first,
47:08name, descriptions. We have the name of
47:09the skill, Excalidraw diagram. Then we
47:11have the description, "To use when
47:12someone asks to draw a diagram, make an
47:14Excalidraw diagram, or build an editable
47:16diagram, default for all diagram
47:18requests, okay?" The whole file here is
47:20done in Markdown formatting because
47:21that's the way the AI knows how to put
47:24the text in hierarchy. So, heading two,
47:25heading one, heading three, right? So,
47:27it knows which one is more important,
47:28right? So, for this, it's a heading two.
47:30This is a heading three. And it goes
47:31through step by step what the skill
47:33actually is. Okay, so step one,
47:34understand the request.
47:35>> [music]
47:35>> Step two, research if needed. Step
47:37three, plan the layout. Step four is
47:39generate elements. Step five, save and
47:41deliver. And then we go through step
47:43six, handle and feedback, critical
47:45rules, design principles, layout
47:47systems, and then color systems,
47:49typography scales, element schema, and
47:51then we have the JSON file, which is the
47:53type of design that Excalidraw, which is
47:56a diagram that we use to make different
47:57diagrams, makes the [music] diagrams
47:59itself, okay? So, that's a skill right
48:00here. On the other hand, let's say I go
48:02here, comment analysis report, which is
48:03a report that I personally made, right?
48:05That I actually made it step by step. We
48:07have the name description, then we have
48:09comment analysis report prerequisites
48:11for the skill itself. This is what it
48:12does. We need this to start, so let's
48:14grab this first. Step one, fetch
48:16comments. So, what it does is that
48:17inside the skill, it goes through the
48:19scripts fetch comments.py. So, scripts
48:21fetch [music] comments.py, which is a
48:22first, you can say, automation of the
48:24whole system. Fetch the top 200 comments
48:26by relevance. Look at these channels
48:28right here. Step two, analyze and write
48:29a report. Step three, generate PDF. And
48:31then it goes to error handling and
48:32partial runs and all that stuff in case
48:34it breaks, okay? And so, that's exactly
48:36how it's made. Now, here I have a empty
48:38folder inside of cloud code, which is
48:40general, which is the one that I used
48:41before. I'm going to go here and create
48:43a dot cloud
48:44folder. Inside dot cloud folder, I'm
48:46going to create another folder here
48:47called skills.
48:49And then, I'm also going to create
48:50another folder inside of dot cloud,
48:52which is called commands,
48:54which is a new thing. Don't worry, I'll
48:55walk you through it. Okay? So, commands
48:57will be the thing that allows me to call
48:59the skills itself. So, for example, if I
49:00want to call comment analysis report, I
49:02don't have to type comment analysis
49:03report. I can do {slash} comment
49:06analysis report, and this is the slash
49:08command. So, I don't have to type the
49:09name, it automatically goes there. It
49:11automatically calls the skill itself,
49:12which is a good thing to have when we do
49:13our skills. So, inside skills now, I'm
49:15going to create a skill, it's a new
49:16folder inside. I'm going to name this
49:18YouTube title
49:21generation.
49:23Boom. And inside of YouTube title
49:24generation, I'm going to create a file,
49:27>> [music]
49:27>> which is called the skill.md.
49:29And the skill.md will be the list of
49:31instructions of that specific skill. So,
49:32first thing I'm going to do, I'm going
49:33to paste this here, which I already
49:34wrote before, and this will be the name
49:36and the description of the skill itself.
49:38So, we're making a YouTube title
49:39generator. We generate YouTube video
49:40title ideas. Use when asked to
49:42brainstorm titles, come up with video
49:45names, or generate title options for a
49:46YouTube video. And then now here, I can
49:48paste this quick sentence saying,
49:50"Generate 10 title ideas for the given
49:51topic. Titles should be under 60
49:54characters, create curiosity, and be
49:56relevant to a tech AI YouTube audience."
49:58And I can also say,
49:59"Here's a few title frameworks
50:03that work well." And I can paste this
50:05here,
50:06which are some title frameworks that we
50:07use for YouTube videos. So, for example,
50:10I failed that topic until I learned
50:11this. I never achieved or never achieved
50:14goal until I learned this. Learn the
50:15[music] skill and never unwanted action
50:17again. If you action like this, you're
50:19going to be negative situation forever.
50:20Cuz these are all title frameworks. And
50:22so, that's what here will be the skill
50:23that I'm going to use. I'm going to
50:24press command S to save or you can go
50:26here, file, save.
50:28And I'm going to test it. Okay? So, I'm
50:30going to go here and say, "I want you to
50:32help me with title generation.
50:34I want to create titles for my new
50:36video, which is about how I broke past
50:39the six-figure mark with my AI agency
50:41step by step." I'm going to press go.
50:43And now, what it will do is it will
50:46think, it will look at the YouTube title
50:47generation skill, because that's the
50:49skill that I need to use.
50:50>> [music]
50:50>> And now, it will give me 10 title ideas
50:52for my video, okay? Which is great. And
50:54so, what did we do? We created the skill
50:56manually, and now we're testing. [music]
50:57So, we're testing that this actually
50:58works. And as you can see, it understood
51:01which skill to use. Of course, we only
51:02have one right now. And by the way, just
51:04because it has YouTube title generation,
51:05it doesn't mean that it's going to use
51:06that every single time. And so, we said,
51:08"Hey, I want to help with titles. I want
51:09to create a title for my new video." And
51:11then, it gave me 10 title ideas, which
51:12is the exact thing that I told it to do
51:14in that skill. I never hit 100K until I
51:16learned this skill. Give me 50 minutes
51:17and I'll show you how to build a
51:18six-figure AI agency. The smartest
51:19strategy to hit 100K with my AI in 2026.
51:22These are pretty good uh titles. It's
51:23not fun, but it'll get your agency to
51:25100K way faster.
51:26>> [music]
51:26>> I like that. And so, now you saw me
51:27actually build a skill step by step. So,
51:29now that you know how to make it
51:30manually, there's actually a trick that
51:31we use when making skills. And that's
51:33using the skill skill creator. So,
51:35there's a skill that actually makes
51:36skills, which makes the whole process
51:38much easier. Now, why does it do this?
51:39It's because I'm going to press go. So,
51:41you can actually find a skill under
51:42manage plugins.
51:45And you'll be able to see skill creator.
51:47So, you can find it here. I can actually
51:48just
51:49install it. I can look at skill creator.
51:52And you can install it here. Okay, so
51:53now [music] it's installed in your
51:56MacBook. Now, it it's asking you, "Do
51:57you want it available for all your
51:58projects? Do you want it available just
52:00for collaboration or just for you? Only
52:02for this exact title or for this exact
52:04project?" I like to have this skill
52:06installed across all my projects, which
52:08means that first of all, I won't be able
52:10to see it here, but it will be installed
52:12in a way where it's always available to
52:14you. And you don't have to reference it
52:16or do anything cuz it's automatically
52:17going to be used whenever you're making
52:19a skill. And so here I'll say install
52:20for you, available in all your projects.
52:22And now whenever I'm making a skill, it
52:24will use this skill to make that skill.
52:25Now, where does the skill creator come
52:27from? So, the skill creator is something
52:28that Anthropic actually came out and
52:30made, which is why we use it all the
52:31time. And it's a skill that allows us to
52:35>> [music]
52:35>> be able to use whenever we're creating
52:36skills. So, you can see skill creator, a
52:38skill for creating new skills and
52:40iteratively improving them. At a high
52:41level, the process of creating a skill
52:43goes like this. Decide what you want the
52:45skill to do and roughly how it should do
52:46it. Write a draft of the skill. Create a
52:48few test prompts and run Claude with
52:50access to the skill on them. Help the
52:51user evaluate the results both on a
52:53quality level but also on a quantity
52:55level. Rewrite the skill based on
52:56feedback, repeat until you're satisfied,
52:58and expand the test and try again on a
52:59larger scale. And now it goes to the
53:01whole process to make the skill, which
53:03is much more efficient and even a higher
53:04quality rather than us making it
53:06manually. So, we're going to make a lead
53:08generation system skill, okay? And this
53:11lead generation system skill will start
53:12by
53:14the input,
53:15which is niche,
53:18location,
53:20and amount
53:21>> [music]
53:21>> of results.
53:22And that will be me just giving it to
53:23Claude like on a chat basis, just fine.
53:25And here we can use Apify
53:28Google Maps scraper. By the way, if you
53:30don't know what Apify is, it's like the
53:32Amazon for scrapers. You go inside, you
53:34look at one scraper, which allows us to
53:36be able to go in Google Maps, extract
53:38the results based on what we want.
53:40[music] So, in this case, we want, let's
53:41say, roofing companies in North
53:43Carolina. It will script all of them.
53:45Obviously, we didn't want a results. And
53:46then, it will go actually filter
53:50leads by website exists.
53:53And then, it will [music] go inside
53:56the website.
53:57And then, it will
54:00give me a personalized
54:03icebreaker
54:05that I can
54:07use
54:09for LinkedIn. Okay?
54:10So, this is an automation, right?
54:12Something you would build in any time.
54:13We're going to build it here on
54:14Clockwork. And we can simply call it
54:16every single time that we want. So, the
54:18first step I'm going to do is explain to
54:19Clockwork roughly what I want, okay? So,
54:21I'm going to say this. Hey, so the task
54:23of today is to make a lead generation
54:24system. The first step is me giving you
54:26a niche that I want to target, the
54:28location of that niche, and the amount
54:29of results that I want. The second step
54:31is you have to go to Apify and use the
54:33Google Maps Scraper to be able to do
54:35this. Then, when you get the list of
54:36leads, you have to filter by which leads
54:38have a website. The ones that don't have
54:40a website, just discard them. Then, the
54:42ones who have a website, just go inside
54:43the website, script the website, and
54:45then finally, based on a personalized
54:46template that I'll give you, you have to
54:48make a personalized icebreaker that I
54:51can use to connect with people for
54:52LinkedIn. The output will be a Google
54:56Sheet with
54:58company name,
55:00location,
55:02website, phone number,
55:05and it's fine if they don't have a phone
55:06number. And then, ice-
55:08breaker. Okay? That's it. Ask me any
55:10questions if you have any, and let's go
55:12from there. Okay? That's the prompt.
55:15Now, bear in mind, I didn't set up
55:16Apify. I didn't set up anything. So,
55:17we're starting from scratch, right? I'm
55:19going to go to plan mode, so so we can
55:20plan things out. And as you can see, I
55:22didn't tell it to create the skill yet,
55:23because I'm going through everything
55:24step by step first, and then I tell it
55:26at the end, "Hey, now that you've done
55:27the whole thing step by step, you know
55:28what went right, you know what went
55:29wrong, just make the skill, okay?"
55:32I'm going to press go. So, how do you
55:33want to interact with the Google Maps
55:34API? Do you already have an API key or
55:37service set? I'm going to say Apify. How
55:39should the output Google Sheet be
55:41created? Google Sheets API. Let's say we
55:43do a CSV file to make things easier.
55:45Tech stack. What language framework do
55:46you want this built in? Python scripts.
55:48Yeah, I definitely do not want to be
55:49building an unattended workflow or else
55:50that would not be the point of this
55:51video. Submit answers. And if you're
55:53wondering whether something like this
55:54would be built in an unattended versus
55:56cloud code, I would definitely build
55:57this in cloud code cuz much easier as
55:59well. And we don't have to [music]
56:01pretty much worry about much. Now it's
56:02asking me, do you already have an API uh
56:04key for Apify? So, I don't so I can
56:06actually go find it. apify.com. And by
56:09the way, if this is the first time that
56:10you hear about Apify, I would definitely
56:12recommend checking out this video up
56:13here which goes through the full
56:14tutorial on how it works. [music] I'm
56:16saying that because it is one of the
56:17most used softwares that I use when
56:20building automations. So, here I am on
56:21the dashboard. I can go down to
56:22settings. I can go to API integration
56:25and I can copy my key. And by the way,
56:27you get $5 of free credits when you
56:29apply or when you actually get into
56:31Apify. I have $29 because that's what I
56:33paid for, but you have the free version
56:35which is $5. And if you're anything like
56:36me and you want to rinse the software,
56:38just make five different emails and then
56:41go through them. So, for the LinkedIn
56:42icebreaker template, do you want to
56:43provide it now or should I build the
56:44system first and you plug in later? Let
56:46me give it to now. So, instead of doing
56:48LinkedIn icebreaker, what if we do this?
56:50Since we're going to get their their
56:51number as well, right? the way, you can
56:53filter by people who have a number like
56:55a phone number itself. I want you to
56:57find from the website so still do the
56:59exact same thing that I asked you to,
57:01but I want you to find from the website
57:02the value proposition of the business,
57:04what makes them unique, [music] and gave
57:06me a case study of a client that they've
57:09helped if you can find it. And this is
57:10part of testing working with cloud code
57:12and so on which is good. For scraping
57:13websites to personalize icebreakers, how
57:15deep should the scraping go? Homepage
57:17only, homepage and about, homepage and
57:18about services. I'll go homepage about
57:22services as well. Cuz I want to know
57:23pretty much everything about it. It'll
57:24be a bit more expensive, sure, but
57:26that's fine. I can submit answers and
57:27now we're going to go ahead. Now it will
57:30remake the plan because I told it to
57:31change a specific thing about uh the
57:34process. [music] So now it's saying that
57:35the input will be niche location number
57:36of results. The Apify Google Maps
57:38scraper to find businesses, filter keep
57:40only leads with both a website and a
57:41phone number. Cool. Scrape each website
57:43home, about, and services. Extract value
57:45proposition, what makes them unique, and
57:47case study of clients it helped if
57:48findable. And output will be a CSV file.
57:50Great, perfect. Okay, so now it's asking
57:51me as well for the Anthropic API key to
57:54analyze websites. So I can go to the
57:56console here, which is the
57:57anthropic.platform.cod.com.
58:00I can log in.
58:04I can go down to API keys. I can make a
58:06key, name it Claude
58:09uh Claude code.
58:11Copy the key.
58:12Paste it here. By the way, the reason
58:14why I'm pasting it into others is
58:15because if I press yes, I have one, I
58:17then have to wait a minute for it to
58:18understand that I told it yes, I have
58:20one, and it will ask me. So I just might
58:21as well just give it the API key itself.
58:23Okay, cool. The plan is ready. Here's a
58:25summary, and it has everything here. So
58:27as you can see, we have the lead
58:28generation system implementation plan,
58:30context, pep in overview, value to
58:31create, implementation details, step one
58:34Google Maps scraping, step two is
58:35filter, website scraping, Claude API
58:37analysis, CSV [music] output, error
58:40handling, dependencies, verification.
58:41Great, amazing. I can press yes, auto
58:44accept.
58:45And now I can go to bypass permissions,
58:47and it will start building it out step
58:49by step. As you can see now, it's
58:50starting to make different files. But
58:53it's not creating at scale, it's just
58:54creating the actual file itself because
58:56we didn't tell it yet to create the
58:57scale. It said it created the files.
58:58Great. Want me to run the test? Yes, run
59:00the test with roofing companies
59:05in
59:06I don't know why I'm very adamant on
59:07roofing companies in North Carolina.
59:09I've never been there. In North
59:10Carolina. And I want and I want 100
59:13results.
59:14Let's just go 50. Actually, 40 to make
59:16things [music] faster. Actually, let's
59:17do 30 cuz I'm not trying to wait all
59:18day. So, there we go. Run the test. Now
59:21it will start working on it. Okay, so
59:22now it's running the actual lead
59:23generation automation. If I go to Apify,
59:25I can see that it just scraped 34
59:27results, right? It just scraped it, and
59:29now it's going to use those results to
59:30then go and scrape each website of those
59:32results, obviously assuming that they
59:34have the website and phone number. All
59:36right, so it just finished running. We
59:37have the leads.csv document, which is
59:39this one right here. I can actually
59:42command A, copy. I can go to a new
59:44Google Sheet right here. I can go here.
59:46I can paste this, and then I can go to
59:48data, text to the columns. There we go.
59:50And now we can have
59:53different
59:54>> [music]
59:54>> things here. So, I can make this bold,
59:56little bit of formatting here and there.
59:57Make this white. Sorry, guys. Have to do
59:59this every single time. Otherwise, I
1:00:00[music] can't work with Google Sheets.
1:00:02And now we have the company name,
1:00:03location, the website, have the phone
1:00:06number, value proposition. So, they do
1:00:08comprehensive roofing gutter solutions
1:00:10with a family-owned approach that
1:00:11prioritizes customer satisfaction and
1:00:13personal attention. They offer financing
1:00:14options. Okay, cool. Unique
1:00:16differentiation. So, this company is led
1:00:17by John Davis with 30 years plus of
1:00:19industry experience. [music]
1:00:20They maintain an A+ BBB rating and don't
1:00:23consider a job complete until a customer
1:00:25is 100% satisfied. Great. Now, as you
1:00:27can see for case studies, interesting.
1:00:29Most of them didn't have case studies,
1:00:31which is fine. And so, this right here
1:00:32is just an extra thing that we can do.
1:00:33I'm going to leave it here, right? Cuz
1:00:34some of them do. So, we have a client
1:00:36testimonial from Tarun Mansingh details
1:00:38on how Brian completed multiple home
1:00:39improvement projects inside a building.
1:00:41And so, the reason why I asked this is
1:00:42because let's say we're cold calling the
1:00:44business. We can see something like
1:00:46this. We can commend the business on
1:00:48having a case study, and we can
1:00:49reference the name of the person who
1:00:50gave the case study, right? Which is
1:00:52always good. Uh cuz personalization over
1:00:54anything is great. And so, that over
1:00:55there is the list, right? That we have
1:00:56here, which is great. What I can do now
1:00:58is this. I can say, "Okay, great. Now,
1:01:00what I want you to do here is the whole
1:01:02process that I went you through. I want
1:01:04you to create a skill." All right, so
1:01:05now in terms of improvements, I want to
1:01:06do a few things. For example, location
1:01:09sometimes gets mixed up with company
1:01:10[music] name. Website sometimes get
1:01:12mixed up with the location. Phone
1:01:13number, and this is pretty much all
1:01:15shifted for no reason.
1:01:16>> [music]
1:01:16>> So, I can go here, and I can say, "There
1:01:18was one problem." Like, for example,
1:01:20Guardian Roofing of North Carolina, the
1:01:22location is LLC, the website and then
1:01:24the value proposition is a phone number
1:01:26and so on, right? And so we are shifting
1:01:28everything by one. So I want to make
1:01:30sure that when you give me the CSV, you
1:01:32check that all the columns that we have
1:01:34are clearly placed, right? There are a
1:01:36lot of them that don't have a location,
1:01:38which is fine. So we can leave it there
1:01:40and you put things in the right place.
1:01:41But there are some of them as well that
1:01:43have different things in different
1:01:44places. So I want you to fix that and
1:01:45then give it back to me. And this is
1:01:47part of iterating the actual skill
1:01:48itself. All right, so now I told it that
1:01:50it updated the leads. The CSV, I can go
1:01:52here, I can copy this, and go back. I
1:01:55can make a new sheet and I can paste it
1:01:56here. I can go to data.
1:01:59Split text to columns and as you can see
1:02:00here, now it fixed
1:02:01>> [music]
1:02:02>> the error that we had before,
1:02:03which is perfect. There we go. Now, with
1:02:06all this said, we can go to the next
1:02:08step and actually create this skill that
1:02:10we need. I'm in a new tab and I can say,
1:02:12"Okay, now that you generated the lead
1:02:14generation kind of workflow, I want you
1:02:16to create a skill so that when I call
1:02:18the skill,
1:02:20it basically runs through the whole
1:02:21process that we just went through. Okay?
1:02:22So let me know if you have any
1:02:23questions, but create a skill called
1:02:24lead generation system and then um add
1:02:27it in the in the folder itself. I can
1:02:29press go. I can put plan mode. But
1:02:31essentially now it will use the skill of
1:02:32skill creator to then be able to create
1:02:34the skill. Okay, so now I can press yes
1:02:36and then I can go from plan mode to
1:02:38bypass permissions. All right, so
1:02:39[music] now it's finished creating the
1:02:40skill. I can go here to lead generation
1:02:42system and I can see that I have a
1:02:43document called skill.md and this
1:02:45contains [music] the full instruction
1:02:46manual for that specific skill. So lead
1:02:48generation system, description,
1:02:49instructions, and we go on and on. The
1:02:52only thing they didn't do is that it
1:02:53didn't make a folder here called script
1:02:56and references
1:02:56>> [music]
1:02:56>> because this right here and this right
1:02:58here, they're both references, even
1:03:00this. And then this right here is a
1:03:01script. So what I can say here is,
1:03:03"Inside the skill folder for the lead
1:03:05generation system, can you also add a
1:03:07references and scripts folder to store
1:03:10the appropriate files? So the lead
1:03:12generation.py needs to be in a script
1:03:14folder
1:03:15um and then the rest as well, like based
1:03:16on what you know and obviously your
1:03:18definition of scripts and references,
1:03:20just include the files there to make
1:03:21things organized. Now I can press go. As
1:03:23you can see now, it just made two
1:03:24different folders, references and
1:03:25scripts. And inside scripts, we can see
1:03:27that we have the lead generation.py. And
1:03:29then inside references, we can see that
1:03:30they have env, requirements.txt, and
1:03:32sample output.csv, which is great. One
1:03:34thing I want to do though is I want to
1:03:35delete these ones. So, I can say, "Okay,
1:03:38now that you added those inside the
1:03:39references and scripts folders, can you
1:03:41delete the other ones that we're not
1:03:41going to use?" As you can see, it
1:03:42deleted both of them and it's also
1:03:44asking me, "Hey, do you also want to
1:03:46delete .env and leads.csv in the upper
1:03:48files?" Yes, [music] delete them. There
1:03:51we go. So, now we have a references
1:03:52folder, we have a scripts folder with
1:03:53the automation, in this case the Python
1:03:55file, which doesn't make [music] any
1:03:56sense to anyone here. It's just a file
1:03:59that runs the automation, all right?
1:04:00Because behind every automation in N10,
1:04:02make.com, it's all Python. It's also
1:04:05JSON. Because as you may know, all the
1:04:06platforms in N10 or make.com, behind the
1:04:09automations, it's all code. And so, this
1:04:11is the code itself that runs the actual
1:04:12[music] system. As you can see, we start
1:04:14to add more and more skills with more
1:04:15and more folders. Now, there's one more
1:04:16thing that I want to do and it's this,
1:04:18which has to do with commands. Now, I
1:04:20have a skill called lead generation
1:04:21system. I want to create a command so
1:04:24that I can put like {slash} lead
1:04:25generation and then it calls that
1:04:26specific skill automatically. Can you
1:04:28make that command? I can go back to
1:04:29permissions and I should be able to see
1:04:31a command being made here. All right, so
1:04:32now I finished making the command. As
1:04:33you can see, we have the lead
1:04:35generation.md here. So, if I go to a new
1:04:37chat, I can press {slash} lead
1:04:39generation
1:04:41and then that will activate the lead
1:04:43generation
1:04:44>> [music]
1:04:44>> system that we have in the skills
1:04:45itself. There we go. And now it's going
1:04:47to ask me, "Hey, what niche do I want?
1:04:48Location, number of leads, and output
1:04:51file?" I was going to say roofing
1:04:51companies again. Med spas in London
1:04:55and I want to find 30 and I want
1:05:00to name the file
1:05:02med spa
1:05:04London
1:05:0630.
1:05:07Boom.
1:05:08And I can go to bypass permissions.
1:05:10Okay? So, that right there is the full
1:05:13process of creating a skill step by step
1:05:15and actually testing it and making a
1:05:16command so that whenever I go to clock
1:05:18code, I can [music] simply just press
1:05:20slash the skill name and it will
1:05:22automatically call the skill itself. All
1:05:23right, so we can see here we have the
1:05:24result match by the London. I can make a
1:05:26new sheet data
1:05:29and then we have the full list right
1:05:30here.
1:05:31And we just call this simply by pressing
1:05:33slash command
1:05:34or that specific skill. So now you know
1:05:36exactly how to build your own clock with
1:05:37skills. But what if I told you that your
1:05:39clock with skills could actually improve
1:05:40themselves 24/7 overnight without you
1:05:43having to do anything. This is where we
1:05:44get to self-improving skills. Like we
1:05:46saw earlier, the first way of actually
1:05:47self-improving your own clock with
1:05:49skills is to do it manually. So we give
1:05:51it the first basic version, it made this
1:05:53right here. Then we said, "Okay, this
1:05:55right here is a bit messed up. This
1:05:56right here is a bit messed up as well.
1:05:58Can we change this?" And then it came
1:05:59and did this as well. And so what it did
1:06:01is that it made it better. It improved
1:06:02itself. And this is great because it
1:06:03allows us to make a near perfect score
1:06:05or the output for our specific skills.
1:06:07But if you want to take your skills to
1:06:08the next level and actually measure how
1:06:11improve by, well, this is where we get
1:06:12to the skill evaluation. So the first
1:06:14step of skill evaluation is making sure
1:06:17that you have the skill creator
1:06:18installed. So you need to have this
1:06:19installed. And this is something that we
1:06:21spoke about in one of the first steps.
1:06:22Then once we have the skill involved, we
1:06:24then have to tell Claude to evaluate our
1:06:25skill. And so here I can say, "Create an
1:06:27eval for my lead generation system
1:06:29skill." I'm just going to do the YouTube
1:06:31title generation just to keep it more
1:06:32simple. So for my YouTube YouTube title
1:06:35generation skill, and I can press go.
1:06:37Make sure it's bypass permissions. As
1:06:38you can see here, it created two eval
1:06:40files. One is eval.json and one is
1:06:42trigger eval.json. Basically, the whole
1:06:44point of evals is to make [music] our
1:06:46skills self-improving. And to make the
1:06:48skills itself improve themselves over
1:06:49time. The first thing it made is three
1:06:51test use cases. So if I go here to the
1:06:53skill itself, it made a folder called
1:06:55evals and we have two. So the first
1:06:57thing is test cases. So it uses specific
1:06:59inputs to run the skill. So let's say
1:07:01we're using the prompt of clock with
1:07:02beginner video. What it tests is a
1:07:03framework usage, character limit,
1:07:05curiosity varied. Then it uses VS code
1:07:06versus cursor versus Win some
1:07:08comparison. So basically these are three
1:07:09video ideas and this is what it tests
1:07:12for, right? In each eval has five to six
1:07:14expectations checking. Exact count 10
1:07:16titles character limit 60, framework
1:07:18coherence, topic relevance, curiosity
1:07:20engagement and variety. Whilst on this
1:07:21side we have 20 trigger tests. 10 should
1:07:23be triggered, 10 should not. Should
1:07:25trigger is give me YouTube title ideas,
1:07:27brainstorm video titles, help me name my
1:07:28video, etc. Should not trigger is
1:07:30writing scripts about YouTube video
1:07:31script, thumbnails SEO analysis, blog
1:07:33titles, channel growth advice,
1:07:35>> [music]
1:07:35>> adjacent but distant tasks. And so now
1:07:37what it will do is it will test the
1:07:39actual skill based on this criteria and
1:07:41it improves if it finds that this output
1:07:43isn't good enough. So it will test all
1:07:44the three use cases and it will test all
1:07:46these 20 triggers. With this said, I can
1:07:48now say run the eval for my YouTube
1:07:49title generation skill [music] and I can
1:07:51go press run. What this will do now is
1:07:53it will go and run these two and then
1:07:55give us a kind of a report. [music] All
1:07:57right, perfect. Now it's finished with
1:07:58the evaluation. It's telling us that
1:07:59with the skill it did 100% correct,
1:08:01without the skill 33%. With the skill
1:08:03100% and then [music] average will be
1:08:0550% here and average will be 100% here.
1:08:07And it also give us full document here
1:08:09with the prompt with the skill, the
1:08:12outputs that we got, formal grades, so
1:08:14six out of six passed, zero failed,
1:08:16which is perfect. Then we also have the
1:08:17benchmark comparison here, which gives
1:08:19us all the data of what the output looks
1:08:22like with the skill and what the output
1:08:24looks like without the skill, okay? And
1:08:26so here we have the pass rate with the
1:08:28skill, without the skill. Obviously here
1:08:29we have with the skill 100%, without the
1:08:31skill 50% plus minus 17. So as you can
1:08:33see, with the skill it's a perfect score
1:08:35of 100. And so what it did here is that
1:08:37it evaluated one, two, and three so that
1:08:39we have the clock code beginner, which
1:08:41is the first topic that it used, the
1:08:42first task use case. With the skill
1:08:44without the skill, you can see all the
1:08:45numbers here. Assertion, so all the
1:08:46different things that we needed inside
1:08:48that specific uh title. With the skill I
1:08:50got everything right, without the skill
1:08:51only two. And it did that for all the
1:08:53three topics. So topic number one, topic
1:08:55number two, and topic number three. And
1:08:57then we have the analysis notes. [music]
1:08:58So with the skill achieves 100% pass
1:09:01rate across all three evals versus a 50%
1:09:03average without the scale. The biggest
1:09:04scale advantage is in the former
1:09:06compliance, exact count, 10 titles, and
1:09:08character limit baseline consistency
1:09:09over produce and exceeds length limits.
1:09:11And so all this is doing is it is
1:09:13testing three different use cases for
1:09:14our scale and it's telling us, "Hey, out
1:09:15of 10 times that we run this, how many
1:09:17times fail? How close is it to the
1:09:19desired output that we asked it to do?"
1:09:20And so you can always have this
1:09:21evaluation for every single skill that
1:09:23you do, you'll be able to test it. Now
1:09:24of course, YouTube title generation is a
1:09:26very easy system, but if we were to do
1:09:28something with the lead generation
1:09:29system, it would work really well just
1:09:30because there's a lot of moving
1:09:31components in the actual skill itself.
1:09:33And so at this point, you can tell it to
1:09:34improve the skill, run the eval again,
1:09:36and see the numbers and how it's
1:09:37improved. And so at this point now, we
1:09:39have 100% accuracy, right? But let's say
1:09:40we had 80%. So what we would do in this
1:09:42case, if in case it has 80% or anywhere
1:09:45less than 100, I can tell it, "Hey, it's
1:09:46missing X, Y, and Z. Can you improve the
1:09:48skill?" It then improves the skill, it
1:09:49tells us, "Hey, it's done." And then we
1:09:51ask it to give us the evaluation again.
1:09:52And this is a whole repetitive loop that
1:09:53it does to make sure that these skills
1:09:56are self-improving. One more thing I
1:09:57wanted to go through when it comes to
1:09:58skills is the ability to call them using
1:10:01{slash} skill, right? So in this case,
1:10:03in the dot claw folder, we have commands
1:10:05and we have skills. I've gone through
1:10:07skills, but commands are the things that
1:10:09we use to call that specific skill,
1:10:11right? So for example, let's say I
1:10:12wanted to run YouTube title generation.
1:10:15Instead of saying, "Hey, can you run
1:10:18YouTube title generation?" I would say,
1:10:20{slash} YouTube title generation and it
1:10:22will simply run automatically, right?
1:10:24Just because we've added the command
1:10:25which allows us to call it directly.
1:10:27Same way that if I wanted to run the
1:10:29views to clicks, I would go {slash}
1:10:31views to clicks analysis and it will run
1:10:33the skill itself. And this is great for
1:10:35us because we simply don't have to
1:10:37manually write the skill itself, you can
1:10:39just put {slash}, call the skill, and
1:10:41you're good to go. And if I go to
1:10:42{slash}, I can see down to the commands,
1:10:46I can go down to the skills right here,
1:10:48and I can see all my skills that I can
1:10:49call automatically and it will simply
1:10:52trigger that specific automation. And
1:10:53the way that we did that is, let's say
1:10:55we have Excalidraw diagram and we have
1:10:57YouTube title generation. If I wanted to
1:10:59call this one here, I would have to put
1:11:01/excalidraw diagram and I simply don't
1:11:03have the command there. So, what needs
1:11:04to happen because of that is I would
1:11:06have to explicitly say, "Hey, can you
1:11:10run my skill called the Excalidraw
1:11:14diagram, right?" Which is quite
1:11:15inefficient, especially when you have 20
1:11:17to 30 skills that you wanted to browse
1:11:19through and just pick one automatically,
1:11:21right? And so, in this case, to turn
1:11:22that skill into a command, I can say,
1:11:25"Hey, can you turn my two skills into
1:11:27two commands and then so that I can call
1:11:29them using / the name of the skill and
1:11:31then add the file inside the commands
1:11:35folder under the doc load folder." All
1:11:37right, now I can press plan mode cuz I
1:11:39wanted to make a plan first. All right,
1:11:40so here we have the plan itself, turn
1:11:42skills into / commands. We have two
1:11:44skills and we're going to create the
1:11:46commands as well. I can just press yes
1:11:48and auto accept. We can see here that
1:11:49it's going to create a file called under
1:11:52the commands, under here. Same thing
1:11:53with this, YouTube title generation and
1:11:56Excalidraw diagram. I can press yes auto
1:11:58accept. I can go to bypass permissions
1:12:01and I can let it do its thing right
1:12:02here. All right, so after a few minutes,
1:12:04it made the two different commands,
1:12:05which are two different files. So, we
1:12:06have this one here, which says, "When
1:12:08someone calls /excalidraw diagram, use
1:12:11the Excalidraw diagram skill to handle
1:12:13the request." Same thing with the
1:12:14YouTube title generation. So, when
1:12:15someone calls /youtube title generation,
1:12:18use the skill right here to handle the
1:12:20request. That way, if I go here and I
1:12:22can put /youtube title generation, it
1:12:24will run the skill itself right here. As
1:12:26you can see, it's looking at all the
1:12:27different references. Then it should ask
1:12:29me what video do I want to make a title
1:12:31for. As you can see, now it's asking me
1:12:33for the video topic, angle, number of
1:12:35titles. And then if I go to a new
1:12:36conversation, I can do the exact same
1:12:38with Excalidraw. So, I can put
1:12:40/excalidraw diagram and it will start
1:12:42running the Excalidraw skill. And now
1:12:45it's going to ask me what concept or
1:12:46system do I want visualized and what
1:12:48level of detail. And the next level
1:12:49would be, obviously, having 10 to 15
1:12:51different skills with 10 to 15 different
1:12:53commands, so that we're able to call the
1:12:56skills directly without us having to
1:12:58type the skills off or have a
1:13:00conversation with Clockwork. We can
1:13:01simply just filter by the skills. You
1:13:03can go here and we can look at all the
1:13:05skills that we have and call them
1:13:06directly right here.
1:13:09All right. So, now you understand how to
1:13:11build custom skills and slash commands
1:13:14that make Clockwork do what you need in
1:13:16one line. But what if sometimes you want
1:13:18Clockwork to do things automatically
1:13:20without you even having to ask? Well,
1:13:23that's exactly what hooks are for and
1:13:24that's what we're covering in the next
1:13:25video. Let's dive in. All right. Now,
1:13:27the one feature that I'm talking about
1:13:28is, of course, Clockwork hooks. If you
1:13:30never heard of Clockwork hooks, they are
1:13:32a bit more on the technical side, but
1:13:34they're actually very, very easy and
1:13:36very powerful when you understand how to
1:13:37use them. So, what they are is automatic
1:13:39actions that run at specific moments
1:13:42like every time X happens, automatically
1:13:44do Y. And it's almost like setting rules
1:13:46that your assistant always follows. Now,
1:13:48the only difference between Clockwork
1:13:49hooks and Clockwork skills
1:13:51>> [music]
1:13:51>> is that hooks are things that Clockwork
1:13:55uses regardless of whether you call them
1:13:56or not. They're on the organization
1:13:58level, right? So, you set a rule that
1:14:00every single time Clockwork runs, it
1:14:02always uses these hooks, okay? And
1:14:04they're automatic things that I'll walk
1:14:06you through what they what they actually
1:14:07are. For example, let's say you give
1:14:08Clockwork a task. You ask Clockwork,
1:14:10Clockwork does the work and it uses
1:14:12these hooks, right? Whenever it does the
1:14:15work. For example, we have pre-tool use.
1:14:17So, we have hooks that we have to use
1:14:20before the tool runs. Post-tool use,
1:14:22which are hooks that we have to use
1:14:23after the code is written. Stop, when
1:14:25Clockwork finishes. Or notification,
1:14:27when Clockwork needs you. So, these are
1:14:28the four different types of hooks that
1:14:29you can have and they're just examples
1:14:31of what it could look like, right? So,
1:14:33you have a hook that runs every single
1:14:35time the workflow finishes or Clockwork
1:14:37starts or when it stops or when it needs
1:14:40notifications. And they're almost like
1:14:41triggers. What starts the actual hook
1:14:43itself. Now, I'm here on the YouTube
1:14:45workspace and if I go to my doc cloth
1:14:47folder, which is where the uh hooks are
1:14:49going to be stored, there's a file
1:14:50called settings.json. If I press on
1:14:53settings.json, you'll be able to see the
1:14:55full code. [music]
1:14:56Now, the code itself is actually very,
1:14:58very easy to understand, right? Now,
1:15:00don't worry about the code too much cuz
1:15:01I'll explain exactly what what it is, uh
1:15:03but essentially, it's stored in a file
1:15:05called settings.json. And so, here I am
1:15:07on a new project. I can go to
1:15:09new folder here. I can name it doccloud.
1:15:12And by the way, doccloud is also the
1:15:13folder that we use to store skills, to
1:15:16store commands, to store different
1:15:17things of cloud code. And inside
1:15:19doccloud, we want to add a new file
1:15:21called settings.json.
1:15:24Now, a few things. First of all, that
1:15:27whenever we're making a hook, a hook is
1:15:29sort of like automation or some sort of
1:15:30plugin, we are all making it under the
1:15:33same exact file. So, this is not like a
1:15:35skill where you would need to add a new
1:15:37uh file for every single hook. We use
1:15:39settings.json as the main file to store
1:15:42all the different hooks that we have
1:15:43across our organization. Okay? Second
1:15:46thing is that when we are uh installing
1:15:49[music] hooks in settings.json, you will
1:15:51be able to use this hook across
1:15:53different projects. This will make more
1:15:54sense when I show you, but essentially,
1:15:56understand those two rules. First thing,
1:15:57you'll be able to write all the hooks
1:15:58here. And then second thing, just as if
1:16:00you were able to to write all the skills
1:16:02in one file without you having to make
1:16:04different folders, but in this case,
1:16:05skills are different. And then, you need
1:16:07to understand that this is on a
1:16:09organization level, not on a project
1:16:10level. With this said, let's go to here,
1:16:13and the way to actually use hooks is you
1:16:15can either press {slash} hooks,
1:16:17which you will be able to only to do
1:16:19using a terminal, as you can see
1:16:22right here.
1:16:23Terminal.
1:16:24Let me go here to terminal, new
1:16:26terminal, [music]
1:16:27and I can press cloud. So, here we are
1:16:28on a terminal view using cloud, and I
1:16:30can put {slash} hooks.
1:16:31View hook configuration for tool events.
1:16:33I can go enter,
1:16:35and here we have zero hooks configured.
1:16:37>> [music]
1:16:37>> So, we have zero hooks that are
1:16:38configured right now in our
1:16:39organization. Okay?
1:16:41And here, we have a read-only menu to
1:16:44add or modify hooks, okay? And so, in
1:16:47the easiest way, what this is is you
1:16:49being able to say, "Okay, I want to run
1:16:52some sort of notification before the
1:16:54tool execution." And so, these are all
1:16:56triggers, right? When I say trigger, I
1:16:57mean, this is a trigger before the tool
1:16:59execution. This is a trigger after the
1:17:01tool executes. This is a trigger post
1:17:03tool use failure. So, when the tool
1:17:05fails. So, let's say you want to send
1:17:06yourself an email whenever a tool fails,
1:17:09you use that. Permission denied. After
1:17:11Automode Classifier denies a tool call,
1:17:13then you use the hook. Then you can make
1:17:15a hook there. And so, you have different
1:17:17options that you can work with whenever
1:17:19you have these kind of statuses [music]
1:17:22inside of Flow code. Let's say I go to
1:17:24notification.
1:17:26Press enter.
1:17:27It says, "We currently have no hooks
1:17:29configured. To add a hook, you either
1:17:31need to edit
1:17:32settings.json directly, or you need to
1:17:34ask Claude." And now, let's go to the
1:17:36normal view,
1:17:37and let's talk to Claude code. Say,
1:17:39"Okay.
1:17:40Hey, I want to make a hook."
1:17:43So, here it's asking me, "What kind of
1:17:44hook do you want to do?" Here are some
1:17:46common hook events. And I have running
1:17:48before the tool is called, run [music]
1:17:49after the tool completes, run when
1:17:51Claude sends a notification, run when
1:17:53Claude finishes a response, or run when
1:17:54sub-agents finishes. Let's do this.
1:17:56Let's run a hook where when Claude
1:17:57finishes a response, it sends me a
1:17:59notification.
1:18:01Okay, so that's going to be the hook
1:18:02that we're going to make. So, when
1:18:03Claude finishes doing something, it
1:18:05actually gives me a notification. I can
1:18:07press yes. As you can see now, it's
1:18:09running [music] the hook. And by the
1:18:09way, when you see something like this,
1:18:11always in Claude code, it just means
1:18:13that this is the previous version, this
1:18:14is the updated version. And it's saying,
1:18:17"Do you agree that we can change this
1:18:19exact file?" And I can say, [music]
1:18:21"Yes." And now, it's going to add the
1:18:22hook inside of settings.json here. And
1:18:24it says that the hook is live. Every
1:18:26time I finish a response, you'll get a
1:18:27Mac OS notification saying that Claude
1:18:30has finished responding with the title
1:18:31Claude code. And I can review or edit
1:18:33anytime via hooks. So, in theory now,
1:18:35what we made is a hook that alerts me
1:18:37every single time ClawKnit finishes
1:18:39doing something so that I'm able to then
1:18:41do whatever I want, work on different
1:18:43things simultaneously, and I'll know
1:18:44when I have to go back to ClawKnit to
1:18:46see the output. So here I can say,
1:18:48"What's the capital of Italy?" And I'll
1:18:49get a notification when it finishes
1:18:51running. Now this is the exact
1:18:52notification that I just got. The reason
1:18:54why I'm showing you here is because I'm
1:18:56using my my monitor so it doesn't
1:18:57actually show you the notification on
1:18:59the monitor as well. But essentially
1:19:00what would happen is that you would be
1:19:02talking to ClawKnit at the same time you
1:19:05can be doing something else and you will
1:19:06know whenever you have to go back to
1:19:08ClawKnit because of this notification
1:19:10right here. And there are there is a
1:19:11hook, right? And this will run every
1:19:14single time that you talk to ClawKnit
1:19:15regardless of whether you have all these
1:19:16documents in your project, it will
1:19:18always [music] run. Right, these are
1:19:19rules that will always run regardless of
1:19:21X, Y, and Z. So in this case, when
1:19:23>> [music]
1:19:23>> ClawKnit finishes, then we get a
1:19:25notification. All right, so another
1:19:26example here could be that let's say we
1:19:29had a .env document. So this .env
1:19:31document is [music] full of API keys for
1:19:33OpenAI, Anthropic, Stripe, all these
1:19:36things. And by the way, they're fake so
1:19:37don't try to copy them. Fake API keys.
1:19:39Let me just put API keys. None of these
1:19:41are real. And what I want is that
1:19:44whenever I talk to ClawKnit, maybe I
1:19:45could be doing a complete custom code
1:19:47full stack project, or maybe I'm doing a
1:19:49full website, or maybe I'm building some
1:19:51some automations, right? Ideally we
1:19:52wouldn't want ClawKnit to impact the
1:19:54.env document because once this is
1:19:56impacted, let's say it deletes
1:19:58everything by mistake, we have to
1:20:00reconfigure everything and all the
1:20:01automations break because they're all
1:20:02reliant on these passwords right to
1:20:05access the softwares for the
1:20:06automations. So what we can do here is I
1:20:08can say, "I want to make a hook."
1:20:12So in this case, we want
1:20:14>> [music]
1:20:14>> to tell ClawKnit, "Hey, don't edit the
1:20:17.env at all, right?" And so in this case
1:20:19it would be a pre-tool hook. So we would
1:20:21say, "Yes, so this is a pre-tool hook."
1:20:23Essentially what we want is [music]
1:20:25I want to tell ClawKnit to not touch the
1:20:28.env document. So don't ever touch the
1:20:30.env document regardless of what it is
1:20:32that we're doing. So, if you have to
1:20:33edit the .env, then you can't do it
1:20:35because you have to stop yourself cuz
1:20:36it's sensitive information. So, now
1:20:38we're making another hook. I can press
1:20:39yes. Let me go to bypass permissions.
1:20:41So, I can let it create the own hook.
1:20:44There we go. So, now it added the
1:20:46pre-tool use onto our settings.json. So,
1:20:49if I go here, I can see now that I have
1:20:50the stop hook, which means that whenever
1:20:52the Claude code stops, it gives us a
1:20:54notification. So, we get a notification
1:20:56here when Claude has finished
1:20:57responding. And then, we also have the
1:20:59write edit one, which is the one that
1:21:03allows us to block .env. [music] And
1:21:05then, we have the pre-tool use hook,
1:21:07which is the one that we used to be able
1:21:10to not let Claude code touch our .env
1:21:12document because, of course, it has all
1:21:14the passwords. So, now here I can say,
1:21:16"Create a .env file with API key
1:21:19test123." [music] And you can see that
1:21:21it says, "I can't do that. Your
1:21:22settings.json has a hook that blocks me
1:21:24from writing to the .env files since
1:21:26they contain sensitive information. You
1:21:27can create it yourself by running echo
1:21:29this [music] this this." right? And so,
1:21:30this is extremely important, right? I
1:21:32wish I knew this sooner because there
1:21:33are many many times where I told Claude
1:21:35code to do something and it like [music]
1:21:37deleted something from here or it added
1:21:39something from here, which messed up my
1:21:41whole project. And so, in this case,
1:21:43what we're saying is, "Whatever happens,
1:21:45don't edit this," right? What I would
1:21:47recommend is you add all the API keys
1:21:49that you have, all the passwords that
1:21:51you have that Claude code needs to
1:21:52access to whenever it needs to run a
1:21:54specific automation. And then, once you
1:21:56have those, put this hook here, make
1:21:58sure that now Claude code is not able to
1:22:00edit .env to be able to make sure that
1:22:02you're safe whenever you're building the
1:22:04project. And last example, let's say I
1:22:06am building a product for a client. At
1:22:07the end of every session, I want to know
1:22:09exactly what files Claude code touched
1:22:11and when, which is an audit trail. And
1:22:13so, what I want is a log of all the kind
1:22:15of edits of different files that it made
1:22:17with the timestamps of when it actually
1:22:18touched them. So, here I can say, "Add a
1:22:20post-tool use hook because post-tool is
1:22:22going to be the trigger." After Claude
1:22:24code finishes, it logs every file Claude
1:22:26code edits or writes to a file Claude
1:22:28Claude/log.txt
1:22:30with a timestamp. And we can press go.
1:22:32And this is great because a lot of the
1:22:33times that I'm building with Claude
1:22:35code, I am making stuff, and then I'm
1:22:37forgetting which files I edited. And
1:22:39sometimes things just get deleted. And
1:22:41so this allows me to know exactly what
1:22:44got edited and what got deleted or what
1:22:46timestamp. All right, so it just
1:22:47finished running. It said the post tool
1:22:49use hook is not configured. I can see
1:22:51here that I have the right to edit type
1:22:53command, and I have the status login
1:22:56change to. All of this makes no [music]
1:22:58sense to me. It should log everything to
1:23:00a Claude/log.txt
1:23:02document, okay? So now if I go here, I
1:23:04say, "Can you make a [music] new
1:23:05markdown format document and name it
1:23:08YouTube title generation?"
1:23:10As you can see now,
1:23:12it's getting the task, but it's also
1:23:14updating the Claude/log.txt.
1:23:16If I go here, you can see that 17:44:09,
1:23:20we wrote
1:23:21YouTube channel generation.md. Let's say
1:23:23I wanted to say,
1:23:24>> [music]
1:23:24>> "Can you also make another document or
1:23:25another five documents naming them uh
1:23:28hook number one, hook number two, three,
1:23:30four, five?" So essentially it's
1:23:31whatever change you make, it will log it
1:23:33here. So as you can see, it made five
1:23:34different documents. If I go here, I can
1:23:36see that we now have five different
1:23:38logs. So 1 2 3 4 5. It'll be hook one,
1:23:42hook two, hook three, hook four, hook
1:23:43five. That way you can have a log of all
1:23:45the edits that you've made, so you can
1:23:46always come back to them in case it went
1:23:48wrong in some way.
1:23:52All right, so now Claude code can
1:23:53actually run custom commands and run
1:23:55actions on its own. But there's still
1:23:56one limitation. And that is that it can
1:23:59only work with what's inside your
1:24:01machine. What if you want Claude code to
1:24:03talk to your browser, talk to your
1:24:04databases, talk to your APIs, or any
1:24:07external tools? Well, as of right now,
1:24:09it can't really do that without this one
1:24:11thing, which is called model context
1:24:13protocol. And it stands for MCP, and it
1:24:15completely changes what's possible. Let
1:24:16me show you. All right, so the best way
1:24:18to explain MCP is they're just plugins,
1:24:21right? Essentially other people build
1:24:23tools and you just plug them in. So, you
1:24:25have Claude Code, which talks to the MCP
1:24:27server, which is running in the
1:24:28background, and the MCP server will be
1:24:30the one controlling the external apps.
1:24:32So, your browser, your databases, maybe
1:24:34it's Notion, maybe it's Superbase, maybe
1:24:36it's your Gmail, right? Whatever that
1:24:38may be. And it makes it so much faster
1:24:40for us to be able to use the softwares
1:24:41because it is trained on how it should
1:24:43talk to the API documentation. And the
1:24:45API documentation is essentially the
1:24:47documentation which shows how you can
1:24:49actually automate some thing with a
1:24:50software. So, you can install the MCPs,
1:24:52they run as servers, right? And Claude
1:24:54gets new tools and talks to real apps.
1:24:56So, you basically don't have to
1:24:57configure things yourself. And so, how
1:24:58are they different from skills and
1:25:00hooks? Well, skills are just
1:25:01instructions. We say, "Do this task like
1:25:04this." Right? We have step-by-step
1:25:05instructions that Claude could uses to
1:25:07fulfill a certain task. Now, I made a
1:25:08full video on skills, you can check it
1:25:09up here. Then, we have hooks, which are
1:25:11guardrails. So, this is giving it rules
1:25:13as to saying, "Don't touch this file."
1:25:15It's sort of like the security guard,
1:25:16right? Which says, "Hey, before this
1:25:18happens, make sure this is in place.
1:25:20After this happens, make sure that is in
1:25:21place." Different rules. I've also made
1:25:23a full video on hooks, you can check it
1:25:25up here. And then we have the MCP. So,
1:25:27the MCP gives us new abilities, right?
1:25:29So, we say, "Now, you can use curl. Now,
1:25:30you're able to connect to Notion, and
1:25:32you know exactly what you should do."
1:25:33Right? And this is a new pair of hands
1:25:35that we give to Claude Code, which is
1:25:36why I say that it completely changed the
1:25:38way that I use Claude Code because it
1:25:39allows me to build things that
1:25:41previously I'd taken a lot more time to
1:25:43set up and things that previously were
1:25:45not possible. So, again, skills are
1:25:46instructions, hooks are guardrails, and
1:25:48MCP are new abilities. They work
1:25:50together, not against each other. So,
1:25:51they're different things, right? When
1:25:52you combine all these three things
1:25:54together, that's when Claude Code starts
1:25:56to become really, really, really
1:25:57powerful. And so, here I'm on my main
1:25:58YouTube workspace, which controls my
1:26:00main YouTube content uh generation
1:26:02system. And inside this workspace, you
1:26:03can see that I have a file called dot
1:26:06mcp.json. So, dot mcp.json is going to
1:26:09be the file that we use to add the MCPs.
1:26:11As you can see here, if I zoom out a
1:26:13little bit, here we have an MCP server
1:26:15called Chrome DevTools. And so, what
1:26:17does this do? This allows us to be able
1:26:19to use the web and access different
1:26:21things outside of Clockwork, right? This
1:26:23doesn't mean script a website, this
1:26:24means going inside a website and taking
1:26:26a screenshot, for example. Now, if you
1:26:28look at this, you're like, "What the
1:26:29hell am I looking at?" That's okay,
1:26:31because you don't really have to
1:26:32understand the code. You just have to be
1:26:33able to talk to Clockwork to ask it what
1:26:35you want, right? As you saw, to set up
1:26:37the MCP, we have to make a file. But
1:26:39before we do that, let me show you why
1:26:40we need it in the first place. Hey, can
1:26:42you go to apple.com? Can you take a
1:26:44screenshot of the main page, and can you
1:26:47then give it to me here? So, in theory,
1:26:48this should be able to then go to Apple,
1:26:50take a screenshot of the page, and then
1:26:52bring it back here as an image. As you
1:26:53see though, it says, "I can't take
1:26:54screenshots of website. I don't have a
1:26:56browser automation or screen capture
1:26:58tools available. I can only fetch the
1:26:59page content, read screenshot files that
1:27:01already exist on your machine." So, this
1:27:03is why we use MCP. And so, what I can do
1:27:05here is I can go to file, I can write
1:27:08.mcp.json.
1:27:10And inside here, I'm going to paste the
1:27:12code that we had before. File, and save.
1:27:14And by the way, I'll show you how to get
1:27:15this code as well, [music] so don't
1:27:17worry. Once we have this here, let me
1:27:18ask it the exact same question in a new
1:27:20tab, and you'll see how now it's going
1:27:22to go to Apple, take a screenshot, and
1:27:24then bring it back. So, you can see here
1:27:25it's saying, "Let me navigate to
1:27:26apple.com, and take a screenshot for
1:27:28you." I can say yes,
1:27:30and let me do bypass permissions, so it
1:27:31doesn't ask me. It said it successfully
1:27:33navigated to apple.com, it took a
1:27:35screenshot, and now we have the
1:27:36screenshot of the page, which is this
1:27:37one here.
1:27:38Okay? And so, just in a matter of
1:27:39seconds, we were able to unlock a new
1:27:41feature in Clockwork code just by adding
1:27:43this mcp.json file here. You can start
1:27:45to see the different applications that
1:27:47come into this. This is a very, very
1:27:48simple example. One of the most used
1:27:50examples is Chrome DevTools, which is
1:27:52the ability to go to websites, take
1:27:53screenshots, and access the web, but we
1:27:55also have other ones as well. And by the
1:27:56way, just so you know, it actually
1:27:58opened up a new Chrome tab here, which
1:28:00is how I was able to take the
1:28:00screenshot, and then bring it back to
1:28:02the workspace itself. All right, so to
1:28:04find the snippets of code that I just
1:28:05showed you, which allow Clockwork to
1:28:07have new features, plugins, we have to
1:28:09go to the mcp.servers.org. Now, if
1:28:11you're aware of skills, which is the
1:28:12ability for us to add instructions to
1:28:15clock code and do different things,
1:28:16skills also have something similar. So,
1:28:18if you go to skills
1:28:20.smp.com, we have what we call a
1:28:22workspace full of skills that we can
1:28:24copy and paste into clock code. So, this
1:28:25right here is Amazon for mcservers. And
1:28:28so, as you can see, we have the Chrome
1:28:29DevTool MCP, which is a an option that
1:28:31we can use. We have Firecrawl, we have
1:28:33Superbase, we have Playwright, we have
1:28:35Xcode Build MCP, and we have a lot more
1:28:37as well. Right across official, search,
1:28:39web scraping, communication,
1:28:41productivity, and so on, right? So,
1:28:43let's say I go to featured. Let's say we
1:28:45go to the Chrome DevTool that we had
1:28:46before. If I go here, I'll be able to
1:28:48see the name, the uh description as
1:28:50well. So, this is the tool that allows
1:28:53your coding assistant, [music] such as
1:28:54Gemini, Cloud, Cursor, or Copilot,
1:28:56control and inspect a live Chrome
1:28:58browser. See, it allows to control
1:29:00Chrome and go to a new browser and do
1:29:01whatever it is that you want to do. It
1:29:03looks at the key features, the
1:29:04disclaimers, usage statistics, update
1:29:06checks, requirements, and we have the
1:29:09code itself. So, here, we can copy this
1:29:11code, and that's the code that you'll be
1:29:13able to paste right here. Go here, you
1:29:15can paste it, and then make sure you
1:29:18press command S. Then it pushes
1:29:19everything back. That's why I kind of
1:29:20changed the structure. But, it's the
1:29:22exact same code. Now, bear in mind, with
1:29:23MCPs, not all of them are vetted, which
1:29:25means that they weren't completely
1:29:26verified. So, use your judgment as to
1:29:29what it is that you can actually be
1:29:30using. If I see that this is in the
1:29:31featured MCPs, then it's probably
1:29:33something that I am going to be able to
1:29:35use because it's trusted. But, let's say
1:29:37I go to cloud service, right, and find a
1:29:39random one, let's say AKS MCP. I don't
1:29:42really know whether this is good unless
1:29:44I look at the stars here, the GitHub
1:29:46stars. The higher the stars, the better
1:29:47it is. If it has two stars, then I'm
1:29:49most probably not going to be able to
1:29:50use this cuz I don't know the impact it
1:29:52will have on my clock code and the
1:29:54things that it does. And if you ever
1:29:55copy some code and you have no clue what
1:29:57it actually means, you can simply just
1:29:58say, "What is the MCP.json that I have
1:30:00right now? What does it do?" It tells
1:30:02you, and then I also told it to give it
1:30:03to me in one sentence and make it very,
1:30:05very simple to understand. And here it
1:30:06says that it lets me control your Chrome
1:30:07browser, clicking, typing, taking
1:30:09screenshots, and inspecting things so I
1:30:11can help you test and debug websites
1:30:13directly from the chat. Now, one use
1:30:14case for this specifically is let's say
1:30:16you are making a landing page. And let's
1:30:18say you wanted to look at the top
1:30:19landing pages from five different
1:30:21websites. And you wanted to take
1:30:22inspiration as to the top header that
1:30:25was on the landing page specifically. So
1:30:26here we have Apollo, we have Mailshake,
1:30:29we have Email List Verify, which are all
1:30:31different SaaS softwares that verify
1:30:33different emails or give you some data.
1:30:35Right? And let's say I wanted to make a
1:30:36new tool and make a landing page similar
1:30:38to these ones. And so here then I can
1:30:40say, "Can you go to each of these
1:30:41websites? Can you take a screenshot of
1:30:43the landing page? And then bring it back
1:30:45and make a new folder under our
1:30:46workspace, name it screenshots of
1:30:48landing pages, and leave the pictures
1:30:50there." And I can press go. I can put
1:30:51this as bypass permissions. Okay, so now
1:30:53it's going to site apollo.io, it's going
1:30:55to mailshake, email list verify, the
1:30:58next one is zerobounce, then we have
1:30:59verify-email.org, and then finally we
1:31:01have hunter.io. As you can see, it just
1:31:03finished everything. It made of
1:31:05screenshots of landing pages, and here
1:31:07we have Apollo, Email List Verify,
1:31:09Hunter, Mailshake, Scrap, Verify-Email,
1:31:12and Zerobounce. Right? And so all of
1:31:14these could be used then when you're
1:31:15making your landing page as inspiration.
1:31:17You can say, "Hey, look at these images
1:31:19of landing pages, kind of emulate their
1:31:21style, and make me a new landing page
1:31:23and call it XYZ name." So another one
1:31:25that I use often is the one of
1:31:27Superbase. So Superbase is a software
1:31:29that allows us to be able to store data.
1:31:31It's sort of like a Google Sheets on
1:31:32steroids, right? It allows us to add
1:31:34data to a table that we can then fetch
1:31:36using its API. So you're throwing a
1:31:37question, all the information is stored
1:31:39in Superbase, and Superbase gives you
1:31:40the answer. So here, what I can do is I
1:31:42can just configure this here. So I can
1:31:44copy this, which again usually is in the
1:31:46bottom, right? Here what I can see is
1:31:48the name.
1:31:49We can see the demo of this of how it
1:31:51works. Then it's giving me the actual
1:31:53setup.
1:31:54What I would have to do here is I would
1:31:56have to go and copy this. You can also
1:31:58look at the setup documentation here,
1:32:00which goes through everything that you
1:32:01have to do, but it's pretty much the
1:32:02same as the one we had here. So, I can
1:32:03go here and by the way, you don't
1:32:05actually have to make a new file, you
1:32:07can use the exact same one. So, now we
1:32:08have this
1:32:09and then I can say, "Can you add this to
1:32:11the MCP?" I can give it the new MCP of
1:32:13Superbase. When I press go,
1:32:15put bypass permissions, it will now add
1:32:18the MCP
1:32:19to our new MCP.json, which will be here.
1:32:22As you can see, it's making the change.
1:32:23I can say yes, and now we have the
1:32:25Chrome DevTools, we have Superbase, and
1:32:27we have more more more more as we start
1:32:29to add new things to our file. Okay,
1:32:31let's say I wanted to have this document
1:32:32right here, which is a fake SOP for an
1:32:34e-commerce store. I wanted to add it to
1:32:36my Superbase, so that we can store all
1:32:38of this in a table and then we can then
1:32:40fetch and can ask questions about the
1:32:42thing itself, like a chatbot. I can go
1:32:44here, I can download as a PDF, and I can
1:32:46upload it right here.
1:32:48And then I can say, "Hey, I uploaded the
1:32:50PDF of the e-commerce store SOP and
1:32:52knowledge base. Can you please now make
1:32:54a vector database inside of Superbase
1:32:56and can you add all the information
1:32:58there, vectorize it, and then uh yeah,
1:33:00make a new table and call it SOP." I'm
1:33:02going to make this plan mode, so plans.
1:33:04What it's now going to do is it's then
1:33:06going to use the MCP of Superbase
1:33:08because it has to access the software
1:33:09itself. As you can see, it's reading the
1:33:11MCP.json file because that's where we
1:33:12have the actual MCP itself. Now, it's
1:33:15giving me a plan for creating the local
1:33:17vector database in Superbase.
1:33:19Sure.
1:33:20Looks all good. I understand half of it,
1:33:22but
1:33:23not even half of it, but it's saying all
1:33:25the things I wanted to do. I can press
1:33:26yes, auto accept, and then I can put
1:33:28bypass permissions to let it run on its
1:33:30own. So, it's going to take the PDF, as
1:33:32we call it chunk it, so it's going to
1:33:33chunk, meaning that it's going to take
1:33:34the actual document, it's going to split
1:33:36it up into different parts, and then put
1:33:38it all into the uh vector database
1:33:40itself. And the vector database, by the
1:33:41way, is just a a Google Sheet which
1:33:44stores all the information. So, it just
1:33:45pulled up this page right here where I
1:33:46have to authorize the MCP connection.
1:33:49So, MCP client connection, authorize API
1:33:51access, I can say, "Yes."
1:33:53Choose your organization, authorize
1:33:55cloud code, authentication successful. I
1:33:57said done. It found my project, which
1:33:59takes 2 minutes to set up, make an
1:34:00account, and then pretty much make an
1:34:02organization, and that's it. And now
1:34:03it's going to create the table in the
1:34:06Superbase account. As you can see, it
1:34:08made a table called SOPs, and we should
1:34:09be able to see the information starting
1:34:11to come in. And as you can see, now we
1:34:13have the information that we had on the
1:34:14Google Doc, we have it here in the
1:34:17Superbase itself, right? And the way
1:34:19that this works again is that we have
1:34:21each text, which is an embedding. An
1:34:22embedding is just a number. And when we
1:34:24say query the vector database when we
1:34:26ask questions, it matches the numbers of
1:34:28the questions or the letters that we
1:34:30that we actually give it to the numbers
1:34:31here, and it gives us the answer, right?
1:34:33But essentially, what it did is that it
1:34:34used the MCP of Superbase itself to then
1:34:37send all the information here without us
1:34:39having to set up anything apart from
1:34:40that MCP code itself. Now, although MCPs
1:34:43are great, there is one fallback, which
1:34:45is the cost. And so, if I go here to
1:34:47{slash} context,
1:34:49I will now be able to see how many
1:34:51tokens the MCPs actually take. So, here
1:34:54I can see that I'm using Cloud Opus 4.6,
1:34:56I'm using 15.6K out of 1 million tokens,
1:34:59so 2%, and only two MCPs use 10,800
1:35:03[music] tokens, which is quite a big
1:35:05amount considering the fact that we only
1:35:06added two. And you can see here that the
1:35:08MCP tools, right? So, each tool, which
1:35:10means that each type of action that it
1:35:12can do is requires tokens to run.
1:35:15>> [music]
1:35:15>> And so, we have click, which is 163,
1:35:17close page 112, drag 155, emulate, and
1:35:21so on, right? [music] And it keeps on
1:35:22going and going. So, one important thing
1:35:23is that the more MCPs you add, the more
1:35:26money you have to spend on tokens.
1:35:28>> [music]
1:35:28>> So, make sure to not use 15 or 20 MCPs
1:35:30because they will take up a lot of your
1:35:33tokens and contexts. Just use the top
1:35:35two to three that I showed you, and make
1:35:37those your main priority ones whenever
1:35:39you're using Cloud Code. And so, with
1:35:40the fact that you are spending a lot of
1:35:41tokens with the MCPs itself, usually
1:35:43what we do is that we install the MCPs
1:35:45here, right? So, let's say we install
1:35:47Chrome DevTools, or let's say we install
1:35:48the Superbase. We check if it works,
1:35:50which means that you test it, see if it
1:35:51did the right thing, and then you turn
1:35:53that MCP itself into a skill because
1:35:55skills require less tokens. And so, with
1:35:57the healthy setup, you typically have a
1:35:59system prompt, which is the prompt of
1:36:01the actual AI itself, Claude code. Then
1:36:03you have built-in tools, and then you
1:36:05have MCP tools. You want to have a few,
1:36:07and then you have your conversation,
1:36:08which means it's a lot of room cuz the
1:36:10more you speak to Claude code, the more
1:36:11the tokens or what we call the context
1:36:13window increases, which means that you
1:36:15can't be having a conversation with
1:36:16Claude code forever because the tokens,
1:36:18the space of the conversation ends,
1:36:20right? And so, you want to have as much
1:36:22space as possible for the conversation
1:36:24instead of having the system prompt,
1:36:25built-in tools, and having 20 different
1:36:27MCP servers, which take up so much
1:36:29tokens and so much space and context
1:36:30[music] where to the point where you
1:36:32can't actually have a proper
1:36:33conversation with Claude code before the
1:36:35tokens runs out. So, you want to be as
1:36:36efficient as possible with your tokens
1:36:38that you are using in Claude code. And
1:36:39by the way, if you want all the
1:36:40resources that I'm going to be talking
1:36:42about in this video, plus being part of
1:36:44a community of over 9,000 people who are
1:36:46actively building with AI, then check
1:36:48out the second link down below.
1:36:54All right. So, congrats on finishing
1:36:55module two. At this point, you have
1:36:58fully customized Claude code. You have
1:37:00custom skills, you have automated hooks,
1:37:02and it's connected to every single tool
1:37:04through the use of MCP. So, now you have
1:37:06a pretty powerful setup. But if there's
1:37:08one thing I learned after working with
1:37:09Claude code for hundreds of hours, is
1:37:11that it's not so much about your setup
1:37:13or your tools, but it's about how you
1:37:15speak to Claude code, which is the
1:37:16prompting. And that's exactly what the
1:37:18next module is all about, which is the
1:37:19prompting strategies that will 10x the
1:37:21quality of everything that you do with
1:37:23Claude code. Let's dive in. Okay. So,
1:37:24the first hack is called splitting
1:37:26prompts. Now, one of the biggest
1:37:27mistakes that I see when people use
1:37:28Claude code is them adding huge and
1:37:31humongous prompts with a ton of
1:37:32information and a ton of tasks all in
1:37:34the same prompt, and then expecting
1:37:35Claude code to give them the output that
1:37:37they want. An example of [music] this
1:37:38could be them saying, "Build me a CRM
1:37:40with off database API dashboard emails
1:37:42and deploy it." Which are basically
1:37:44seven different tasks, right? Or four
1:37:45different tasks within the same kind of
1:37:47major thing that they want to build. And
1:37:48there's a few problems with this. The
1:37:49first one being that there's a lot of
1:37:50context or tokens used [music]
1:37:52in that same session that you're using
1:37:54Claude code. The second thing is that
1:37:55the output is not going to be the best
1:37:57just because you're feeding it and
1:37:58you're telling it 20 different things at
1:37:59once. The same way that you tell an
1:38:00employee, "Hey, can you build me a full
1:38:02software with this, this, this, and
1:38:03that?" without giving the person any
1:38:04clear structure. And the issue here,
1:38:06even when I started, was that I thought
1:38:08Claude code would build anything, right?
1:38:09It could build anything and anything if
1:38:10I give it one prompt and that's it. But
1:38:12it's much better to split prompts and
1:38:14take a more structured approach. So,
1:38:15instead of putting this prompt right
1:38:16here with seven different tasks, we
1:38:18split the tasks up, right? And we talk
1:38:20to Claude code task by task. So, we look
1:38:22at this and we say, "Okay, the first
1:38:23task is setting up the project." Then,
1:38:25it will set up the project and then we
1:38:26review whether it did a good job or not.
1:38:28If it did a good job, then we go to the
1:38:29next thing. If it didn't do a good job,
1:38:31then we go back and fix it. Then we go
1:38:32to task two, which is knowledge base,
1:38:34and then we review if it worked. If not,
1:38:35then we look back. And we do this for
1:38:37the other tasks as well. And the reason
1:38:38this works a lot better is because when
1:38:40we have to fix task four, we don't have
1:38:42to fix the whole thing cuz we reviewed
1:38:43each one and it worked very, very well.
1:38:45And so, we go task by task, making sure
1:38:46that everything works until the end.
1:38:47Now, is it going to take more time to
1:38:49prompt Claude code at the start? Yes.
1:38:50But actually, you'd spend more time
1:38:52fixing this after it didn't do a good
1:38:54job rather than setting things up
1:38:55properly from from the start. So, to
1:38:56make this more practical, I can give you
1:38:58like an actual prompt that I built. So,
1:39:00you can see the bad approach and then
1:39:01the good approach that we take. So, on
1:39:02the one hand, we have a prompt that
1:39:03says, "Build me a full customer support
1:39:05chatbot with Next.js front end and a
1:39:06knowledge base that uses embeddings,
1:39:08escalation logic that routes tickets to
1:39:09Slack, an admin dashboard with
1:39:10analytics, user authentication with
1:39:12OAuth, and deploy the whole thing onto
1:39:13Vercel." Use Tailwind for styling and
1:39:15make it responsive. Now, I'm not going
1:39:17to run this for the sake of not having
1:39:18to wait 3 hours, but this right here, it
1:39:21will overwhelm Claude code with so much
1:39:23context and so many things to do that
1:39:25the output, don't get me wrong, it will
1:39:27be okay, but it won't be as good as if
1:39:29you split things up step by step, right?
1:39:31And so, what we could do here is we have
1:39:34the first prompt, which would be this,
1:39:36let's say. Prompt one, set up the
1:39:38Next.js project with the TypeScript and
1:39:39Tailwind, create the basic folder
1:39:41structure with components, and we give
1:39:42it the first part. So, when prompt one
1:39:44is finished, it did a good job, [music]
1:39:45then we go to prompt two right here.
1:39:47Then, once prompt two is finished, so
1:39:48this is added to the knowledge base,
1:39:50it goes to prompt three,
1:39:52and then we keep going until prompt four
1:39:54and prompt five, right? So, prompt four
1:39:55will be this. So, all we did here is
1:39:57took this prompt right here, which has
1:39:58Next.js front end and knowledge base,
1:40:00escalation logic, it has admin
1:40:02dashboard, it has OAuth, it has
1:40:03deploying, right? So, Next.js front end
1:40:05will go here,
1:40:06then we had knowledge base will go here,
1:40:08then we had escalation logic will go
1:40:09here,
1:40:10then we had dashboard, which will go
1:40:11here, and then finally we would have
1:40:13another one, which is deploying the
1:40:14whole thing. And so, you can see that
1:40:15it's it's kind of like a spiderweb, and
1:40:17the reason why we do this is because you
1:40:18take a more structured approach, right?
1:40:20And so, I walked you through how you
1:40:21review it and so on, but essentially you
1:40:22would run this, right? And don't worry
1:40:23about all these terms cuz they don't
1:40:25mean much to me either, but it's more
1:40:26the structure that we have to
1:40:27understand, right? And so, you have
1:40:28something simple with seven different
1:40:30tasks, you split them up into different
1:40:31prompts, prompt one, prompt two until
1:40:33prompt seven, you review them
1:40:34individually to make sure that when you
1:40:35get the output, it's going to be
1:40:36correct, and you don't have to re-prompt
1:40:37Claude all over again and fix it and
1:40:40debug it and spend two, three hours
1:40:41asking it why it did the wrong thing.
1:40:42The next hack is using plan mode. Now, I
1:40:44know plan mode is not something
1:40:45revolutionary, but there's a big
1:40:46misconception as to what it actually
1:40:48does and when it's important. So,
1:40:49whenever we have a new task in Claude
1:40:51Code, we ask ourselves whether it's hard
1:40:52or not hard. If it's hard, so if it
1:40:54requires different steps, then we use
1:40:55plan mode, then we review the plan, and
1:40:57then we approve it, and then we execute
1:40:58the actual thing to give us the output
1:40:59that we want, right? An example of this
1:41:00could be an API with five end points, or
1:41:02when you're building a full workflow.
1:41:04And I'll show you exactly what that
1:41:04looks like. Because it has different
1:41:05steps to it, and you want it to plan
1:41:07accordingly, take the time to show it to
1:41:09think through how it's going to build it
1:41:10before it builds it. On the other hand,
1:41:11if you have something very easy like
1:41:13fixing a typo or changing some sort of
1:41:14color in a button that it made in a
1:41:16software, then you can just execute
1:41:17directly without needing the plan mode
1:41:19because there's not much planning to it,
1:41:20it's just a very, very easy task. All
1:41:22right, so here I'm on my main hub in
1:41:23Claude Code, which is called the YouTube
1:41:24workspace. Um I automate pretty much 30%
1:41:27of everything that I do on a day-to-day
1:41:28basis with like YouTube. I would even go
1:41:30as far as saying like 40%. Now, here
1:41:32under dot Claude, and by the way, if
1:41:33you're new to Claude Code, if you have
1:41:34no idea what I'm talking about, if this
1:41:35looks very confusing to you, watch this
1:41:37video up here. It will give you a much
1:41:38better idea. But, if you understand
1:41:39skills, which I'll walk you through in
1:41:41the last talk as well. It's Cali Draw
1:41:43diagram. So, that skill right there
1:41:44allows me to be able to make these
1:41:46diagrams automatically. To break down
1:41:47the topic for you guys to understand,
1:41:48[music] I typically break it down in a
1:41:50visual way, but also make it practical.
1:41:52So, you see how I'm showing you the
1:41:53diagram, plus also how you can use the
1:41:55actual concept here. And so, in this
1:41:57case, let's say I wanted to make a video
1:41:59about anti-gravity versus Cloud Code. By
1:42:00the way, if you want the video, let me
1:42:01know in the comments down below if
1:42:02you're interested. If not, also let me
1:42:04know.
1:42:05>> Uh-huh.
1:42:05>> But, if I wanted to do a video on
1:42:07anti-gravity versus Cloud Code, well, I
1:42:08would want to basically first of all
1:42:10explain what anti-gravity is and how
1:42:12Cloud Code is different from
1:42:13anti-gravity. And so, I would want to
1:42:14make maybe three diagrams, I'd say,
1:42:16using Excalidraw, which is this platform
1:42:18right here, which allows me to explain
1:42:20my concept, the differentiation between
1:42:21anti-gravity and Cloud Code. Okay. So, I
1:42:23told it to research anti-gravity. I told
1:42:25it to research Cloud Code. I wanted to
1:42:27look for the differences. I wanted to
1:42:28make three visuals, and then give me
1:42:29back the visuals. Okay. So, a task like
1:42:31this has a plan, right? And so, instead
1:42:34of going bypass permissions, which is
1:42:36basically where we go yolo mode, we say,
1:42:37"Hey, take care of it yourself. Don't
1:42:38ask me for any help. Don't make any
1:42:40plan. Just go after it, right?" We can
1:42:41go and use plan mode to make a plan
1:42:44first before it gives us the output.
1:42:46Okay. So, now I'm going to press run.
1:42:48And you know it's on plan mode when it's
1:42:49blue right here. Bypass permissions is
1:42:52going to be red. Now, it's saying that
1:42:53it's going to research both products in
1:42:54parallel to understand what we're
1:42:55working with, then plan the diagrams.
1:42:57Okay. So, that's how it plans through.
1:42:58That's how it thinks through actually
1:42:59planning the the whole concept here.
1:43:00Okay, there you go. So, after a few
1:43:01minutes, it made the plan, which is
1:43:03Google's anti-gravity versus Cloud Code
1:43:05three visual comparison diagrams. It
1:43:06gives me the context, so Michael A wants
1:43:08to create XYZ, research summary, so the
1:43:10summary that it made, three key
1:43:12differentiators, so interface, agent
1:43:13orchestration, verification, and
1:43:14feedback, and deliverables, which is the
1:43:16three things that we want, right? With
1:43:17the color system, the layout, diagram
1:43:19one, diagram two, diagram three, diagram
1:43:20four, and the implementation steps, and
1:43:22the files and verification. One more
1:43:23reason why I do also like this in
1:43:25specific is because when I use plan
1:43:26mode, I can also add comments. So, I can
1:43:28say interface paradigm. Paradigm is a
1:43:30very strange word. Make it
1:43:32beginner-friendly. I know it's a very
1:43:33stupid comment, but let's say you add
1:43:34the comment. When we press approve plan,
1:43:36then we will approve the plan, but it
1:43:38will also take the comments that we add
1:43:39in the plan as context to change the
1:43:41whatever it is that we told it to
1:43:42change. And but also I get to see
1:43:43exactly what it came up with. Cuz if I
1:43:45didn't like interface paradigm or if I
1:43:47didn't like agent organization, I could
1:43:48say, "Hey, can you find something else?"
1:43:50Right? And we can go back and forth.
1:43:51Now, let's say I I like the plan,
1:43:52everything's good. Now, I can press send
1:43:54feedback and keep planning. So, send
1:43:56feedback because we we added one
1:43:57comment. And if you add more comments,
1:43:59so if you say here, let's say
1:44:00orchestration, you add turn this to
1:44:01organization. Turn this word.
1:44:05So, we're just playing around with the
1:44:06names.
1:44:07I can go here and I can send feedback
1:44:08and keep planning. Or, actually now that
1:44:10I thought of it, color system. I know
1:44:11antigravity is not purple, so we can do
1:44:13use dark blue. Clockcode will be not
1:44:18blue, it will be
1:44:20uh use orange.
1:44:22That's it. Now, I can go back. It will
1:44:24say we added four comments. It will take
1:44:25these four comments and we can send
1:44:26feedback and keep planning. Okay, so now
1:44:28it will take the feedback, it will apply
1:44:30the feedback, and then give us back
1:44:31another plan with the updated feedback.
1:44:33It's a lot of feedback there, but you
1:44:34get the point here. It's basically
1:44:36crafting the whole plan before it gives
1:44:37it to us. All right, so here we have the
1:44:38new updated plan. If I go back here, we
1:44:40can see that we have dark blue, orange,
1:44:42we have interface style, and we have
1:44:43agent organization. So, these things
1:44:45changed. It updated the actual plan. And
1:44:46now I can say yes and now to accept. And
1:44:48now instead of making it edit
1:44:49automatically, I can now put bypass
1:44:51permissions because I know that the plan
1:44:53is good. I know that it's smart enough
1:44:55to be able to do whatever I want it to
1:44:56do. And now it's just about doing it.
1:44:58All right, so it just finished making
1:44:59the three diagrams. Now, let's look at
1:45:00the output. So, I can look at the first
1:45:02one. I can go here, I can import scenes.
1:45:04So, antigravity, I believe this is an
1:45:06interface. So, let's open it up.
1:45:08There we go. Not crazy. Insane. So, here
1:45:11we have the actual diagram that it made.
1:45:12Google antigravity and clockcode. This
1:45:14is the diagram itself to show the the
1:45:15differentiation in the interface. The
1:45:17next one is agent organization, which is
1:45:19this one right here, which is on the
1:45:21left-hand side is antigravity, on the
1:45:23right-hand side is clockcode, showing
1:45:24the difference. And finally, we have the
1:45:26agent verification, which is this one
1:45:27right here. So, the verification and how
1:45:29they both work. Now, the output was
1:45:30significantly better because we took the
1:45:32time to plan things ahead. On the other
1:45:34side, you could have just used bypass
1:45:35permissions, and it would have given you
1:45:36something, and I've tried it before, and
1:45:38the output just wouldn't be as good as
1:45:39you'd want it to be. So, if a task
1:45:41requires more thinking, if it requires a
1:45:42plan, just make sure to use plan mode
1:45:44every single time.
1:45:47So, the difference between a kitchen
1:45:49sink versus a clean session is the
1:45:50amount of tokens that it uses. So, if
1:45:52you didn't know, a Claude could session,
1:45:53right? One session means one kind of
1:45:55conversation here. Look here. I just
1:45:57press this, it will compact it. I'll
1:45:58explain what that means, but
1:45:59essentially, each conversation has X
1:46:01amount of tokens. So, if I go here and I
1:46:02put context,
1:46:04we get the context usage, right? So, we
1:46:06get how many tokens have we used so far.
1:46:08So, we have a limit of 200,000 tokens,
1:46:10and we have used about 83,900 tokens in
1:46:13one session. So, the whole session that
1:46:15we just did, it [music] used about
1:46:1683,900, which is about 42% of each one.
1:46:19And it also gives us the actual like
1:46:20percentage broken down into like what
1:46:22actually takes the most amount of
1:46:23tokens. So, the system message is 3.8,
1:46:25[music] the system tools is 9.3, the MCP
1:46:27is 5.7, the system tools is 10.8, and so
1:46:30on, right? And so, you have all the
1:46:31tools broken down, and you can see
1:46:32everything here. My point is that your
1:46:34prompts are better if you separate them
1:46:36in different sessions, right? So,
1:46:37instead of fixing a bug, making a new
1:46:39feature, refactoring something, or
1:46:40writing a doc, all in the same session,
1:46:42right? Even though they have nothing to
1:46:43do with each other, you would create a
1:46:44session for each one. And when I say
1:46:45session, I mean a new session here. So,
1:46:47you press plus, this is a session.
1:46:49Conversation. You can call it a
1:46:49conversation, really, not a session. So,
1:46:51conversation one, conversation two. Cuz
1:46:52the more tokens you use in a session,
1:46:54the more stupid Claude could becomes.
1:46:56The more it starts to think less, and
1:46:57the more it starts to get overwhelmed
1:46:58with information. And so, split them up
1:47:01in different conversations, and then you
1:47:02can start over. Now, I call this prompt
1:47:03hack because sometimes you think that
1:47:05the problem is your prompt, but really
1:47:07it's the session tokens that you're
1:47:08using. So, right here, we basically made
1:47:09three diagrams for a video that I'm
1:47:11planning to make. Now, we're using about
1:47:1242% of the actual usage. If I keep going
1:47:15with this conversation, we could do
1:47:16that, right? But the level of
1:47:19intelligence of Claude could will
1:47:20decrease the more you increase the
1:47:22tokens in that conversation. So, if I
1:47:24wanted to deploy another workflow, let's
1:47:25say comment analysis, it goes to my
1:47:26competitors, it goes through all their
1:47:28comments, see what people want, and it
1:47:29gives you back a report. Is it better
1:47:31for me to keep going on the session, or
1:47:32is it better for me to go to a new
1:47:34session and start from there? Well,
1:47:35obviously it's better to go to a new
1:47:36session and start from there because you
1:47:37start from scratch, and if I go here and
1:47:39I put context,
1:47:41then this will now give me the context,
1:47:42and we're only using about 9% token
1:47:44usage. So, now it has a fresh mind and
1:47:46it can start doing the task that you
1:47:47wanted to do without being overwhelmed
1:47:48with all the information from the
1:47:49previous conversation that we had, which
1:47:51did something completely different. The
1:47:52next hack is using Claude.md. So, let's
1:47:54say you wanted to make an app, some sort
1:47:55of software app, okay? And you wanted
1:47:57the app to use TypeScript, you wanted
1:47:58all the words to follow camel case,
1:48:00>> [music]
1:48:00>> you wanted to use Tailwind, you wanted
1:48:02to use Superbase, and you wanted to use
1:48:04an app router, not pages. Now, don't
1:48:06worry if you're not really understanding
1:48:07what all these words mean, but
1:48:09understand that these are all rules,
1:48:10right? And so, these rules, you can
1:48:12either tell Claude code every single
1:48:13time to do all these things, or you can
1:48:15add them to something that we call a
1:48:16Claude.md document where you basically
1:48:18build a one-time prompt, and it will
1:48:20auto-load that prompt every single time
1:48:22in Claude, so it has context, and we
1:48:23don't have to repeat everything over and
1:48:25over again. So, Claude already knows,
1:48:26you just ask. So, if we had to make a
1:48:27PDF document, and we wanted the color to
1:48:29be blue, we wanted to see the the
1:48:30letters to be small, we wanted the I
1:48:32guess that the font to be italic font,
1:48:35right? We can either tell it
1:48:36individually every single time, or we
1:48:38can add it to a prompt like this, and
1:48:39then it will use that every single time
1:48:41>> [music]
1:48:41>> without us having to repeat ourselves.
1:48:42Now, lucky for you, I made a skill in
1:48:44Claude code called the Claude.md
1:48:45generator. And this includes a pull
1:48:47prompt on how to make the best Claude.md
1:48:49document, plus it gives you the example
1:48:51structure of one of them, for example,
1:48:52here, right? So, all you have to do is
1:48:54download the skill, which you can find
1:48:56in the second link down below. And here
1:48:57we can say, "Hey, I want to create a
1:48:58Claude.md prompt for my project uh for
1:49:01my app that I'm building. I just want
1:49:02you to ask me any questions of what you
1:49:04need from me to be able to make this
1:49:05document." And we can press go. So, now
1:49:06what it's going to do is it's going to
1:49:07look at the skill of all the
1:49:08requirements that it needs to be able to
1:49:09make the Claude.md document, and then
1:49:11we're going to start working on it. So,
1:49:12as you can see here, it says, "I have
1:49:13found the Claude.md generator skill.
1:49:15Now, let's walk through the process."
1:49:16So, now it's going to ask me where I
1:49:17should save the cloud.md file. I can
1:49:19press uh recommended, it's fine. And
1:49:21then, tell me about your project. So,
1:49:22what is the project name? What does it
1:49:23do? And who is it for? I can press other
1:49:25and I can start talking to it. So, let's
1:49:26say here I say, "Okay, so I want to
1:49:27build a website." So, this is for a
1:49:29website and the website needs to be for
1:49:31my company, JM Solutions. Uh we're an AI
1:49:33automation agency that helps marketing
1:49:35agencies to be able to build their
1:49:37back-end delivery systems. And uh yeah,
1:49:39that's what that's what the actual thing
1:49:40is for. So, you give a context and then
1:49:42the tech stack we'll use is I'm not too
1:49:44sure. Whatever you think is the best for
1:49:46build a website, let's just do that. You
1:49:48can submit the answer. And now, it will
1:49:49start thinking about what to put in the
1:49:50cloud.md document. Okay, so it's not JM
1:49:52Solutions, it's JM. But, now when it
1:49:54makes the cloud.md document, it will be
1:49:55an overview of what it is that we want
1:49:58to build. For your JM Solutions website,
1:50:00what kind of pages features do you want?
1:50:01Landing plus services. Keep it simple.
1:50:04Keep it simple. Actually, for
1:50:06conventions,
1:50:07actually, let's keep it simple.
1:50:09>> [music]
1:50:09>> And rules, are there going to be any
1:50:10hard rules Cloud must always or never do
1:50:12in this project? I have specific rules.
1:50:14There you go. And now, it's most likely
1:50:15going to ask me for rules. Because all
1:50:16these rules, I don't want to repeat
1:50:18myself every single time I use Cloud
1:50:19Code. I want Cloud Code to use the
1:50:21cloud.md document when it's actually
1:50:23building the thing that I wanted it to
1:50:24build. What are your hard rules?
1:50:26Mobile-first design. No inline styles.
1:50:28There we go. Submit answers. [music]
1:50:30And also other, my company name is JM
1:50:32Solutions, not JAM.
1:50:35Okay, so here we have a plan to create
1:50:36the cloud.md document. We have the
1:50:38context, project of the JM Solutions AI
1:50:40automation agency website helping
1:50:42marketing agencies to build back-end
1:50:44delivery systems. The tech stack is
1:50:45going to be this. Pages, conventions,
1:50:47file structure, rules, actions,
1:50:49verification, and we're good to go. Yes,
1:50:51and now to accept. All right, so here it
1:50:53made the cloud.md document. If I open it
1:50:55up, I can see that we have a full
1:50:57prompt, which is the one that we had
1:50:58before. So, these are information,
1:51:01basically a summary of what it is that
1:51:02we want to build. And this is great for
1:51:04us because every single time that we
1:51:05have to go out of Cloud Code and come
1:51:06back in, we don't have to re-explain to
1:51:07Cloud Code what we want. We can just say
1:51:09it knows that every single time that we
1:51:11start talking to it, it will always look
1:51:13at the Claude.md document. It will get
1:51:14context and then only it will respond to
1:51:16us. Sort of like an employee having
1:51:18context to what it is that it has to do
1:51:19every day. I mean, imagine you give the
1:51:20AI the task to build a website, you give
1:51:22it all the different features and what
1:51:23it needs to do, then you go and close
1:51:26Claude Code, you come back to it later,
1:51:27and then you tell it, "Hey, keep going."
1:51:29And then it's going to ask you, "Keep
1:51:30going with what? What it is you want me
1:51:31to build?" And you have to re-explain
1:51:32everything again. That's what this
1:51:33prompt is for. The last prompt hack is
1:51:35using skills. So, we call it reusable
1:51:37prompts. So, initially, we had the
1:51:38manual way, which is same effort every
1:51:40time. So, we had session one where we
1:51:42write a full prompt. Session two, we
1:51:43write another full prompt. And we keep
1:51:45going session three, session four, and
1:51:46session five. And each session will take
1:51:47about five minutes, and we take about 25
1:51:49minutes that we wasted writing every
1:51:51single prompt manually. The second way
1:51:53is using skills. So, we write the skill
1:51:55once and we use it forever. So, if you
1:51:57remember before, I showed you how I made
1:51:59a Excalidraw diagram using what we call
1:52:01skills. The skills are preset
1:52:03instructions that we give Claude Code to
1:52:05perform a specific thing. And so, all we
1:52:07have to do is create the skill one time,
1:52:08which will take about five to 10
1:52:10minutes. And we also have a skill to
1:52:11create the skill, which is which is
1:52:12awesome. And so, whenever you want to do
1:52:13a specific thing, you can always create
1:52:15a skill and then call the skill whenever
1:52:17you want to do that specific thing
1:52:19whenever you want. And so, let's say we
1:52:20had the Excalidraw diagram thing. So,
1:52:22inside here, we have a prompt. This is a
1:52:24prompt that explains how to make the
1:52:26Excalidraw diagrams. So, this is a full
1:52:27prompt. Now, I could either copy and
1:52:29paste this prompt every single time into
1:52:30Claude Code into a new session, or I
1:52:32could always just save it here under the
1:52:35skills, and it could always read this
1:52:37when it needs to make the actual thing.
1:52:39Now, this is great for us because it
1:52:40saves me so much time whenever I want to
1:52:42make any of these, to be honest. So, for
1:52:44example, if I go back here and I go to
1:52:45the command, I can see here
1:52:47that I can call the Claude.md generator,
1:52:49I can call the comment analysis, I can
1:52:50call the comment analysis report,
1:52:52competitor analysis, which are all
1:52:53skills which perform different things.
1:52:55Now, I've covered skills step by step in
1:52:57my Claude Code tutorial. So, if this is
1:52:58the first time that you hear about this,
1:53:00just check out this video up here. It
1:53:01will give you much more context. But, if
1:53:03you already know what skills are, then
1:53:04let's actually to one step by step and
1:53:06show you how it works. Now, let's say I
1:53:07wanted to go on YouTube every day and
1:53:08check the amount of views that I have
1:53:10for the videos. Now, I could either
1:53:11prompt Clocko to do this every day or
1:53:13whenever I want to. And by the way,
1:53:14before we build the skills, you can find
1:53:16all of them at ai.tmpl.com.
1:53:19These are preset skills where we have
1:53:21the skill creator, which is a skill that
1:53:22makes skills. We have the senior front
1:53:24end, we have the UI design, we have all
1:53:25these different skills that do different
1:53:27things. But sometimes I want to make my
1:53:28own skill, which is my own kind of task.
1:53:30And so here, I can say, "Hey, I want to
1:53:32be able to make a skill, and the skill
1:53:34needs to be
1:53:35the skill that goes on the internet
1:53:37every time and it finds the top news or
1:53:40top features that came out about Clocko.
1:53:42Then it goes on YouTube and it checks
1:53:44the highest viewed video on Clocko in
1:53:46the moment, and then it gives me content
1:53:48ideas that I could be using for my
1:53:50channel for Clocko." Make a plan for it
1:53:52first and then we can
1:53:53go ahead and make a skill. So this is
1:53:55the initial thing that we give it. I can
1:53:56just press go, and now it will start
1:53:58making the plan for the skill that it
1:53:59needs to do. All right, so as you can
1:54:00see here, it made the plan for the skill
1:54:01that it's going to make, and it's going
1:54:02to be called Clocko content ideas. So
1:54:04what does is that it search the internet
1:54:06for the latest Clocko news, features, it
1:54:08searches YouTube for the highest viewed
1:54:09recent Clocko videos. It then generates
1:54:1110 to 15 ranked content ideas
1:54:13cross-referenced against Michael's
1:54:14content pillars, and then it outputs a
1:54:16markdown report plus styled PDF to
1:54:18outputs. So now it's going to be making
1:54:19all the skills for that specific thing,
1:54:22and then we can see phase one, phase
1:54:24two, phase three, phase four, phase
1:54:25five, phase six, and that's it. So from
1:54:28a single prompt, it made this full plan
1:54:29of execution that we can now simply
1:54:31press yes and now to accept, and I'm
1:54:33going to press bypass permissions so it
1:54:35does everything on its own. So as you
1:54:36can see here, it's going through the
1:54:37to-do's, tasks that it needs to make,
1:54:39and then it's adding the folder called
1:54:41Clocko content ideas, which is a folder
1:54:43where it's going to store our skill that
1:54:45we can always call when we want to to
1:54:46use it. All right, so as you can see
1:54:47here, it finished making the skill. If I
1:54:49go to Clocko content ideas,
1:54:51>> [music]
1:54:51>> I have the references, so this will be
1:54:52the report in PDF that it needs to give
1:54:54us, uh which is the output. Then the
1:54:56scripts, which is the actual Python
1:54:58scripts, it's kind of the the automation
1:54:59that it needs to execute, generate PDF,
1:55:01and search YouTube videos, and then it
1:55:03has the skill.md, which is a main
1:55:05document that explains to it exactly
1:55:07what we want to do. So, with that said,
1:55:10I can
1:55:11go here, and I can say / Claude content
1:55:15ideas. So, it also made a command, and
1:55:17the command allows us to be able to to
1:55:18just find it easily. So, I can press
1:55:20Claude code content ideas. I can put
1:55:21this in plan mode, and now it will start
1:55:23looking at all the different skills
1:55:25here, and it will go through each
1:55:26skill.md right here, and it will know
1:55:29that it needs to use the Claude code
1:55:30content ideas because it's associated
1:55:32with that command in specific. And so,
1:55:34what it will do here is it will start
1:55:35calling the Claude code content ideas,
1:55:37>> [music]
1:55:37>> and it will start executing everything
1:55:38that it needs to do. It says, "I have
1:55:39all the context needed. Let me write the
1:55:41plan and exit plan mode so we can
1:55:42execute." All right, it made the plan of
1:55:44what it needs to do, all the steps, the
1:55:45key files, the verification. We can go
1:55:48and press yes and auto accept, and we
1:55:50can put this on bypass permissions. So,
1:55:52the idea behind this is that whenever
1:55:53now we have to go back to Claude code,
1:55:55instead of prompting it again saying,
1:55:56"Hey, can you go on the web? Can you
1:55:57search this? Can you search that?" We
1:55:58can simply just press / and call the
1:56:00Claude code content ideas, and it will
1:56:02execute that thing that we made
1:56:04previously, which again will take us 10
1:56:06minutes at the start, but it will take
1:56:07us 3 seconds to be able to set this up
1:56:09or to call the tool whenever we have to
1:56:11run it. All right, so now it gave us the
1:56:13top five content ideas. It finished
1:56:15running the actual thing. Agent teams
1:56:17for agency operations, Claude code HTTP
1:56:19hooks, why don't delete my Claude.md,
1:56:21skills build a business command center,
1:56:23and Claude code for non-coders. And if I
1:56:25go to outputs, I can see that it also
1:56:27made
1:56:28uh let's go to here. It also made a
1:56:30report in PDF format. And the PDF looks
1:56:31like this. So, we have news and updates,
1:56:33the loop command which came out on
1:56:34Claude code, the Claude API skill, the
1:56:37effort levels and ultra think, auto
1:56:38memory, security, agent teams,
1:56:41enterprise AI agent ROI, and we have the
1:56:43top YouTube videos latest by 30 days.
1:56:45So, Next Drive is crushing it. Claude
1:56:47code work, there we go. Dan Martel, he's
1:56:49a legend. Uh Brian Cassell, and we have
1:56:51the other ones as well. And it keeps on
1:56:52going and going and going, and we have
1:56:54the patterns observed, content ideas,
1:56:55and so on. So, it took us like 10
1:56:56minutes to be able to make this whole
1:56:57skill, but now it's going to save us a
1:56:59lot more time when we have to run it
1:57:01again. Cuz now when I go to a new Claude
1:57:02session, I can simply go and press the
1:57:04Claude code content ideas. And when I
1:57:07press this, it runs it over and over
1:57:08again. So, instead of running the prompt
1:57:10over and over again, we have one time
1:57:11and we save ourselves an eternity every
1:57:13time we have to run the whole thing.
1:57:20All right, congrats on finishing module
1:57:21three, which is the prompting guide.
1:57:23Now, you know how to actually
1:57:24communicate with Claude code to make it
1:57:26do things that you want it to do in a
1:57:28very very effective way. And that
1:57:29matters because every single thing that
1:57:31we kind of covered from now on requires
1:57:33that skill. So, now we're going to be
1:57:34moving into agentic workflows, which is
1:57:36sub-agents, agent teams, workflow hacks,
1:57:38and the deployment side. And this is
1:57:39where you get Claude code to stop acting
1:57:41like a normal coding agent, but start
1:57:43becoming the system that runs the entire
1:57:46operations for you. And this is the
1:57:47highest leverage thing that you can
1:57:49build with Claude code right now. So,
1:57:51let's dive in. Okay, so think of Claude
1:57:52code as the business owner. Right now,
1:57:54what's happening is that you're asking
1:57:56Claude code to do everything. Research,
1:57:58write, test, organize, [music]
1:57:59right? And that's like running a
1:58:01one-person business forever. And so, you
1:58:02can think of sub-agents like hiring
1:58:04[music] different employees. Each one
1:58:06has their own brain, each one has their
1:58:08own job description, their own tools,
1:58:10and each one reports back to the main
1:58:11boss, which is Claude in this case,
1:58:13right? And so, what happens here is that
1:58:14you tell Claude a task, you give it a
1:58:16task of what it has to do, and then
1:58:19Claude delegates, which means that it
1:58:21gives the work to the right employee, to
1:58:23the right agent [music] in this case.
1:58:24And the employee works in their own
1:58:26space, and then hands back a clear
1:58:27summary to the actual agent, which then
1:58:29gives it back to you. So, you can think
1:58:31of [music] this as a research assistant,
1:58:33a copywriter, and a QA employee. And so,
1:58:35these are three different employees,
1:58:36which are able for us to split up the
1:58:38task across different people, so that
1:58:40each one is specialized in their own
1:58:41thing. And they all report back to the
1:58:43main agent [music] here. So, now that
1:58:44you know exactly what it is, how do you
1:58:46actually create sub-agents? So, there's
1:58:48the first way, which is using Claude
1:58:49code, and the second way, which is doing
1:58:51it manually. And we're going to look at
1:58:52both ways. Okay, so here I'm on Claude
1:58:54Code. I made a new folder called sub
1:58:56agents. And for us to be able to create
1:58:58the sub agent, I can type /agents. Okay.
1:59:02Now, if you're on this view right here,
1:59:04you will then see a symbol which is this
1:59:07one right here. This symbolizes that to
1:59:09run this exact command to make sub
1:59:11agents, you have to run it through your
1:59:12terminal. And to run it through your
1:59:13terminal, go to here. You can go up,
1:59:16terminal, you can press new terminal.
1:59:18And then you will be able to see the
1:59:19terminal here. And I can press Claude to
1:59:21initiate Claude to actually make Claude
1:59:23work. And then I'll have this screen
1:59:25right here. Let me actually go maxed
1:59:27out. And then I have this page right
1:59:29here. So now we're going to be working
1:59:30with this page specifically to make the
1:59:32actual sub agent. Okay. Now that we're
1:59:34here, we can press /agents even though
1:59:36the spelling is wrong. You can press
1:59:38enter. And now it's telling us, "Hey,
1:59:39agents. Do you want to create a new
1:59:41agent?" I can press enter again, which
1:59:43means yes. It's asking me the location.
1:59:44So project or personal. I can do
1:59:47project. That's fine. So now it's asking
1:59:49me how I want to create the agent. In
1:59:50this case, I'm going to do generate it
1:59:51with Claude. And now I have to explain
1:59:53what it is that I want. For the sake of
1:59:55just showing you how it actually creates
1:59:57the agent, I'm going to keep it very
1:59:58simple. I want a sub agent that can
2:00:00review or that can take an an idea of a
2:00:03of a YouTube video and then it can spin
2:00:05up 10 different YouTube ideas, give me
2:00:08the 10 ideas plus hooks that I can use
2:00:10for that specific video. Now, I'm using
2:00:12just a very simple content creation kind
2:00:13of use case, but at the end of the
2:00:15video, I'll also show you three
2:00:16different use cases where you can
2:00:18actually apply this inside a real
2:00:19business. Right? Now that we have this
2:00:21here, I can press enter. And now it's
2:00:22going to generate the agent from
2:00:23description. [music]
2:00:24And we should be able to see the files
2:00:26in the folder being created here, which
2:00:28is basically how it stores the actual
2:00:29agent [music] itself. Okay, so now it's
2:00:31asking me to select the tools. And I
2:00:32think of this is like the level of
2:00:34accessibility that you give to the agent
2:00:36itself. Like so for example, do you want
2:00:37to give it access to all the tools, read
2:00:39only, edit tools, execution tools, other
2:00:41tools? I'm going to press continue cuz I
2:00:44can give pretty much [music] access to
2:00:45everything. It's going to ask me to
2:00:47choose the model that we're going to use
2:00:49to be able to run the sub agent and I'll
2:00:51walk you through like the model and
2:00:52which one should you choose in the first
2:00:53place. I'm going to go for Sonnet cuz it
2:00:55says balance performance, best for most
2:00:57agents. Let's go. And then it's asking
2:00:59me for a color, so let's make this nice
2:01:01and pink. Then it's asking me where do I
2:01:03want to configure agent memory? Keep it
2:01:05very very simple, use product scope
2:01:07because then it's going to add it to
2:01:08your actual project and then now you can
2:01:10see that it made a folder called dot
2:01:12cloud and then I can press S to save. If
2:01:15I go here, I can see that I now made the
2:01:17agent memory which doesn't really have
2:01:19much in it and then it made the agent
2:01:21itself. So the agent itself, just so you
2:01:23know, each sub agent all it is is a
2:01:25prompt, right? The same way that you
2:01:27know skills, sub agents are just prompts
2:01:29as well. And a prompt is a list of
2:01:31instructions of what the agent should
2:01:33actually do. Okay, so now that we've
2:01:34made a sub agent step by step, let's go
2:01:36through the structure of the actual sub
2:01:38agent. Okay, I'm going to press X here.
2:01:39I'm going to expand this
2:01:40>> [music]
2:01:40>> and here you can see you have the name,
2:01:42description, model, color, memory and
2:01:44then we have a whole description of the
2:01:46actual sub agent and what it needs to
2:01:48do. At this point, you have it ready, go
2:01:50to your chat window and then you can
2:01:51press file, you can press save or you
2:01:53can press command S and it's saved to
2:01:55your computer as well. And so what
2:01:57happens is that you'll start to have
2:01:58more and more agents, which is more and
2:02:00more prompts, right? And so you can run
2:02:02agents simultaneously if you want. All
2:02:04right, so now you know how to make it
2:02:05automatically using Claude. But just
2:02:06like everything, if you don't really
2:02:07understand the fundamentals of how it
2:02:09all works, then if you're stuck on
2:02:11something or if Claude doesn't do it the
2:02:12right job, you won't know because you
2:02:14don't know what good looks like. So, for
2:02:15us to create the sub agents, we first
2:02:18have to create a folder called the dot
2:02:20cloud folder. Now in case you haven't
2:02:22watched the skills video, check it up
2:02:23here, but I use that exact folder itself
2:02:25to store skills as well. And so you have
2:02:27agents, you have skills, you have agent
2:02:29memory, you have commands, you have all
2:02:30these things in the same exact folder.
2:02:32And then we have to make a folder called
2:02:33agents,
2:02:34>> [music]
2:02:34>> make a folder called agent memory, which
2:02:36is the memory of the actual agent and
2:02:37then we have the actual markdown file,
2:02:40which is dot md because it's the way
2:02:42that is written like this with hashtags
2:02:44with asterisks
2:02:46to organize the text which is the
2:02:47instructions of the actual agent. So I
2:02:49deleted the previous sub agent that we
2:02:50had. First things first to make our own
2:02:51sub agent is make a folder called the
2:02:53dot Claude folder. And inside this
2:02:56folder we can make another folder called
2:02:58agent /memory. Boom. And then inside the
2:03:03dot Claude folder as well, we can put
2:03:05the agents. And so this looks like the
2:03:07exact same thing that we had before. And
2:03:08then inside the agent here we can then
2:03:10press new file and call this the name of
2:03:13the actual agent. So let's say we're
2:03:14doing email writer.md.
2:03:18Okay.
2:03:19And then you have the actual markdown
2:03:21file where you'll be actually writing
2:03:23these instructions. Now the instructions
2:03:25for a sub agent prompt are divided into
2:03:28different parts. We have the name, we
2:03:29have the description, we have the tools,
2:03:31the model, and then we have the actual
2:03:32instructions here. First things first we
2:03:33have the name which is email writer in
2:03:35this case.
2:03:36>> [music]
2:03:36>> It's what you call this agent. It's how
2:03:38you will be able to refer to it later.
2:03:40Okay, so remember the name. Then we have
2:03:41the description. So in this case, write
2:03:43professional client emails based on the
2:03:44context you give it. You tell Claude
2:03:45when to use this agent. Claude reads
2:03:47this to decide if it's the right fit and
2:03:49the description is the most important
2:03:51part because you tell Claude when to use
2:03:53this agent. So when you tell it, "Hey, I
2:03:55want to do XYZ." Claude will read this
2:03:57exact description to decide if this is
2:03:59the right agent for your request. Then
2:04:01we have tools. In this case we just
2:04:03wrote read or write. The tools are the
2:04:04permissions, right? So what this agent
2:04:06is allowed to do. Well, it's like
2:04:07onboarding someone into your own project
2:04:09management system but not giving them
2:04:10access to everything. Then we have the
2:04:12model which is in this case Sonnet which
2:04:13is the brain that we're going to use. So
2:04:15which Claude could brain do you want to
2:04:16use for this exact [music] use case? Now
2:04:18bear in mind that we have three
2:04:19different models. We have Haiku which is
2:04:21fast and cheap. We have Sonnet which is
2:04:23why it's at balance recommended. Most
2:04:25agents use this. [music] And then we
2:04:26have Opus which is the most powerful.
2:04:27Bear in mind that the most powerful they
2:04:29are, the more costly they are. Haiku is
2:04:31very very fast but it's only used for
2:04:33things like reading files, searching
2:04:34code, or quick lookups. So looking for
2:04:36different things that don't necessarily
2:04:38require a lot of brainpower. Then we
2:04:39have Sonnet, which is typically good for
2:04:41most things. It's medium cost, it's
2:04:43cost-effective, hence why it's we say
2:04:45balanced. But it's good for writing
2:04:46code, code reviews, and most of the
2:04:48daily tasks that you do out there, like
2:04:49writing emails or making content. And
2:04:51then we have Opus, which is the most
2:04:53powerful. You're rarely going to use
2:04:54this, but you would use this if you're
2:04:55doing very hard debugging, if you're
2:04:57making some sort of full-stack
2:04:58application, and you need a agent that
2:05:01can go there and debug, meaning fix the
2:05:03code that you're writing, or if it needs
2:05:05really, really complex reasoning for it,
2:05:07right? So, basically, the overall
2:05:08summary here is use Haiku for simple
2:05:10tasks, Sonnet for most work, and Opus
2:05:12only when you need the big brain. And
2:05:13once we're done with the model, then we
2:05:15have the instructions. So, these are the
2:05:16rules and personality and guidelines of
2:05:19the actual sub-agent. So, for example,
2:05:21we have uh you're a professional email
2:05:22writer for an AI automation agency. When
2:05:24asked to write an email, you ask who
2:05:26it's for and the goal. You write a
2:05:28clear, professional email, and you keep
2:05:29it under 150 words. Match the tone to
2:05:32the situation, always end with a call to
2:05:33action. So, this right here is the
2:05:35personality and step-by-step
2:05:36instructions. Think of this as an
2:05:38employee handbook. You tell it exactly
2:05:39how to do its job, right? So, these are
2:05:41the instructions that you give it. And
2:05:43so, that right there is the structure of
2:05:44the .md file, which is the file where
2:05:47we're going to give everything that we
2:05:49need uh to the sub-agent, to Claude in
2:05:51this case, to know that, "Hey, this is
2:05:52the sub-agent that we want to run." I'm
2:05:54just going to paste the example that I'd
2:05:55already showed you. We have the name,
2:05:57description, tools, model, and then we
2:05:59have the instructions here as well. So,
2:06:00now we can press command S to save it.
2:06:02We can go to Claude, and I can paste
2:06:04this prompt right here, which is "Write
2:06:06a follow-up email to my client John
2:06:07about the chatbot project." Now, what it
2:06:09should do is it should look at the
2:06:11emailwriter.md, which is this right
2:06:12here. It will read through the
2:06:13instructions, it will look at what it
2:06:15has to do. In this case, the first thing
2:06:16it should ask me is ask who the email is
2:06:18for and what the goal is. In this case,
2:06:20we already told it who it's for, not
2:06:22entirely because it doesn't have the
2:06:23email. So, it should ask us the email
2:06:25itself. And as you can see, it just made
2:06:27the email called emailjohnfollowup.md,
2:06:29where where have [music] the to John.
2:06:31And by the way, we could actually make
2:06:32it so that it can send the emails for
2:06:34us, but in this case, I just wanted to
2:06:35to write the email. We have the subject
2:06:37line, checking in chatbot project and
2:06:39next steps. Hey John, hope you're doing
2:06:40well. I wanted to touch base on the
2:06:41chatbot project and to see how things
2:06:43are progressing on your end. We've made
2:06:45solid progress so far and want to make
2:06:46sure that we keep the momentum going. If
2:06:47there's any questions, blockers, or
2:06:49feedback you've been sitting on, now's a
2:06:50great time to surface them. I'd love to
2:06:51align on the next steps and make sure
2:06:53we're on track with the timeline. Would
2:06:54you be available for a quick call this
2:06:55week or early next week? Happy to work
2:06:57around your schedule. Just let me know
2:06:58what works best for you. Looking forward
2:07:00to hearing from you. Best, Michaela.
2:07:01Yeah, very very good email. And as you
2:07:02can see, we just gave it a simple prompt
2:07:05to write a follow-up email to my client
2:07:06John about the chatbot project. Then he
2:07:08knew that it had to use the agent email
2:07:10writer.md to fulfill the actual task.
2:07:12Now, in the grand scheme of things, you
2:07:13will have the output. So, some of them
2:07:14are outputs, right? Which is this right
2:07:16here, which is the output of the task
2:07:17that we gave it. But sometimes it can
2:07:19get a very very messy. So, what we
2:07:20usually do is we make a folder
2:07:23which is not in the doc cloud folder
2:07:25called outputs. And then we store this
2:07:27right here.
2:07:29Move. So, the all the outputs go under
2:07:31outputs folder. It's also keep things
2:07:33very very structured. With that said,
2:07:34you got to see how I built it plus how I
2:07:36ran it step by step. Now, in terms of
2:07:38triggering, which is like how do you
2:07:39start the agent itself? You saw me
2:07:41talking to the cloud agent and not
2:07:43really referencing the name of the
2:07:44actual agent itself. But that doesn't
2:07:46work too well when it comes to you
2:07:47having skills that are very very
2:07:48similar. And so, the main ways that we
2:07:50actually run the actual agent is three
2:07:53different ways. We have the natural
2:07:54language, which is similar to what I
2:07:56just did, but in this case, we ask it to
2:07:58use the code reviewer to check my
2:08:00changes. Use the name of the actual sub
2:08:02agent to do XYZ, to run the task. So,
2:08:04cloud figures out which agent to use and
2:08:06then it runs the task. The second and
2:08:08the most common one is mentioning.
2:08:09[music] So, we're going to use at the
2:08:11name of the actual sub agent to tell it
2:08:13to initiate XYZ or to do whatever it has
2:08:15to do, right? Create the task. Now, this
2:08:17is guaranteed to trigger because you're
2:08:19referencing that specific agent only,
2:08:21right? And then finally, we have session
2:08:23wide, which is use one agent for
2:08:25everything. So, you cloud {slash}
2:08:26{slash} agent code reviewer. So, my
2:08:28advice is to start with the first one,
2:08:29natural language, cuz that is the
2:08:30easiest to start with, and then move on
2:08:32to the second one mentioning, right? The
2:08:34only reason for that is because most
2:08:36times, you guys are not going to have 35
2:08:38different sub-agents that you have to
2:08:40manage, right? And so, you want to keep
2:08:41it as simple as possible. So, right
2:08:42here, to reference a sub-agent, we can
2:08:44just say, "Can you make an email using
2:08:47the email writer agent to Jessica
2:08:50telling her we have dinner tomorrow."
2:08:51And then you put [music] the actual
2:08:52name, right? You can also not put the
2:08:53slash itself. It just a
2:08:55I like to do to keep it more precise.
2:08:57And in this right here, it will know
2:08:58that it needs to use the agent itself.
2:09:01It's going to look through the agents
2:09:02here. Of course, there's only one,
2:09:04right? And it drafted the email itself.
2:09:05The second way, like I mentioned, is you
2:09:07go to tag, and you can see the different
2:09:09options, and you would use the email
2:09:11writer.md. And then you can say, "Can
2:09:13you make an email for James and Jack
2:09:16telling them, 'Thank you for the
2:09:17business meeting that we had before, and
2:09:19tell them that the next steps is to have
2:09:20an onboarding call tomorrow 5:00 p.m.
2:09:22EST to go through the next steps.'" And
2:09:24so, because I was very specific, and
2:09:25turn this off so you know that you're
2:09:27not reading it, I'm not cheating. But
2:09:28we're exactly referencing the agent that
2:09:30we want to use, and so there's no way
2:09:32that it will fail in using the agent
2:09:34because we give it the exact
2:09:35instructions to use that agent. And as
2:09:37you can see, it made the email itself,
2:09:39James and Jack, saying, "Great meeting,
2:09:41onboarding call tomorrow 5:00 EST." And
2:09:42then it made the whole thing, right?
2:09:44Obviously, you can change this by
2:09:45putting your name. I didn't really give
2:09:46it the main folder, which is a Claude.md
2:09:48folder, just cuz I didn't want to spend
2:09:50too much time on that, but you can give
2:09:51it more information as well. Now, in
2:09:52case that you are using the terminal to
2:09:54run Claude, you can simply go to {slash}
2:09:56agents again, and then you'll be able to
2:09:58see the email writer agent, which is
2:09:59this one right here. And you can simply
2:10:01press enter again. You can view, edit,
2:10:03delete agent, you can back agent, and
2:10:05you can do all sorts of things here. And
2:10:06again, you can simply just give it
2:10:07natural language, or you can press {at}
2:10:10email writer agent, and then you can
2:10:12tell it whatever you want. Make an email
2:10:14to James tomorrow. And you can run this,
2:10:15[music] and it pretty much does the
2:10:16exact same thing. I just want to show
2:10:18both things because I know some people
2:10:19get When I see some Some I mean me at
2:10:20the start. I got a bit confused when it
2:10:22came to using the terminal versus using
2:10:24the normal Claude chat and how those
2:10:26actually work. So, as you can see now,
2:10:27one thing we didn't have before, I told
2:10:29Claude that if I don't tell it who it
2:10:30specifically have to send the email to
2:10:32and what the goal of the email is, then
2:10:34it should ask me. And so, here it's
2:10:35saying, "Who is James and what's the
2:10:36purpose of the email?" James is a client
2:10:38and the purpose of the email is follow
2:10:39up after a meeting. WhisperFlow has
2:10:40changed my life, guys. You should
2:10:42definitely try it out. As you can see,
2:10:43we have the full email, James, subject,
2:10:45and the body itself. Now, one more thing
2:10:46that I spoke about before is
2:10:48restrictions and safety. Cuz I think
2:10:49about sub-agents is that they get
2:10:51really, really powerful because you can
2:10:53control what they can and cannot do. The
2:10:55same way that you wouldn't give an
2:10:56intern access to the whole company
2:10:58database, here's the exact same thing.
2:10:59And so, we have the read-only agent, we
2:11:01have the balanced agent, and we have the
2:11:02full access agent. Now, when you say
2:11:04read-only, it can look at your files, it
2:11:06can search through the code, it can
2:11:07explore uh kind of what you have, but it
2:11:09cannot edit or write, right? It can
2:11:11look, but can't touch. It's perfect for
2:11:12research that you have to do without it
2:11:14actually doing something on your
2:11:16documents or changing anything in your
2:11:17computer. Then we have the balanced
2:11:18agent, it can read, search, it can also
2:11:21edit and run commands, but it cannot
2:11:22delete files, right? It can build
2:11:24things, but can't break things. It can
2:11:26only add, but it can't remove, which is
2:11:27great because a lot of times when you
2:11:29have an agent that's out of hand, that's
2:11:30going everywhere, it can delete things
2:11:32that later on you can't get back, right?
2:11:34And so, you want to keep it normal and
2:11:36clear. And then we have the full access
2:11:37agent, it can read, edit, write, so full
2:11:39power, only use this when you really
2:11:41trust the task you're doing. Typically,
2:11:42it's for tasks that are easy, that don't
2:11:44require a lot of brain power, that are
2:11:45very, very structured, you have a
2:11:46step-by-step, it's done in this way, and
2:11:48it can be done pretty fast. So, here you
2:11:50decide what each agent can do. The less
2:11:52access the safer it is, the more access
2:11:54the more powerful, but the riskier it is
2:11:56for you as the person who's actually
2:11:57talking to the agents. And the way that
2:11:59you would change the actual permission
2:12:01is if I go here to the MD, you would
2:12:02have tools, right? The tools is the
2:12:04permission. So, you can literally think
2:12:05of this as if the word is permissions,
2:12:07but we use tools, right? Now, if you
2:12:09only want this to read, you can copy
2:12:12this, read, rep, and glob. That's it.
2:12:15Now, if you want a balanced, which is
2:12:17the most common, then you want this
2:12:19here. Read, edit, bash, grep, and glob.
2:12:21It's also going to give you suggestions,
2:12:22but I'm going to keep it normal for now.
2:12:24And then, if you want the full access,
2:12:26you don't have to specify tools. It
2:12:28inherits everything. So, if this is
2:12:29empty, then it will use any tool that it
2:12:31has access to. It's only if you specify
2:12:33the level of permission that you want
2:12:35the sub agent to have that it actually
2:12:36uses it per se. Now, when it comes to
2:12:38sub agents, there's one issue, which is
2:12:41the issue of memory. Cuz every single
2:12:42time that you have a new session with
2:12:43Clockcode, so a new conversation, these
2:12:46sub agents start from scratch, which
2:12:47means that they don't have a previous
2:12:49memory of the conversation that you had
2:12:50before. Now, as you can see here, we
2:12:52have a folder called agent memory. And
2:12:54this is where this comes into play.
2:12:55Because now we're able to actually have
2:12:57the sub agent remember the previous
2:12:59conversations, and [music] we don't have
2:13:00to re-explain things again. So, all we
2:13:02have to do is go to the actual agent
2:13:04itself, the prompt, and then we can
2:13:05simply right below model and above this,
2:13:08we wanted to include. But, all you have
2:13:09to do here is put memory, colon,
2:13:11project. Okay? And you press save. So,
2:13:13let me show you how it actually works.
2:13:15So, if I go to a new tab, of course,
2:13:16again, [music] save the project, so it's
2:13:18here. So, I can go to a new session, and
2:13:20I can say emailwriter.md.
2:13:22That's fine. [music] I can paste this
2:13:24right here, which says, "Make an email
2:13:26for John saying that we have dinner
2:13:27tonight or tomorrow night at 5:00 p.m.,
2:13:29and make it really, really, really,
2:13:30really, really happy with an exclamation
2:13:32mark. Also, add the tone to the memory."
2:13:37And I can press go. And so, now what it
2:13:38does, it's going to launch the email,
2:13:39right? To actually make the email that
2:13:40we asked it. But, it's also going to
2:13:42save the tone preference in parallel, so
2:13:44at the same time. Write the email and
2:13:45save the tone into the memory itself.
2:13:47So, you can see now it's starting to
2:13:48work with memory.md. It's going to save
2:13:50the tone preference to the memory
2:13:51itself, which is this folder right here.
2:13:52I can press accept. That's fine. Accept
2:13:54again. So, as you can see here, it made
2:13:56the email, dinner email.md, with all the
2:13:58exclamation marks, its own voice. And
2:14:00then we have the agent memory. The agent
2:14:02memory has two different things. It has
2:14:03feedback tone matching.md, and it has
2:14:05memory.md, which is the main one. So,
2:14:07feedback tone matching means this is the
2:14:09explanation of the actual tone of the
2:14:12email that we need to use. So that we
2:14:13don't have to re-explain to it again to
2:14:15use exclamation mark, to use X, Y, and
2:14:17Z. This list of tone for emails is going
2:14:21to be used again and again when we're
2:14:22making new emails, right? And then we
2:14:24have memory.md, which has a memory
2:14:25index, which means that every single
2:14:27time that we give feedback to the actual
2:14:29agent itself, it's going to store the
2:14:30feedback here. And so we'll have more
2:14:32and more feedback so that it remembers
2:14:34the previous things and it gets better
2:14:35over time. Self-improving sub-agents.
2:14:37And now that we have the memory in the
2:14:38prompt itself, I can go here and I can
2:14:40say I can reference uh the email
2:14:42writer.md. Is it this? Yep. And then I
2:14:45can say, "Can [snorts] you send an email
2:14:48to Jack saying that we have football
2:14:50tomorrow 5:00 p.m. Bring the boys over
2:14:52as well so all of them can have a good
2:14:53time." And uh yeah, make it exciting. So
2:14:55as you can see now, as it's running,
2:14:57it's going to read the memory.md5 and
2:14:59it's going to do that every single time
2:15:00because it's with this specific agent
2:15:02itself. The more memories you have, so
2:15:03the more self-improving the system will
2:15:05become over time, right? And the more
2:15:06effective and yeah, accurate the results
2:15:08are when it comes to doing the exact
2:15:10task. It's also looking at the feedback
2:15:11tone matching.md. And then we have the
2:15:13actual email itself with exclamation
2:15:15mark here, here, here. Kind of looking
2:15:18at the tone matching here as well.
2:15:22All right. So you can now spin up
2:15:24sub-agents to handle specific tasks in
2:15:26parallel, working at the same time. But
2:15:27what happens when you need multiple
2:15:29agents working together on a bigger
2:15:31project? That'd be coordinating, sharing
2:15:33context, dividing the tasks, right?
2:15:35That's exactly what agent teams is all
2:15:37about. And the way that I like to think
2:15:38about it is that agent teams is like
2:15:40sub-agents on steroids. So, let's dive
2:15:43in. All right. So here I broke down
2:15:44Claude chat versus sub-agents versus
2:15:46agent teams. What they all are and how
2:15:48they're different. All right. So the
2:15:49first one here is Claude chat. This is
2:15:50you talking to Claude directly with no
2:15:52tools, no code, just a conversation,
2:15:53simple back and forth, and nothing in
2:15:55between. And typically you use Claude
2:15:56chat only when you have a question or
2:15:58want some help writing something, just a
2:15:59normal conversation. Maybe you can write
2:16:01your emails. An example of this could be
2:16:03that you're building automation in,
2:16:04let's say, an A10, you have an error,
2:16:06you drop the error message into Claude
2:16:08code and say, "Hey, can you explain this
2:16:09error message?" Or, can you help me
2:16:11write my cover letter, brainstorm video
2:16:12ideas. Basically, the way the 99% of
2:16:14people are using AI right now, it's
2:16:16using Claude chat. Then we have
2:16:17sub-agents. And sub-agents just [music]
2:16:19means that you have a boss, and the boss
2:16:20has three employees, but this is a bot.
2:16:23The employees cannot talk to each other.
2:16:24Okay, so if employee A is making
2:16:26something that then depends on employee
2:16:28B, well, that's a problem because they
2:16:30can't speak to each other and they can't
2:16:32communicate, right, when making a task
2:16:33together. Which could decrease the
2:16:35quality of output, and it just wouldn't
2:16:37work well. The results can only return
2:16:38to parent agent only, and they have a
2:16:40shared context window. So, they have a
2:16:41shared context window, which means that
2:16:43they all operate under the same window,
2:16:44the same kind of tab in Claude code. And
2:16:46when do we use sub-agents? Well,
2:16:47sub-agents are used when you're building
2:16:48code, or you need Claude to research,
2:16:50explore, or [music] plan different
2:16:51things in parallel at the same time.
2:16:54This could be authentication, write
2:16:55test, right, where you have two
2:16:57different things which are not dependent
2:16:58on each other running at the same time,
2:17:00or explore this code base for me, plan
2:17:02this feature then build it. And when
2:17:03you're building with sub-agents, you
2:17:05ideally want to have two different
2:17:06tasks, or three, or four, or five,
2:17:07whatever it is that you want, uh but
2:17:09make sure that these tasks do not
2:17:11correlate with each other. So, for
2:17:12example, let's say you have a YouTube
2:17:13transcript, and you want to repurpose
2:17:15that YouTube transcript into a LinkedIn
2:17:16post, a blog post, and maybe a Twitter
2:17:18post, right? All these three things, you
2:17:20might think they're not dependent on
2:17:22each other, but they are because the
2:17:24style of writing, the tone of voice, has
2:17:26to be the same across the different
2:17:28content that you make. And so, with
2:17:30sub-agents, you're not able to have the
2:17:32LinkedIn post talk to the blog post,
2:17:33talk to the Twitter post to make sure
2:17:35that they all have the same exact tone
2:17:37of voice. And then we have the one and
2:17:38only agent teams, okay? So, the way the
2:17:41agent teams works is that you have a
2:17:43team lead. So, you give team lead a
2:17:44task, say, "Hey, can you make [music]
2:17:46this team of people that can do X, Y,
2:17:48and Z, right?" What this does is that it
2:17:50spawns a team, so it creates a team, and
2:17:52assigns tasks across the different
2:17:55teammates. [music] And just like
2:17:56sub-agents, it will start creating agent
2:17:58one, agent two, and agent three. And
2:18:00these agents all have the same shared
2:18:02task list, but the agents communicate
2:18:04with each other and each one has [music]
2:18:05its own context. So it has its own
2:18:07context window and operates under its
2:18:09own kind of environment, right? And so
2:18:11you have teammate A talking to teammate
2:18:12B, talking to teammate C, which can all
2:18:14see the same tasks that they have here.
2:18:16And the team leader is the one that kind
2:18:17of operates the whole uh the whole team.
2:18:19And this would be like an actual
2:18:20company, right? all the employees are
2:18:22talking to each other, same departments,
2:18:23maybe different departments, but they
2:18:24can still speak to each other,
2:18:25correlate, and communicate with the head
2:18:27of department kind [music] of in that in
2:18:29that role. Now, you only use agent teams
2:18:31really when you need multiple cloud
2:18:32working on separate repos or separate
2:18:34kind of instances, or when you have a
2:18:35big tasks, right? Usually it's
2:18:37development based, so it's more
2:18:38technical. For example, you can build a
2:18:40front end plus back end plus tests.
2:18:41Front end does depend on the back end
2:18:43and it does depend on the tests. So you
2:18:44want to have two different agents under
2:18:46the same team that are building the
2:18:47exact thing that you want, but bear in
2:18:49mind they can also communicate with each
2:18:50other. So the output will be much
2:18:52higher, right? Because obviously the
2:18:53front end would depend on the back end
2:18:55working. And the test would depend on
2:18:57front end and back end working. So do
2:18:58you want them to communicate with each
2:19:00other at the same time? Or migrate five
2:19:01services at once, but also one cloud
2:19:03orchestrating the rest. Now, in terms of
2:19:05the actual build itself, there is a
2:19:06repeat and improve process, which is a
2:19:08four-step cycle that we go through every
2:19:09single time when it comes to the
2:19:10communication, when it comes to speaking
2:19:12to [music] the agent team for whatever
2:19:13it is that we want to build, right? The
2:19:14first thing is collect. So we want to
2:19:16brainstorm ideas, and it's going to list
2:19:17every tasks and set clear goals. Just
2:19:20like a head of department would or just
2:19:21like a CEO would. Then we organize. So
2:19:23we rank by priorities, we set deadlines,
2:19:25and we block the time slots for the
2:19:27agents itself. Then we execute. So each
2:19:28agent will work through the tasks, will
2:19:30track the progress, and will have focus
2:19:32on what we have to do cuz each agent has
2:19:34its own task that it has to do. And in
2:19:35review, which is checking what's done,
2:19:37um not what's left, and adjusting the
2:19:38plan in case we need to change, all
2:19:40right? And then we repeat and improve,
2:19:41and we keep going on and on and on. And
2:19:43that's exactly how the team actually
2:19:44operates, just like you would for a real
2:19:46company. And one more thing before we
2:19:47get to the build is that this is great
2:19:49because we have what we call
2:19:51parallelization advantage. Because a
2:19:52traditional workflow would have six
2:19:54tasks, and let's say they all take one
2:19:56minute, right? So you have 6 minutes
2:19:57total. When you have an agent team, we
2:19:59have what we call parallel processing.
2:20:01So instead of each task being done
2:20:02individually at different times, here
2:20:05you'd have three different agents all
2:20:06working at the same time on their own
2:20:08tasks. Task 1, task 3, and task 5 will
2:20:10be done at the exact same time, as well
2:20:12as task 2, task 4, and task 6, which
2:20:14means that instead of spending 6
2:20:15minutes, or waiting 6 minutes, then you
2:20:16will wait 2 minutes because we have what
2:20:18we call parallel execution. [music]
2:20:20Now, these are all fancy terms just to
2:20:21say that this basically saves us a lot
2:20:23of time. I'm speaking in in like
2:20:24minutes, but sometimes it can be hours
2:20:26that you'd wait for a workflow to to
2:20:28work, and here you would literally get
2:20:30it within minutes, which is kind of
2:20:31mind-blowing, and I think that's where
2:20:32the future of agent workflows are going
2:20:33to be. It's having one main agent which
2:20:36controls all the others, and having them
2:20:38[music] do the task for you without you
2:20:40having to be there. Okay, so here I am
2:20:41in Antigravity, our IDE that we're going
2:20:43to use. And by the way, Antigravity VS
2:20:45Code basically the exact same. You can
2:20:47pretty much use VS Code as well, but I
2:20:49just chose Antigravity [music] cuz
2:20:50that's what I have here. First thing I
2:20:51want to do here is press open folder. So
2:20:53I typically like to have a Claude Code
2:20:54main folder, and then have different
2:20:56folders inside. Just something that I
2:20:57do. And I can press new folder here,
2:21:00and name this agent, so our Claude agent
2:21:03teams. And I can press open.
2:21:05I already downloaded the extension, but
2:21:06you can go here to extensions. You can
2:21:08look for Claude
2:21:10Code.
2:21:11Make sure it's the one by Anthropic,
2:21:12which should be the one here, Anthropic.
2:21:15And then you can install it, and then
2:21:17you'll be able to see a little button
2:21:18here which has Claude Code open. You can
2:21:20press this button.
2:21:21X out of here, X out of here. Let me go
2:21:23back to here. Okay, so here I am on the
2:21:24Claude Code documents. This is the
2:21:26official Claude Code documents which
2:21:27tells us how to basically start the
2:21:29team. I found this by going to
2:21:30code.claude.com/docs/en/agent/teams,
2:21:34[music]
2:21:35and you'll see run agent teams right
2:21:37here. And if I scroll down, we can see
2:21:39when we use it, when we don't use it,
2:21:40how it works. And then here it says to
2:21:42enable agent teams. So to enable agent
2:21:44teams, we have to make [music] a file
2:21:45called settings.json, and then we have
2:21:47to put this right here inside. I'm going
2:21:49to go back here. I'm going to press
2:21:51[music] new file.
2:21:52I'm going to name this setting.
2:21:54Is it setting or settings?
2:21:56settings.json settings.json
2:21:59And you'll see that this will now be
2:22:00turned into JSON because that's the data
2:22:02structure. [music] And if I go back, I
2:22:03can copy this and then copy it back.
2:22:07And I can press file, save.
2:22:09What you could have also done is you
2:22:11could just actually copy this or you can
2:22:12just you go command A, copy the whole
2:22:14thing and then you can paste it into
2:22:16Cloud Code and say, "Hey, based on this
2:22:18documentation, can you please enable
2:22:20agent teams?" And it will do it for you.
2:22:21But I like to be different, so
2:22:23there you have it. Okay, so
2:22:23settings.json, we have this here. And
2:22:25now, this is where we can prompt Cloud
2:22:28Code to be able to make the agent. All
2:22:29right, so I already have my prompt here
2:22:30that I'm going to use. Uh you can find
2:22:32it in the second link down below in my
2:22:33school community. So what this prompt
2:22:34says is create a multi-agent team that
2:22:36turns sales discovery call transcript
2:22:38into useful sales assets. When I drop
2:22:40the transcript into the {slash} input,
2:22:42the system to generate the client
2:22:44summary, the product proposal, the deal
2:22:46updates, and follow-up emails. So we'll
2:22:47have four different agents. The first
2:22:49agent is called Insights Agent. So we
2:22:51analyze the transcript and extract the
2:22:52client company summary, core problem
2:22:54pain points, desired outcomes, timeline,
2:22:56and the output will be a client
2:22:57summary.md, which will be a file, which
2:23:00will give us a summary. Then we have
2:23:01[music] agent two, which will be a
2:23:02proposal builder agent. So this will
2:23:03generate a proposal. And then we have
2:23:05CRM agent, which will create a
2:23:07structured JSON CRM update including
2:23:09company, contact, main problem, and a
2:23:10bunch more other things. And here we'll
2:23:12have the output, which will be deal up
2:23:14to JSON for the CRM. And then we have
2:23:16the follow-up email agent, which will
2:23:17create a follow-up email. So basically,
2:23:19what I'm telling the agent to do is I'm
2:23:21going to drop a sales discovery call
2:23:23transcript. So let's say I had a call
2:23:24with the client, I drop the transcript
2:23:26of the call, so what we said on the
2:23:27call. And then the team will be
2:23:28consisted of four different agents, one,
2:23:30two, three, and four. So this agent will
2:23:32make the client summary document. This
2:23:34agent will make the proposal. This agent
2:23:35will make the CRM kind of update. And
2:23:37this agent will make a follow-up email.
2:23:38With that said, I'm going to copy this.
2:23:40I'm going to go back here and I'm
2:23:42actually not going to be using this
2:23:43chat, but I'm going to be using the
2:23:45terminal.
2:23:47Now, I was a bit confused when I saw
2:23:49people using the terminal versus using
2:23:50the chat. The reason why we use the
2:23:52terminal here is because we can see
2:23:53more. So, if I press cloud here, uh my
2:23:55cloud will open up. Like this tells us
2:23:57everything. This shows all the settings,
2:23:58it shows all the small things. It might
2:24:00look a bit more complex, but it's very,
2:24:01very easy when you start using it and
2:24:03you you get the hang of it. I'm going to
2:24:04press X.
2:24:06I'm going to make this full screen. And
2:24:07we can use the terminal. And by the way,
2:24:09you can use the other one as well. It's
2:24:10just that I can't see all the features
2:24:11that I'm going to show you. Um but it
2:24:12still works the exact same. I'm going to
2:24:14go here, copy this,
2:24:15and then paste the prompt here. [music]
2:24:16I'm going to press enter.
2:24:18And now Clockwork is going to start
2:24:19working on this. This tool loaded.
2:24:22We have this settings.json, which means
2:24:23that we are enabling agent teams, and so
2:24:25it knows that it has to make a team of
2:24:27agents. It's now saying, "Hey, we have
2:24:28to create the input and output directory
2:24:30structure." I can just say yes, proceed.
2:24:32[music] I could also say yes and don't
2:24:33ask again, but I want to show you all
2:24:35the different uh things that it's asking
2:24:37me to do. Which is fine, so so I can
2:24:38press yes. As you can see, it made the
2:24:40output and the input. The input is where
2:24:42we have the transcript, so we'll be
2:24:43adding the files here, and the output is
2:24:45where we get the CRM, the emails, the
2:24:47insights, and the proposals as well for
2:24:49each transcript that we use. Now, it's
2:24:51saying that it's going to create a
2:24:52sample transcript and the orchestration
2:24:53script in parallel. So, it's also
2:24:55testing that it does work. Now, it's
2:24:56asking whether we can make a
2:24:58transcript.md document to basically test
2:25:01a transcript of a call and make sure
2:25:03that all the agents work simultaneously.
2:25:04I can press yes.
2:25:06This is now writing the transcript.md,
2:25:07which I believe it's going to be here.
2:25:08Yeah, the input cuz that's the input. Uh
2:25:10and then it's saying, "Do you want to
2:25:11make this edit to runpipeline.sh?" I can
2:25:14just press yes. Now, it's asking me all
2:25:15the different small things that I
2:25:17typically would not want to be asked uh
2:25:19cuz it's all code. Now, I can go to yes
2:25:21and don't ask again. Okay, now it's also
2:25:22making the folder called agents.
2:25:25So, all the different agents that we
2:25:26have, the four different agents, will
2:25:27all have an MD document. So, MD just
2:25:29stands for markdown. Um it's the way
2:25:31that they basically put titles and
2:25:32subheadings and bullet points and so on
2:25:34[music] to structure the information.
2:25:36And so, each agent will have an MD
2:25:37document. In the MD document, it's
2:25:39simply just instructions of what the
2:25:41[music] agent is. So, then when we give
2:25:42it a transcript, it basically goes back
2:25:44to the MD documents and says, "Okay,
2:25:46what does this agent have to do? What
2:25:47about this agent? What about that agent
2:25:49and the last agent as well?"
2:25:50>> [music]
2:25:50>> And it will start running all of them at
2:25:51the same time. Okay, there we go. So,
2:25:52this is the structure we have dot cloud
2:25:53slash agent. Then we have agent one. Let
2:25:57me just take this out. Here, we have
2:25:59agent two. We have agent three. And then
2:26:01we have agent four to make the
2:26:02proposals. Okay, so I'm pulling back up
2:26:04the terminal. I can actually just
2:26:07clear this conversation.
2:26:10Clear. I can start from zero. So, now
2:26:11what we'll do is we'll actually test it.
2:26:13So, we have everything done. We have the
2:26:14cloud agents set up. We have the input
2:26:17and the output folders. We have the
2:26:18script and then we have the the thing
2:26:20that allows us to be able to make the
2:26:21team. So, I can just go back here. I can
2:26:23then go to this example discovery call.
2:26:26This is an example like example input
2:26:28that I would give the actual agent. And
2:26:30I can say, "Hey, can you use this
2:26:32discovery call transcript to be able to
2:26:34activate the team and yeah, activate the
2:26:36whole team agent team in cloud code and
2:26:38use this transcript as the input." Not a
2:26:40question. Paste it here.
2:26:42And now we press run. So, this basically
2:26:44pasted the whole the whole call that we
2:26:45have. It's a short call, so nothing
2:26:47crazy, but it's enough for it to be able
2:26:50to make all the things that we wanted to
2:26:51make. So, it's saying, "I can see the
2:26:52full agent team pipeline. Let me first
2:26:54update the transcript file with the
2:26:55Collins marketing discovery call and
2:26:56then run the pipeline." Pipeline just
2:26:58means run the agent team. Now it's
2:26:59saying, "Hey, can we replace the
2:27:00transcript MD, which is the one here,
2:27:02with the transcript that you just gave
2:27:03us?" I'll just say yes, allow all edits
2:27:05during the session just because I don't
2:27:06have to then approve everything. So, it
2:27:08uploaded here and now it's saying, "Hey,
2:27:10can we run the pipeline?" I can just
2:27:12press yes and don't ask again.
2:27:14Now it's saying, "Hey, let me kick off
2:27:15agent one, call insights agent, first
2:27:17since the other three depend on its
2:27:18output." And this is great because you
2:27:19get to see exactly how agent one now
2:27:21talks to the other agents. Agent one is
2:27:23a call insights and based on the call
2:27:25insights, we then feed it to the other
2:27:26agent to make the email and the proposal
2:27:28and the CRM as well. Cuz now it's
2:27:29telling me that it made the client
2:27:30summary, company overview, core problem,
2:27:33main pain points, desired outcome,
2:27:34timeline, red flags, objections and
2:27:36concerns and key quotes. I can say, yes,
2:27:39I want to proceed, and now I should get
2:27:40the insights here, client summary.md.
2:27:42And by the way, I'm doing this obviously
2:27:43manually to show you how to test it and
2:27:45how everything works, but you can pretty
2:27:46much upload this or publish this to a
2:27:49platform like Modal, and then all you
2:27:50have to give it is the input, and then
2:27:52the whole agent team will work on its
2:27:54own. And one thing I'll also cover it
2:27:55later on is the pricing, cuz you can see
2:27:57I'm getting completely rinsed with my
2:27:59tokens. It's going faster than I've ever
2:28:02seen. So, this is one drawback of agent
2:28:04teams, but we'll get into that in the
2:28:05later part. So, agent one was complete.
2:28:07It then ran agent two three at the same
2:28:09time. So, two agents finished. It built
2:28:11the product proposal from insights and
2:28:12the CRM deal for Jason from insights.
2:28:14[music]
2:28:14And now that agent two and three are
2:28:16done, now we're going to be launching
2:28:17agent four, which is going to be
2:28:19creating the follow-up email. So, I'm
2:28:20saying now that all four agents have
2:28:21been completed,
2:28:22>> [music]
2:28:22>> here's the summary of what was
2:28:23generated. So, we have the call insights
2:28:25agent, proposal, CRM, and follow-up
2:28:27emails. We have the files, which you can
2:28:29find emails here, CRM here. We have
2:28:32insights here, and we have proposal
2:28:33here, which I'll go into in just a
2:28:35second. And it says, all outputs files
2:28:36are ready in the output directory. Want
2:28:38me to open them for review? I can just
2:28:40say, yeah, let's see what it does,
2:28:42really.
2:28:43Yes, open them.
2:28:45And by the way, you can also edit the
2:28:46permissions. So, right now, if you press
2:28:48shift tab, shift tab, you can see that
2:28:51I'm I'm changing this. Plan mode accepts
2:28:53edits. You can also do bypass
2:28:55permissions, which is letting it just do
2:28:57everything on its own. So, client
2:28:58summary, this is a document, product
2:28:59proposal, CRM deal update, and follow-up
2:29:01emails. So, I'm going to make this
2:29:03smaller so you can see. All right, and
2:29:05I'm going to go through each one. So,
2:29:06CRM,
2:29:07this is the deal update that you upload
2:29:09in your CRM to be able to include all
2:29:11the fields. So, this included the
2:29:12company name, industry, size indicator,
2:29:14growth rate, contact name, title, email.
2:29:16There was no email, so we couldn't find
2:29:17it. The deal stage, discovery summary,
2:29:19main problems, pain points, desired
2:29:21outcomes, and a bunch more other things
2:29:22as well.
2:29:23Then we have the emails. So, this is the
2:29:25follow-up email that it made, and it
2:29:26said, to Sarah Collins from Ren Gym
2:29:29Solutions. Subject line, your road map
2:29:31from 18 to 40 clients, proposal inside.
2:29:33Thank you for the candid conversation
2:29:34Wednesday. The way you described the gap
2:29:36between where Collins marketing is today
2:29:38and where you want it to be, 40 clients
2:29:39without the operations breaking, gave us
2:29:41a clear picture of what needs to happen.
2:29:43A few takeaways I want to make sure we
2:29:44capture correctly. Onboarding is a
2:29:45bigger time drain. I now have a client
2:29:47with manual folder creation, document
2:29:49sends, and tool setups [music] across
2:29:50Google Drive, ClickUp, and Slack, and
2:29:51things still slip through the cracks.
2:29:53Reporting is costing your VA 3 to 4
2:29:55hours per week. The knowledge bottleneck
2:29:56is real. I've attached a proposal which
2:29:58outlines a two-phase 8-week engagement
2:30:00and [music] talk soon, Ron. We also ask
2:30:02you to book a call. Now, if I go to
2:30:04client summary, I can see the insights.
2:30:06So, this is the summary of the actual
2:30:08conversation that we had on the call.
2:30:09And in proposal, we have the proposal
2:30:11that we have that it just made based on
2:30:12the conversations. Understanding your
2:30:14challenges, manual client onboarding,
2:30:15reporting, tools, knowledge, [music]
2:30:18proposal solution, and we have scope of
2:30:20work, deliverables, timeline,
2:30:22investment, why us, next steps, and so
2:30:25on. And these are all things that you
2:30:26could easily turn into like a PDF. So,
2:30:28let me go here to Clockwork. And I'm
2:30:30going to say, "Can you turn the project
2:30:32proposal.md file into a PDF that I can
2:30:35use? Make sure the formatting is
2:30:36correct." Obviously, if you look at
2:30:38this, you'd be like, "Okay, how do we
2:30:39use this now?" You could easily copy
2:30:41that into Notion or copy that into
2:30:43Google Docs, but I just want to show you
2:30:45how Clockwork does it for me. All right,
2:30:46so I just finished making the product
2:30:47proposal PDF. So, we can go here. This
2:30:50is the markdown format, and it turned it
2:30:52into a PDF. Of course, if you look at it
2:30:53this way, it also wouldn't make sense.
2:30:55>> [music]
2:30:55>> But, because it's a file on our
2:30:56computer, I'm just going to open it up,
2:30:58and we get this right here, which is PDF
2:31:01version with the executive summary,
2:31:02understanding your challenges. [music]
2:31:03Then we have tools, proposal solution,
2:31:05scope of work, deliverables, timeline,
2:31:07investments. It also made the table,
2:31:09very nice formatting, the price, why us,
2:31:11and next steps.
2:31:12>> [music]
2:31:12>> And so, that's exactly how the whole
2:31:13Agent Team ran through to make the
2:31:15output, which you can then use for
2:31:17whatever it is. And this right here is a
2:31:18seven out of 10 use case, I'd say,
2:31:20because you could still do with sub
2:31:21agents in some way,
2:31:22>> [music]
2:31:22>> but the output is better here just
2:31:23because you have the client summary,
2:31:25which is made, and then passes the
2:31:27client summary to the proposal and to
2:31:29make the [music] email and the CRM to
2:31:31then give you the full output. All
2:31:32right, so there's one aspect of this
2:31:33that I briefly mentioned, but I didn't
2:31:35really dive deep into, which is price.
2:31:37Cuz if you saw before, for me to run all
2:31:39the four agents, it cost me 10,000
2:31:41tokens, and it was increasing way too
2:31:44fast for my liking. And so here we have
2:31:46token cost and here we have speed.
2:31:47Obviously, regular cloud will take a lot
2:31:49more time, it'll be cheaper, the faster
2:31:50you get and it more expensive it will
2:31:52get until you get to agent teams, which
2:31:54is 5 minutes and 80 bucks. This is
2:31:56because
2:31:57>> [music]
2:31:57>> each employee in the agent team has its
2:31:59own context. It has its own kind of
2:32:01place where it runs. And so it will cost
2:32:04you a lot more to be able to run this
2:32:05agent team. But if you are building
2:32:07something a bit more complex, well,
2:32:09let's say you're building a software
2:32:10application. And for the software
2:32:11application, you want the main, you
2:32:14know, team leader, and then you want the
2:32:15different agents to be able to talk to
2:32:17each other. Cuz when you're making the
2:32:18front end, the back end, the privacy,
2:32:20the the OAuth, all these things, you'd
2:32:22want the different agents to communicate
2:32:25to build a much better output rather
2:32:27than using sub agents. Now, for the most
2:32:29part, if you're working with Clock code
2:32:30and you want to run different tasks in
2:32:32parallel, you're better off going with
2:32:34sub agents rather than spending six
2:32:36times more, but also getting six times
2:32:38the speed.
2:32:41All right, so you now know how to build
2:32:42sub agents and agent teams inside of
2:32:45Clock code. But after having built
2:32:46hundreds of agent workflows, I've
2:32:48discovered specific techniques, hacks,
2:32:51really, that make your Clock code
2:32:52dramatically and exponentially better.
2:32:54And I've divided these hacks into eight
2:32:56different hacks, which is why in this
2:32:58video I'm going to break them all down
2:33:00step by step. With that being said,
2:33:01let's dive in. All right, so we're going
2:33:02to divide these hacks into three
2:33:04different types of hacks. We have system
2:33:05design hacks, then we have execution and
2:33:07performance hacks, and then we have
2:33:09control and optimization hacks. All
2:33:11right, so the first hack is actually
2:33:12setting up the folders for your project.
2:33:14All right, so here we have Clock code,
2:33:16and the mistake that people make when
2:33:17starting out to build is that they go
2:33:18straight to building, right? But the
2:33:20problem is that right here on the
2:33:22left-hand side where we store all the
2:33:23folders and the files and all these
2:33:24things, we have to structure clock code
2:33:26in a way where it knows where to store
2:33:28all the different folders and the files,
2:33:30which is where we go to the three layer
2:33:31agent architecture, which is a way for
2:33:34us to be able to structure the AI agent
2:33:36to be way more reliable. The way that we
2:33:37do this is dropping a file inside of
2:33:39clock code, which is called the cloud.md
2:33:42file. This right here I can think of it
2:33:43as a blueprint that you get to clock
2:33:45code and contains a three layer
2:33:47architecture, a three layer structure
2:33:49that clock code needs to use to be able
2:33:51to then build these things and store all
2:33:53the files [music] in the same place. The
2:33:55first layer is the directives, so this
2:33:56is what to do. So we give it SOPs, we
2:33:58define objectives, we list the inputs
2:34:00and the outputs, we specify which tools
2:34:02do we have to use, and we cover all the
2:34:04edge cases of the workflow. It's like a
2:34:05brief that you give to someone working
2:34:06in your company. Then we have layer two,
2:34:09which is the orchestration, and this is
2:34:10for the decisions. So the actual LLM
2:34:13lives inside this folder. So this is
2:34:15good when giving it instructions as to
2:34:17where to store the right scripts, right?
2:34:19How to handle errors, how to ask
2:34:20questions when it's stuck. It's
2:34:21self-healing, right? That's what we call
2:34:23clock code because when something
2:34:24breaks, it knows exactly what's broken,
2:34:26and then it goes back and fixes it so
2:34:28you don't have to. And this is the glue
2:34:29between what you want to build and
2:34:31what's actually built. And then layer
2:34:33three is the execution. This is doing
2:34:35the work. So this includes the Python
2:34:36scripts, the API calls, all the data
2:34:38that's going within the automation, all
2:34:40the files, all the databases, all the
2:34:41techy stuff that it needs to do to
2:34:43actually do the thing. And this is all
2:34:44code, right? Of course, clock code is
2:34:46just building code um that we use. Now
2:34:48instead of doing this, why don't we let
2:34:49the LLM do everything and structure
2:34:51everything? It's because LLMs are about
2:34:5490% accurate per step, right? It sounds
2:34:56good until you realize that when you put
2:34:58step over step, you have 0.9 to the
2:35:00power five, which is 59% success rate,
2:35:02[music] and it's simply not good enough
2:35:03at all if you have anything below a 99%
2:35:06rate, which is simply not good enough
2:35:08when we're building agent workflows. And
2:35:09so the solution is to give it more
2:35:10frameworks, more guardrails in clock
2:35:12code to be able for it to understand how
2:35:14to structure things when it's building,
2:35:16okay? Now if that didn't make too much
2:35:17sense for you, I'll show you what it
2:35:18means. So, right here I have the
2:35:20cloud.md file that I just mentioned. I'm
2:35:22going to copy this. And by the way, if
2:35:24you want access to this full prompt,
2:35:25second link down below, my community,
2:35:27you can find it in the classroom
2:35:28section. So, we go here and we can make
2:35:29a file and name this cloud.md.
2:35:33And then you can paste this whole
2:35:34prompt, right? Which again, is just a
2:35:36verbal explanation of what I just said.
2:35:38So, layer one, layer two, and layer
2:35:39three. And then you want to press file
2:35:41and save.
2:35:43And then you can say, "Initialize this
2:35:45project based on the cloud.md file that
2:35:47I just put in the project or in the
2:35:49files." There you go. And I put plan
2:35:51mode and then I press run. Okay, so now
2:35:53what it's going to do is it's going to
2:35:55go inside the cloud.md file, it's going
2:35:57to read it, and then it's going to start
2:35:58structuring all the different folders of
2:36:00what it needs to do. All right, so it
2:36:01just made a plan here. Plan to
2:36:02initialize product structure from
2:36:04cloud.md. The context, the changes. So,
2:36:06if you see this right here, don't get
2:36:08too intimidated. What this means is the
2:36:10structure of all the folders here. All
2:36:12right, so this is the project root. We
2:36:13have directives, we have execution, we
2:36:15have .tmp, and we have .env, and then we
2:36:17have .gitignore, right? These are all
2:36:19different things that are part of the
2:36:20actual structure that it will need to
2:36:22do. You can just press yes and auto
2:36:23accept. [music]
2:36:24Now we'll have tasks. And then on the
2:36:25left-hand side here, you can see that
2:36:26it's adding the different files and
2:36:27folders that it's going to use when
2:36:29we're going to be able to make the
2:36:30actual workflow to make things more
2:36:31structured. Okay, there we go. It just
2:36:32finished. So, all the folders are here.
2:36:34We have directives, we have readme.md,
2:36:36which is a basically a file that
2:36:38explains exactly what this folder is.
2:36:40So, this folder contains markdown SOPs,
2:36:41standard operating procedures that
2:36:43define objectives, inputs, tools and
2:36:44scripts, outputs, and edge cases. And in
2:36:46the easiest way possible, in this way,
2:36:48it's like an architect going to build a
2:36:50house, right? Either he goes to build a
2:36:51house and starts building immediately as
2:36:53soon as he goes on site, or what he does
2:36:54is that he sits down, he basically makes
2:36:56a road map or a blueprint as to how he's
2:36:58going to build the house, and then we
2:36:59get to the building part, which is
2:37:01exactly what this is. This is the
2:37:03structure and the blueprint that we're
2:37:04going to use, well, the clock code is
2:37:06going to use, when we have to make all
2:37:08these things. To give you a practical
2:37:09use case, I I built this YouTube
2:37:11strategy system for myself, [music] and
2:37:13I made a full video as well. You can
2:37:14check it out here. But essentially, we
2:37:16have the cloud.md file that we dropped,
2:37:18[music] which is the exact same one that
2:37:19we did for the other one. And then, it
2:37:20started making all the folders. And if
2:37:22you go to directives, we get to see all
2:37:24the different kind of SOPs, right?
2:37:26Explaining the different automations. So
2:37:28for example, this right here, analytics
2:37:30tracking. So it's an MD file, so it's
2:37:31markdown file. Markdown because we have
2:37:33hashtags. It's just a way of writing. So
2:37:35AI or computers can understand better.
2:37:37But it's a markdown file that tells the
2:37:39actual cloud code exactly what to do,
2:37:42the SOPs. This is what the objective of
2:37:45this thing is. This is the objective of
2:37:47what it is that we want to do. What is
2:37:48the input and what is the output, right?
2:37:49The input is a channel ID and the output
2:37:51is performance dashboard data. So a
2:37:52dashboard with all the data and all the
2:37:54different things. Now to be fair, it
2:37:56might seem a bit intimidating when
2:37:57you're starting out and you're wondering
2:37:58still wondering why is it like this and
2:38:00why is it like that. But just understand
2:38:01that this is the way that we structured
2:38:02all the different folders and files. The
2:38:03second hack is using cloud code skills.
2:38:06made a full video breaking cloud code
2:38:07skills down in much more detail. But
2:38:09essentially, what they are is
2:38:10instruction manuals that cloud reads
2:38:12before doing a task. So the same way
2:38:14that you give instructions to a chef to
2:38:15mix some sort of food, you give cloud
2:38:17code the skills to then make better
2:38:18PDFs, better websites, better
2:38:20presentations and so on. And so for
2:38:21example, let's say you give cloud code a
2:38:22task to make a website. Cloud code,
2:38:24which is the agent, will look at its own
2:38:25skill that we just added to build
2:38:27effective websites and it will give us a
2:38:28much better result rather than us not
2:38:30giving it a skill, right? Because
2:38:32without a skill, cloud guesses and it's
2:38:34inconsistent in the results. Whilst with
2:38:36a skill, it follows a pattern and it
2:38:37follows instructions and it gives us a
2:38:38much better output. And we can find the
2:38:40skills at ai.tmpl.com.
2:38:42I'm going to leave the link down below
2:38:43in case you want to check it out. And
2:38:45here, we have over 683 skills and
2:38:48there's more added every single day. And
2:38:50these skills, like senior backend web
2:38:52app testing, senior architect, code
2:38:53reviewer, git commit helper, are the
2:38:56attachments or the things that we can
2:38:57add to cloud code for it to be able to
2:38:59do a better job at the thing that it
2:39:00needs to do. So I'm going to be working
2:39:01with this agenda workflow that I just
2:39:03mentioned, which gives me a full
2:39:04presentation of all the competitors, the
2:39:06analytics and everything like that. And
2:39:07so, I'm going to be using this skill
2:39:09right here, which is Canva Design, which
2:39:11allows me to be able to give Claude Code
2:39:13the extra skill of making better
2:39:15presentations and better PDFs and
2:39:16visuals for any kind of thing that we're
2:39:18doing. So, I actually already added the
2:39:19skill. I'm just going to go ahead and
2:39:21delete it. I'm going to delete the
2:39:22folder as well, cuz then you can see the
2:39:24output before and after. I'm going to
2:39:26say, "Can you please make the
2:39:27presentation and give me back the result
2:39:29analyzing my competitors and the
2:39:31analytics and everything like that?" I'm
2:39:32going to press go, and now it's going to
2:39:34start to do what it needs to do. And
2:39:35I'll catch you when it gives me the
2:39:36output. All right, so it just finished
2:39:37making the presentation with no skill.
2:39:39As you can see, it's very plain and
2:39:40simple, no effects, no colors, and it's
2:39:43all right. The actual structure of it is
2:39:45a bit bit off, but nothing crazy, right?
2:39:47And so, [music] ideally we want
2:39:48something that's branded, that's better
2:39:51in terms of structure. I feel like I
2:39:52want some effects here, I want some
2:39:54colors to make it easier on the eyes,
2:39:56but also to make it appear better when
2:39:57I'm using it with my team and so on. So,
2:39:59ideally we want something that's a bit
2:40:00more professional, that has colors,
2:40:02structure, and the layout looks a bit
2:40:04better. What we're going to do here is
2:40:05go back to the Claude Code templates.
2:40:07I'm going to scroll down to the Canva
2:40:08Design. And by the way, if you're
2:40:09looking for a PDF or for an MCP Builder,
2:40:12which I use all the time, and for other
2:40:13things, you can just simply search it
2:40:15here and you'll find it. You can, for
2:40:16example, we can look for a website, web
2:40:19design reviewer, Google Analytics, SEO
2:40:20Optimizer, Web Experience, a lot more
2:40:22things that are there. So, I can look
2:40:24for Canva
2:40:25Design.
2:40:26Canva Design. And if I go to View
2:40:28Details,
2:40:29I will get to this page right here. Find
2:40:31out that the skill is actually a prompt.
2:40:33So, every skill is a prompt, and the
2:40:35prompt is very detailed, which allows
2:40:36Claude Code to be able to build better
2:40:38presentations based on this little
2:40:40prompt that we have here, which someone
2:40:42made. Now, I would recommend looking at
2:40:43the amount of downloads because if it
2:40:44has good downloads, then obviously it's
2:40:46good. Not all the skills are great, but
2:40:47it depends on the prompt, and prompting
2:40:49depends on the person. So, all we have
2:40:50to do here is copy this code. I can go
2:40:52back to Claude Code. I can clear
2:40:54conversation, and I can say, "Install
2:40:57this skill here."
2:40:59And I can press go. And now, what this
2:41:02will do is it will start adding the
2:41:04skill inside of Claude here. We just
2:41:05have to prove a few things. As you can
2:41:07see it's saying that the canvas design
2:41:08skill has been installed successfully. I
2:41:10can go to Claude, I can go to skills,
2:41:12and here we have canvas design. And if I
2:41:14go to skill.md here, we have the full
2:41:16prompt that we had here on the code.
2:41:18Let's run it again. So let me go to
2:41:20clear conversation. I can say, "Okay,
2:41:22can you now make the presentation for
2:41:23the analytics comparing me and my
2:41:25results and my competitors and give me
2:41:27the back to full presentation on a
2:41:29statue." I'll press play mode.
2:41:30It will now start working on the
2:41:31presentation. All right, and after a few
2:41:33minutes it made this full presentation
2:41:34right here with colors and designs, and
2:41:37it looks a lot more cleaner than the one
2:41:39that we had before. We have colors, we
2:41:41have different types of colors, we have
2:41:42a better structure now, better way that
2:41:43we can look at things. We have headers,
2:41:45which is a lot cleaner than the one we
2:41:46had before. I can see there's a few
2:41:48mistakes though on the top as well, but
2:41:49it's something that we can fix just by
2:41:51prompting it uh to do a better job in
2:41:53those specific areas. And we have until
2:41:55the end. Now Claude's skills are by far
2:41:57one of the most important things that
2:41:59you need to know when it comes to Claude
2:42:00code because Claude code is great, but
2:42:02it's even better when you add
2:42:03enhancements. It's sort of like the Iron
2:42:05Man suit, which is amazing, but it's
2:42:06even better if it has the controller, if
2:42:08it has the laser, if it has all these
2:42:09things that make it 10 times more
2:42:10powerful, and it makes you build things
2:42:12at 10 times the speed instead of you
2:42:13having to go back and forth and
2:42:15prompting it saying, "Hey, it's wrong.
2:42:16It's not wrong. Make it better." and
2:42:18things like that. The next hack is using
2:42:19the spec to do and code workflow. Okay,
2:42:22so now that we added all the structure
2:42:23of the files, do we go straight to
2:42:24building? Well, no. We have to plan even
2:42:26more if you want to build something
2:42:28that's actually scalable and that's more
2:42:29structured because yes, you will spend
2:42:31more time now, but it will save you a
2:42:33lot more time later on when you want it
2:42:35to actually work. And the problem with
2:42:36most people using Claude code is that
2:42:37they paste the prompt like this, which
2:42:39is Claude, "Build me a family activity
2:42:41finder app." AI is great, but but if
2:42:43you're not specific about what you want,
2:42:45then it's not going to do a good job.
2:42:46And so [music] we replace this prompt
2:42:48with something that looks like this,
2:42:50which is asking it to make a spec.md for
2:42:52a family activity finder app. We have
2:42:53the requirements, which is parents enter
2:42:55city, kids' ages, availability, distance
2:42:57willing to travel. The app returns five
2:42:59activities with title description and
2:43:00emojis. The tech stack that we're going
2:43:01to use is simple HTML, CSS, and
2:43:03JavaScript. And we have milestone one,
2:43:05which is building the UI with dummy
2:43:06data, and milestone two, which is
2:43:07connect to real activity API. Now, the
2:43:09reason why we do this is because we want
2:43:11a detailed blueprint with requirements,
2:43:13tech choices, and clear milestones as to
2:43:15what it needs to do to be able to build
2:43:16[music] this thing before we even start
2:43:18building anything. And so, I'm going to
2:43:19press go, and as you can see here, this
2:43:21spec.md is sort of like an SOP. And we
2:43:24told previously the to CloudCode that
2:43:26all the SOPs should be stored under the
2:43:28directives. Now, it's going to ask me to
2:43:29accept this plan. I'm just going to
2:43:31press out just because I don't want it
2:43:32to do anything. And if I go to the
2:43:34directives, I should be able to now to
2:43:35see a spec.md file, which is a full
2:43:37file. Let me zoom out a little bit. It
2:43:39tells us the overview of the exact app
2:43:42that we're going to use, right? So, we
2:43:43have the overview, tech stack, user
2:43:45inputs, output, UI layout, milestone,
2:43:48and we have dummy data examples,
2:43:49milestone two, and so on. This gives a
2:43:51lot more clarity and structure to
2:43:52CloudCode later on when you are building
2:43:54something. Because bear in mind that
2:43:55when you talk to CloudCode, you're
2:43:57talking here. Then when you leave
2:43:58CloudCode and you come back, this will
2:44:00be empty, and so it will need some sort
2:44:02of memory to understand exactly what you
2:44:04were doing in the first place. And so,
2:44:05you always want to make sure that you
2:44:06are structured as possible when building
2:44:08the actual thing. Now, we're done with
2:44:09step one, which is creating a detailed
2:44:11spec. Now, we're going to go to step
2:44:12two, which is generating to-do lists.
2:44:14And so, now I can say, "Create a to-do
2:44:15list for milestone one based on the
2:44:17spec." And we'll be able to go through
2:44:18milestone by milestone to make sure that
2:44:20everything is all right and everything
2:44:21is structured, and it doesn't build a
2:44:23full thing and then we're like, "Hey,
2:44:24man, not the thing that we wanted."
2:44:26[music] It's going to make to-dos uh for
2:44:28that specific milestone. Now, it has
2:44:30made all the to-dos for the first
2:44:31milestone in specific, and I can see
2:44:33here that it says, "Here's the milestone
2:44:34one to-do list. Create project
2:44:36directory, build index.html, it can do
2:44:38XYZ." And it gives us all the steps in
2:44:40specific because we right now are acting
2:44:42as a supervisor for the actual developer
2:44:44that's building the thing. And now I can
2:44:46say, "Build milestone one using the
2:44:47to-do list." I'm not going to press
2:44:49enter cuz I don't want to build
2:44:49anything, but I can tell you that once
2:44:51it finishes milestone one, you can then
2:44:52go and say, "Okay, can you not build a
2:44:54to-do list for milestone two, and
2:44:55milestone three, and milestone four?"
2:44:56until you're finished. And, yes, you've
2:44:58taken a bit more time and to structure
2:45:00things, but you've made the output 10
2:45:02times better just because you've been
2:45:04more structured. Remember that Clawcode
2:45:06is only as good as your input and the
2:45:07prompts that you give it. This is a
2:45:09developer, but a developer is good at
2:45:10coding, but it might not be the best to
2:45:12understand the context behind the code
2:45:14and what it is that you actually want,
2:45:15which is why we want to be as structured
2:45:16and explicit as possible when building
2:45:18these agenda workflows or apps. All
2:45:19right. So, now we get to the execution
2:45:21hacks. The first one being permission
2:45:23mode settings. So, down here when we're
2:45:25talking to Clawcode, we can actually
2:45:26change the mode or the settings of the
2:45:28permissions that we're giving Clawcode.
2:45:29You can either press and change, or you
2:45:30can simply go shift and tab, and it'll
2:45:32change automatically. So, the main ones
2:45:34is bypass permissions, and then we have
2:45:36the plan mode. Bypass permissions is
2:45:38used when you're making a very simple
2:45:39task. This is what we call the YOLO
2:45:40mode, which is telling Clawcode, "Hey,
2:45:42you do everything. Don't ask me any
2:45:43questions. Don't plan. Just like go at
2:45:45it." This is great for something very,
2:45:46very simple. Whereas, plan mode lets you
2:45:47collaborate with Clawcode to explore
2:45:49solutions and create detailed structures
2:45:51without wasting tokens of going back and
2:45:53forth. What we usually say is a minute
2:45:55of planning saves you 10 minutes of
2:45:56building. The next hack is using
2:45:58sub-agents for parallel processing. So,
2:46:00what this means in a very non-techy way
2:46:02for beginners is that when you talk to
2:46:04Clawcode here, you're talking to the
2:46:05main agent. But, what you can also do is
2:46:07talk to the main agent and then have
2:46:08other agents working in parallel as all
2:46:11these ones, which means that you can do
2:46:13four times the work at the same amount
2:46:14of time. All right. So, here I have my
2:46:15partnerships email inbox. Let's say I
2:46:17want to process all these emails
2:46:19together. Of course, I only have about
2:46:2184 right now, but just let's say I had a
2:46:23thousand or two thousand or whatever it
2:46:24is, cuz most people do have that. Then,
2:46:26instead of having one agent trying to
2:46:27process all of them, you can split the
2:46:29work between agents. And so, what you
2:46:31could do is you can split the task
2:46:33across these four agents. So, you can
2:46:34say to this agent, "Hey, can you label
2:46:37emails from 1 to 200? This one from 200
2:46:39to 400? And then 400 to 500?" and so on,
2:46:42right? And so, when you split the task,
2:46:44they all work simultaneously at the same
2:46:45time, and it'll will so much faster.
2:46:47Here's my client ID and client secret.
2:46:49Uh I want to create a email classifier
2:46:51system where it goes through my Gmail
2:46:53and it basically classifies the emails
2:46:55on whether there it's it's a full full
2:46:56caps, an editor request, or software
2:46:58partnership. So, I'm going to give you
2:46:59my client ID and client secret so you
2:47:00can use that and then authorize my
2:47:02Gmail, like ask me to authorize my
2:47:03Gmail, and let you in, and then you can
2:47:04we can start the task. All right, and
2:47:06these four agents just finished running.
2:47:07I can see here that they've all been
2:47:08labeled other a software partnership,
2:47:10other, or uh editor request as well.
2:47:12Now, the example I showed you because
2:47:13it's 80 emails, it wouldn't be
2:47:15realistic. But, if you have thousands
2:47:16and thousands of emails, then it would
2:47:17work really well because you have these
2:47:19agents running at the same time doing
2:47:20the exact same thing. And all you have
2:47:22to do is make sure that you're there to
2:47:24be able to conduct them and answer any
2:47:26questions if you need to answer. All
2:47:27right, the next hacks are going to be
2:47:28the control and optimization hacks. The
2:47:29first one is mastering context
2:47:31management. So, think of Claude's brain
2:47:33like a notebook with only 200,000 pages,
2:47:35right? Every time you chat with it with
2:47:36Claude, it writes down your message, its
2:47:37response, all the files it reads, every
2:47:39tool it uses. As the notebook fills up,
2:47:41Claude gets dumber. It gets more stupid,
2:47:43right? Because it's trying to remember
2:47:45too much at once from a conversation and
2:47:46files. And so, if we keep having a chat
2:47:48with this agent right here, it becomes
2:47:50very sluggish because you've used up
2:47:5250,000 plus tokens just on conversation
2:47:54history. The solution to this is using
2:47:56the smart reset, right? The smart reset
2:47:58is when you put {slash} dot compact,
2:48:01right? And this compresses your
2:48:03conversation into a summary, but it
2:48:05keeps the most important stuff, right?
2:48:06And you use this when your chat is
2:48:07getting way too long, but you want to
2:48:08keep working on the same project. And
2:48:10it's compacting the conversation that we
2:48:11had here. Of course, most of this is
2:48:13just Claude code just working on its
2:48:14own, but let's see how you conversation,
2:48:17it would extract all the summary points
2:48:18of this, and then it would give us the
2:48:20summary, which we'll then use to be able
2:48:22to go to the next steps. As you can see
2:48:23here, it says, "The session has been
2:48:24continued from a previous conversation
2:48:26that ran out of context. The summary
2:48:27below covers the earlier portion of the
2:48:29conversation." And so, this right here
2:48:31summarizes the conversation that we had
2:48:32to then be able to then keep going with
2:48:34the conversation. And once you keep
2:48:35talking to it, now you have used about
2:48:37180,000 tokens free again, but Claude
2:48:40still remembers what you were working
2:48:41on. And the next command is actually
2:48:42something that you've seen me use over
2:48:44time and time again, which is the
2:48:45{slash} clear conversation. Cuz when I
2:48:47want to have a new conversation with
2:48:48Claude code, instead of having to spin
2:48:50up a new agent, I want to have the same
2:48:51conversation here, but the conversation
2:48:53will be cleared and you can start from
2:48:54zero. And one more thing you can do
2:48:56within the context management is check
2:48:57your context usage. So, if you're
2:48:59worried about how much context you're
2:49:00using within a conversation, you can
2:49:02simply press {slash} context here. So, I
2:49:05actually have to say, "Hey, can you give
2:49:06me the context usage?" {slash} context.
2:49:07And here it says, "Hey, here's your
2:49:09current context usage. Model is Claude
2:49:10Opus 4.6. Tokens is 55.5 thousand out of
2:49:13200,000. We're using about 28%. Again,
2:49:16200,000 is the maximum of tokens that
2:49:18you can use within one conversation."
2:49:19And it tells us the system prompt, the
2:49:21system tools, the memory files, the
2:49:22skills, the messages, the free space,
2:49:24basically giving us an overview of how
2:49:25much we can use more than what we're
2:49:27using right now. This context hack is
2:49:28important because it allows you to be
2:49:30able to understand exactly how much more
2:49:32you have left, how much runway you have
2:49:33left, and also how much you've used so
2:49:35far. And the last hack here is using the
2:49:36Pixel Agent extension. So, there are
2:49:38some of you here, including myself when
2:49:40I was starting out, that look at this
2:49:41and think that it is very intimidating
2:49:42and it looks very, very complex. And
2:49:44there is actually a way to view your
2:49:47agent workflows or your Claude workspace
2:49:49in a much more visual way. We can go
2:49:50here to extensions, we can look for
2:49:52Pixel Agents,
2:49:54and then you'll choose this one here.
2:49:56You can press install, and now this will
2:49:58be installed in your workspace. Then
2:49:59we're able to view our workspace in this
2:50:01manner right here. Instead of like
2:50:02different emojis or LEGO blocks that
2:50:04show us all the different agents working
2:50:06simultaneously, which is actually
2:50:07really, really cool. And the way to find
2:50:08it is to go to command shift P, I can
2:50:10look for Pixel Agent show panel. There
2:50:12we go. And on the right-hand side we can
2:50:14see here that we have a visual
2:50:15representation of a workspace. And if I
2:50:16press new agent on the bottom, you can
2:50:18now see that this person just appeared,
2:50:19and he is idle because we haven't given
2:50:21him any task to do. And I can press more
2:50:23and more agents, which will add more and
2:50:24more people here. And these are the
2:50:26different agents that we can use to be
2:50:29able to run up and spin up different
2:50:30things. So, maybe the first agent will
2:50:31be, "Hey, can you research the web on a
2:50:33top AI tools uh out there in 2026?" And
2:50:36then the second agent will be taking
2:50:38research the best workflow automation
2:50:39tool past 10 years. Hey, can you do
2:50:41extensive research on human psychology
2:50:43and tell me the top 10 trends that you
2:50:45see over time, right? And now, using
2:50:47agent needs approval, so I can see
2:50:49exactly when something needs approval,
2:50:50and I can go back to the agent itself,
2:50:52right? And just give them approval on
2:50:53what they need. But all in all, you're
2:50:54not able to visually see your workspace
2:50:56and the agents working at the same time
2:50:58all simultaneously through a very very
2:51:00visual representation. I do love this
2:51:01because it also makes it fun for me to
2:51:03work with Clayco, and I can see
2:51:05different things. And also, in addition
2:51:06to that, I can go to the layout, and I
2:51:08can start moving things around. I can
2:51:10put the plant here, or I can put the
2:51:11boxes here, or I can put the desk here,
2:51:13or the computer, and so on. In case
2:51:15you're bored and you want to change
2:51:16things around for no reason.
2:51:17But you have some you have some options.
2:51:19Uh you can also choose the floor. You
2:51:20can basically create your own game, and
2:51:22then work with Clayco at the same time.
2:51:26Quick one before we get back to the
2:51:27video. If you're looking to work with me
2:51:29one-to-one to be able to start and scale
2:51:31your own AI automation agency in a
2:51:3390-day one-on-one mentorship program,
2:51:34then check out the first thing down
2:51:35below. With that being said, let's get
2:51:37into it. All right. So, you've now got
2:51:38eight proven hacks to build better agent
2:51:40workflows. But everything we built so
2:51:42far lives in your machine, lives in your
2:51:45computer, which is why in the next
2:51:46lesson, I'm going to show you the top
2:51:48four ways that you can deploy your agent
2:51:50workflows onto the cloud, so they run
2:51:5224/7 without you having to touch a
2:51:54single thing. Let's dive in. All right.
2:51:55So, first of all, what's deployment?
2:51:57Because I know it's a word that
2:51:58everybody says, but only a few people
2:52:00actually know what it means. But
2:52:01essentially, when you're building an
2:52:03agent in Clayco, so let's say you want
2:52:05to build an agent that can run every
2:52:06single morning at 3:00 a.m. When you're
2:52:08testing, it's hosted on your computer.
2:52:10That means you tell it when to run. You
2:52:12press the button, you extract it to run
2:52:14the actual workflow to give you the
2:52:15desired output. The problem with this
2:52:17method is that if you close your laptop,
2:52:19it stops. And if the Wi-Fi goes out, it
2:52:21stops. It's sort of like a plant that
2:52:23only lives
2:52:23>> [music]
2:52:23>> in your desk, and it can only survive if
2:52:25you water the plant. And if you want the
2:52:26workflow to run every 3:00 a.m. in the
2:52:28morning, then you would have to be in
2:52:30your computer turned on with Wi-Fi to be
2:52:32able to make it work, or else it just
2:52:33wouldn't run. And so, in very easy
2:52:34terms, deployment means that we're
2:52:36taking this exact agent that we built in
2:52:37Cloud Code, and we're hosting it, we're
2:52:39putting it in someone else's computer.
2:52:40And this is great because the computer
2:52:42is always on, it's always connected to
2:52:43the internet, but most importantly, it
2:52:45runs 24/7. So, you can actually run the
2:52:47workflow at 3:00 a.m. every single day
2:52:49without you having to be in your own
2:52:50computer. Now, getting more in-depth
2:52:52into it, a computer here is called a
2:52:54server. So, this right here is a server.
2:52:55And servers live in big warehouses
2:52:58called data centers. So, data centers
2:53:00right here, they
2:53:01basically store all these computers,
2:53:03right? Each one is a computer right
2:53:04here. And a computer hosts your
2:53:06workflow, and it's always on.
2:53:08Okay, that's essentially what I'm
2:53:09getting at. Now, the good news about
2:53:10deploying it to these servers is that
2:53:11you don't have to buy a server, and you
2:53:13don't have to know exactly how it works.
2:53:14Cuz what they do is they rent server
2:53:16time, so you pay for the time that the
2:53:18workflow is in the actual server, and
2:53:20they handle all the complicated stuff
2:53:21for you. So, all you have to do is build
2:53:22it in Cloud Code, you have to take the
2:53:24code, and then upload it in the app, and
2:53:26then it runs it for you. And these are
2:53:28the companies that we're going to be
2:53:29talking about today. So, here's an
2:53:30agenda workflow that I built in Cloud
2:53:31Code. Now, I'm not going to show you the
2:53:33full build step-by-step, cuz if you want
2:53:34a full video, you can check it out up
2:53:35here. But essentially, it is a strategy
2:53:37system that runs every single week on a
2:53:40Monday, and it gives me a full
2:53:41presentation, basically comparing my
2:53:44analytics with my competitors on
2:53:45YouTube, and then it gives me content
2:53:46ideas. So, I'm going to show you the
2:53:47output first, so you can see exactly
2:53:49what it is that we need.
2:53:51I can put {slash} start. So, I just ran
2:53:53the actual workflow. All right, so the
2:53:55workflow just finished running. It gave
2:53:56me the weekly report right here that I
2:53:58can copy,
2:53:59and then I can paste in my browser right
2:54:01here.
2:54:02And this is a presentation that we get
2:54:03every single uh week. So, I get an
2:54:05analytics report, which tells me my
2:54:07stats for my channel,
2:54:08my top-performing video right here,
2:54:11my benchmark comparisons against other
2:54:12competitors in my space with views,
2:54:14engagement, and so on. Key insights for
2:54:16my channel that I can see. Then I get
2:54:17five ideas, so 1 2 3 4 5 [music]
2:54:20that I can use, and then finally I get
2:54:21next steps as well. So, now we get to
2:54:23the part where we deploy the actual
2:54:24workflow. So, there are four main ways
2:54:26to deploy the agent workflow. The first
2:54:28way is using modal. The The second way
2:54:30is using railway. Then we have render
2:54:32and then we have replit. So we're going
2:54:33to start with modal right here because
2:54:35it is one that a lot of people in the
2:54:36cloud community are already using and
2:54:38for good reason because it was literally
2:54:40built for AI workflows. So first thing
2:54:41we want to do is press get started. You
2:54:43can then sign up to Google right here.
2:54:45You can choose your account.
2:54:47Just continue. You can do personal and
2:54:49then you can get started. And one
2:54:50important thing here is that we get $5
2:54:52of free credits. But you can also have
2:54:54$30 of free credits if you add your card
2:54:56details. So once you add a payment
2:54:57method you should be able to see that
2:54:58you have now $30 of free credits right
2:55:01here. Now you do get charged 50 cents
2:55:03for the actual verification, but it's
2:55:05worth it because then you get six times
2:55:07the amount of money uh that you can use
2:55:09on your own account. So the first step
2:55:10here is going to Claude code and saying,
2:55:12"Hey, I want to host the um agent
2:55:14workflow to run every morning
2:55:18on a Monday at 9:00 a.m. London time and
2:55:21I want to host it in modal. And all I
2:55:23have to do is go back here and I can run
2:55:25this.
2:55:26So I can copy this and then simply paste
2:55:28it here.
2:55:29And I can press go. We could go down
2:55:30different ways, but this is the easiest
2:55:32way to just tell Claude code to figure
2:55:34it out before it starts adding it to the
2:55:36modal deployment. So after a few minutes
2:55:38you will get Claude which will make a
2:55:39plan and will go step-by-step as to how
2:55:41it needs to go through deploying the
2:55:42workflow. All you have to do is press
2:55:44yes and auto accept. And now it's going
2:55:46to start executing on the actual plan.
2:55:47All right, so the first thing it does is
2:55:48that it creates tasks that show us
2:55:50step-by-step what it has to do before
2:55:52starting the plan of the deployment in
2:55:54modal. All right, and here's the plan
2:55:55that Claude has to deploy the system on
2:55:57modal. We have step one until step
2:56:01four. All we have to do is press yes and
2:56:03auto accept. And you can also read
2:56:05through this if you want, but of course
2:56:06everything is in code, so you won't
2:56:08really understand exactly what it is.
2:56:09But one important part is as you can see
2:56:11we just made a file called modal app.py
2:56:14which is going to be the Python file.
2:56:15Not that you should know exactly what it
2:56:17says here, but it'll be the Python file
2:56:18that it uses to then deploy the actual
2:56:20workflow on modal. And we also see that
2:56:21it made tasks to then deploy to modal
2:56:23itself. All right, so once everything is
2:56:24done, you should be able to see that it
2:56:26finished all the to-do's [music] and it
2:56:27says test runs results all passed. And
2:56:29now you can press this link right here,
2:56:30which will take you to the actual app
2:56:32[music] in the place as well. And if you
2:56:34want to see all the apps, all the
2:56:35different workflows that you have, you
2:56:36can simply just go to apps and then
2:56:38press on each one in specific. Now, here
2:56:39we have the overview, which is the app
2:56:41function, so the thing that we actually
2:56:43need to do. Then we have the deployment
2:56:45history. So let's say you want to change
2:56:46something in the actual agenda workflow,
2:56:48you don't actually go here and change
2:56:49it, you go back to cloud code, you go
2:56:51right here, you then tell it to change
2:56:53[music] things, and then it
2:56:53automatically updates and it makes a new
2:56:55version of the actual agenda workflow.
2:56:57And that's what we can see V1, V2, V1.3,
2:56:59V1.4, and so on. [music] And you get to
2:57:01see all the different deployment
2:57:02histories here, so you can always come
2:57:04back to this and maybe change to the
2:57:05previous version. Then we have app logs,
2:57:07and by the way, replace the word app
2:57:09with workflow, so workflow logs. So you
2:57:11can see all the different logs, all the
2:57:12different runs that we had before. And
2:57:14then we have one weekly YouTube report,
2:57:17which is the action that the actual
2:57:18workflow needs to do. We get to see all
2:57:20the function calls, and then up here we
2:57:22get to see that it is scheduled for 9:00
2:57:24a.m. and only on Mondays, London time.
2:57:26And the next one will be February 16th
2:57:28as today is Wednesday at 1:00 p.m. Now,
2:57:30to show you how it does look like when
2:57:32it runs, we can press run now. And what
2:57:34this will do is it will schedule an
2:57:35immediate run, and we should be able to
2:57:36see the run show up somewhere here. As
2:57:39you can see, it just showed up. So
2:57:41queued, started, start up, and running.
2:57:43So that means that it's executing the
2:57:44actual workflow. And just so you know,
2:57:45the workflow finishes by sending myself
2:57:47an email with the actual report every
2:57:49single week. So we should be able to get
2:57:51an if it succeeded. All right, as you
2:57:53can see here that it took 56 seconds for
2:57:55the presentation to actually run, and I
2:57:57can see the status is success.
2:57:59We can see the time here. And if I go to
2:58:00log, I can see all the steps that it
2:58:02took to be able to make the actual
2:58:03presentation. Now, to show you that it
2:58:05did actually make the presentation, I
2:58:06can go to my email, and I can see that
2:58:08the latest email we got is weekly
2:58:09YouTube analytics report, and I get an
2:58:11email saying, "Hey, this is your
2:58:13report." I can press view presentation,
2:58:15and this will take me to the actual
2:58:16report with all the different ideas and
2:58:18so on. As you can see, it took us less
2:58:19than 2 minutes to set everything up and
2:58:21a good thing is that you were now able
2:58:22to run it every 5 days, so it will run
2:58:24every single Monday, but you can also
2:58:26test it the same way that you're in a
2:58:28workflow platform like Any Ten or and
2:58:30make.com. You can simply just manually
2:58:32press this and [music] see if it works
2:58:34in real time. Now, there is a chance
2:58:35that when you do run this, it doesn't
2:58:36work, which is why you can go inside
2:58:38here and it will show you the uh message
2:58:40as error, so there's something wrong
2:58:42with it and you can simply copy this, so
2:58:44copy the whole thing, bring it back to
2:58:46Cloud Code and say, "Hey, I just ran
2:58:48this on Model. It didn't work. It gave
2:58:49me this error." And you can paste it
2:58:50here. Okay? Now, of course, this is
2:58:52success, so I'm not going to run this,
2:58:54but if you have an error, that's what it
2:58:55will show you. You copy this, you paste
2:58:56[music] it into Cloud Code and you let
2:58:58Cloud Code debug and self-heal its own
2:59:00workflow. So, in terms of pricing, the
2:59:02way that Model works is that we are
2:59:03actually able to pay per second of the
2:59:06workflow running. Now, there are three
2:59:08plans. The first one is starter, team,
2:59:10and enterprise. The good thing about
2:59:11Model is that it is very easy to set up.
2:59:13Plus, because we get $30 of free
2:59:14credits,
2:59:15we're pretty much able to run this every
2:59:17single month for completely free, unless
2:59:19you have some astronomical automation
2:59:21and you have like 50 agents that are all
2:59:22running at the same time, but you're
2:59:24most likely not. You get three
2:59:25workspaces, you get 100 containers, and
2:59:2710 GPU concurrency. Now, what this means
2:59:29basically is the size of the automation,
2:59:30so the space that you have. So, I
2:59:32definitely recommend starting with the
2:59:33plan right here and just know that it is
2:59:35free because you get $30 of free
2:59:36credits. So, who is Model for? Now,
2:59:38Model is great for beginners who want
2:59:40the easiest path to deploy and maintain
2:59:43a workflow built in Cloud Code. It's
2:59:45also best for people who have a workflow
2:59:46that is Python-based, which it probably
2:59:48is if since you're using Cloud Code. It
2:59:50is also best if you have a workflow that
2:59:52is not running 24/7, but it runs in
2:59:54bursts. So, maybe it starts the
2:59:55automation when it gets triggered by
2:59:57something. So, it can be every single
2:59:59day at 9:00 a.m. it gets triggered. And
3:00:00also, it's great if you want to pay only
3:00:02for what you use. So, the second way
3:00:03that you can host your workflow is using
3:00:05a software called Railway. Now, Railway
3:00:06is a slightly more traditional hosting
3:00:08platform, but it is still incredibly
3:00:10simple to use. Think of it as a very
3:00:12friendly modern version of what
3:00:13developers used to use when deploying
3:00:15workflows. All we have to do is press
3:00:16deploy.
3:00:18And we can see here that we have the
3:00:19option of GitHub, database, template,
3:00:21Docker image, function, bucket, and
3:00:22empty project as well. So, we're going
3:00:24to be using GitHub. Now, GitHub is a
3:00:26software that we now have to make an
3:00:27account as well, which is a software
3:00:29where we basically upload the code
3:00:30there, and it's going to be the
3:00:31middleman between Cloud Code and Railway
3:00:34itself. So, the first step is going here
3:00:36and going to github.com.
3:00:38github.com
3:00:39All right, so here's GitHub, and think
3:00:40of this as your Google Drive folder. So,
3:00:42we're going to host all the different
3:00:43files in a Google Drive folders, and
3:00:44[music] then we're going to connect the
3:00:45Google Drive folders to Railway, so it
3:00:47can look inside the code, and then be
3:00:48able to host it there. So, first things
3:00:50first is you want to sign up.
3:00:52You want to continue with Google.
3:00:54Choose your email. Press continue.
3:00:57Create an account.
3:00:58And once you've made an account, you
3:00:59will get on a page that looks like this.
3:01:01So, the first step that we want to do is
3:01:02you want to create a repository, because
3:01:04we're making the Cloud Code agent, we're
3:01:06hosting it on GitHub, and we're
3:01:07deploying it on Railway. We can create a
3:01:09repository. We can name this
3:01:11uh agentic workflow. We can do public.
3:01:14Everything else is fine. We can create
3:01:15the repo, and then you will get on a
3:01:17page that looks like this.
3:01:19All we want to do here is you want to
3:01:20copy this URL, bring it back to Cloud
3:01:22Code, and say,
3:01:24"Hey, can you um upload all the code to
3:01:26this GitHub repo?"
3:01:28And then you want to paste the link.
3:01:30>> [music]
3:01:30>> Okay, so we're saying, "Hey, take all
3:01:31the code of this and connect it to
3:01:33GitHub repo,
3:01:34and so that we can then use Railway
3:01:36here, so that we're able to see the code
3:01:38in GitHub through Railway." Cuz there's
3:01:40no way of connecting it directly to
3:01:41Cloud Code, that's why we're using this.
3:01:42And I can press go.
3:01:44All right, so here we have the plan to
3:01:45push code to the new GitHub repository.
3:01:47We can press yes, auto accept. All
3:01:49right, so now it's telling me that I
3:01:50need to get something called an SSH key.
3:01:52Now, if you've heard of API keys or
3:01:54passwords, then this is exactly what
3:01:55it's asking us. So, you can press add
3:01:57key to account.
3:01:59Submit answer. So, what you want to do
3:02:00here is you want to go to
3:02:01github.com/settings/keys,
3:02:04and you will see the SSH and GBG keys.
3:02:07All you have to do is press new SSH key.
3:02:10You can name this MacBook.
3:02:12Authentication key is fine. And then
3:02:13make sure to ask Cloud Code to give you
3:02:15the actual key that you can paste. Once
3:02:17you have the key, you can copy it. You
3:02:18can paste it here and then you can add
3:02:20an SSH key. And that's pretty much all
3:02:22you have to do here. I can go back and
3:02:23say, "Okay, I just finished [music] and
3:02:24added it there. Keep going." And it says
3:02:26that all my code has been pushed to my
3:02:28Agent Workflow. So, I can press this
3:02:29link right here.
3:02:31Which will take me to my repo, which is
3:02:32this one right here. Or if you're stuck
3:02:34on here, you can simply go to three
3:02:35dots. You can go to home and you'll be
3:02:37able to see that you have the Agent
3:02:38Workflow here that you can then see. So,
3:02:39all the files that we have here are all
3:02:41the files that we have here. And now I
3:02:42can go back to Railway. I can press
3:02:44GitHub repo. I can log in with GitHub.
3:02:47I can authorize the Railway app.
3:02:50And here you can press I agree. I will
3:02:51not deploy any You can press this button
3:02:53right here. And you want to press
3:02:54configure GitHub app. All [music]
3:02:56repositories, press install and
3:02:57authorize. And then you want to press
3:02:59new GitHub repo.
3:03:01And now you'll be able to see the
3:03:02Michaela Partnerships one. Press this
3:03:04right here. It's deploying the repo in
3:03:06Railway. As you can see now it was an
3:03:07error. So, I can press this right here.
3:03:09It's error creating build plan with real
3:03:10pack. So, what I would do here by
3:03:12instinct is go back to Cloud Code and
3:03:14say, "Hey, I've tried to upload it to
3:03:16GitHub and I got everything working. I
3:03:19got all the all the files. And then I
3:03:20tried putting it everything into
3:03:21Railway, so connecting Railway to
3:03:23GitHub, and it said error creating build
3:03:25plan with real pack. So, I don't know
3:03:27what to do here." It's going to ask me,
3:03:28"What do you want Railway to do with
3:03:29this project?" This will determine how
3:03:30we set it up. We can run it on a
3:03:32schedule.
3:03:33So, we can just [music] yes and auto
3:03:35accept the plan that he uses to then fix
3:03:37the Railway. All right, so he just fixed
3:03:39it so I can go back here and I can see
3:03:40that this is now building. So, he just
3:03:42fixed the thing for me by actually
3:03:44modifying the GitHub repo, which
3:03:46automatically updates the Railway app.
3:03:48And if I go here, I can see that I have
3:03:49building status. As you can see here,
3:03:51the deployment was successful. And now
3:03:52we have the Agent Workflow, which is
3:03:53online, which is now be triggered
3:03:55automatically and is uploaded and
3:03:57deployed. If I go back here, I can see
3:03:59that now I have the actual app in
3:04:00production. Now, bear in mind that this
3:04:03is also running on a cron timeline,
3:04:05which means that we are running it every
3:04:06Monday at 9:00 a.m. UTC. Up here you
3:04:08have the architecture, which is where
3:04:09you can actually edit the Agent
3:04:11workflow. Then we have this tab right
3:04:12here, observe the environment, so you
3:04:14can monitor different projects and look
3:04:16at the metrics. Then we have log, which
3:04:17is very similar to the one that we're
3:04:18talking about before, which is
3:04:19essentially all the logs that we have
3:04:21here. And whenever something goes wrong,
3:04:22whenever there's an error, you can
3:04:23always go here and copy this and then
3:04:25paste it into Cloud Code and it will fix
3:04:26it automatically. And then we also have
3:04:28settings here as well. All right, so in
3:04:29terms of pricing, we have the $0 per
3:04:31month plan, which is actually isn't a $0
3:04:34a month plan, cuz we start with a 30-day
3:04:35free trial with $5 credits, and then
3:04:37it's $1 per month. Now, if you're
3:04:39confused about VCPU, what that means is
3:04:41basically how fast your computer can
3:04:43think. So, one VCPU means that you only
3:04:45get one brain for working on your tasks,
3:04:48whilst 0.5 GB RAM is how much the
3:04:50computer can remember all at once.
3:04:52Basically like a short-term memory or
3:04:53the space that it has to contain
3:04:54everything. And I would recommend this
3:04:56plan to start out, but as you start to
3:04:57build more and more agents, you ideally
3:04:59want to go to the hobby plan, just
3:05:00because there isn't enough RAM and CPU,
3:05:02but also there's rate limits on the
3:05:04things that you can actually build. So,
3:05:06for one agent, it's okay. For multiple,
3:05:08I'll recommend going to the hobby, but
3:05:09you also get a $20 a month plan as well.
3:05:11Now, when it comes to Railway itself,
3:05:12it's best if your workflow needs to run
3:05:14continuously. So, maybe you need a bot
3:05:15that runs and checks the workflow 24/7,
3:05:18Railway is good for that. It's also good
3:05:19if you want the database support built
3:05:21in in the software. And it's also good
3:05:23if you like the idea of putting
3:05:24everything in GitHub and then
3:05:26automatically deploying it on Railway.
3:05:28Now, in terms of debugging, so when
3:05:29something goes wrong, the way that it
3:05:30tactically works, so you understand, is
3:05:32that So, let's say there's an error, we
3:05:34go to Cloud Code, we say, "Hey, there's
3:05:35an error." Cloud Code fixes the code,
3:05:37which then goes to GitHub and updates
3:05:39the code with the new updated version,
3:05:41and then it [music] goes to Railway and
3:05:43redeploys the actual Agent workflow. All
3:05:45right, so the next platform is called
3:05:46Render. Render is often described as the
3:05:48modern Heroku. And if that means nothing
3:05:50to you, then just think of it as a
3:05:52really reliable, no surprise hosting
3:05:54platform. So, all you have to do here is
3:05:55you want to get started.
3:05:57You want to go to Google, choose your
3:05:58account, continue,
3:06:00and you will get on the page that looks
3:06:01like this. You want to press on
3:06:02background workers as this is the Agenta
3:06:04workflow uh builder. And then, you want
3:06:06to connect this to GitHub. So, this is
3:06:07very very similar to Railway in the
3:06:09sense that we use GitHub as a middleman
3:06:11between Claude code and the deployment.
3:06:13We press GitHub here.
3:06:14We can authorize this. So, authorize
3:06:16Render. And when we go back, we can
3:06:17install. And right here, I can see that
3:06:19now I can choose my repo.
3:06:21So, choose this here.
3:06:22I can name this here Python 3. Leave
3:06:24this all the same. And now, I have to
3:06:26choose between one of the plans. Because
3:06:27Render does not have a free plan, you
3:06:29either start with the $7 a month plan,
3:06:30or you don't start at all. So, Render,
3:06:32in contrast to the other ones, does not
3:06:33have a free plan. So, you have to pay
3:06:35for the $7 a month plan if you want to
3:06:37use the software here. We can go here.
3:06:38We can deploy background worker. And
3:06:40now, this will take you to this right
3:06:42here, which is now going to GitHub. It's
3:06:44extracting all the code. So, it's taking
3:06:45all the code and it's [music] building
3:06:47the actual workflow and hosting it and
3:06:48deploying it inside Render right here.
3:06:50All right. And after a few minutes, we
3:06:51can see that it is now successful. It
3:06:53says the service is live. And we're now
3:06:55able to use this workflow automatically.
3:06:56And now, this is hosted easily on Render
3:06:58right here. If I go to logs, I'll be
3:06:59able to see all the logs of the actual
3:07:01workflow. And metrics here is where you
3:07:02can track the memory and CPU that the
3:07:04workflow is using over time. And you
3:07:06also have other things as well here. And
3:07:08now, the workflow is going to be
3:07:09triggered at 9:00 a.m. every single
3:07:10Monday every week as [music] that is
3:07:12what we added to GitHub. The process of
3:07:14debugging is the exact same as the one
3:07:16that we had on Railway. You basically go
3:07:18to Claude code. You fix the code, which
3:07:20then updates the code on GitHub, and
3:07:21then finally redeploys on Render right
3:07:24here. So, who is Render for? Render is
3:07:26great if you want predictable flat rate
3:07:27pricing for your workflows. It's also
3:07:29good if you need background workers for
3:07:31your agents. So, basically sub agents
3:07:32for your main actual agent to work. It's
3:07:34a platform that has been around for a
3:07:35while, and it has been battle tested.
3:07:37So, if it's important to you to have a
3:07:38platform that has a good reputation and
3:07:40has been for a while, then Render is the
3:07:41choice. And also, it's great if you
3:07:43value automatic rollbacks when something
3:07:45breaks. All right. And the last platform
3:07:46here that we're going to be using to
3:07:47deploy is called Replit. Replit is one
3:07:50of the most popular platforms as well
3:07:51because it does actually have the
3:07:53agentic feature. Very, very good. So,
3:07:56first thing you want to do is press
3:07:56create an account, choose your email,
3:07:58continue, put your name.
3:08:02Press next, personal project, that's
3:08:03fine.
3:08:04And now, we can go straight to pricing.
3:08:06So, we have the free version here where
3:08:08you can publish one app. Then we have
3:08:09the Replit Core, which is the most
3:08:10recommended one, is the one that most
3:08:11people use where you can publish more
3:08:13apps and you have more money as credits
3:08:14that you can use. And then finally, we
3:08:16have the Teams plan as well. And it's
3:08:17only something that you use if you're
3:08:18trying to build something more complex
3:08:20and you have more and more in your
3:08:21organization. So, I'm going to go ahead
3:08:23and start with the starter.
3:08:25And here we have the app right here. I
3:08:26can press import code or design. I can
3:08:28go to GitHub.
3:08:29I can connect my GitHub account.
3:08:31>> [music]
3:08:31>> I can press continue to GitHub source
3:08:33control. I can authorize Replit
3:08:34integration. It just connected my GitHub
3:08:36here. I can choose agentic workflow.
3:08:38And I can simply press import from
3:08:40GitHub.
3:08:41As you can see now, the workspace is
3:08:42loading and it's now importing all the
3:08:43code from GitHub to Replit to then be
3:08:45able to deploy the agent here. As you
3:08:47can see now here, it says status okay,
3:08:49service is used strategist and it says
3:08:51the app is running successfully. Let me
3:08:52verify it responds correctly and then
3:08:54set up the deployment. All right, so now
3:08:55it's finished and it says everything is
3:08:56set up and running. Let me finalize the
3:08:58import and it basically imported all the
3:09:00different files here. And all we have to
3:09:01do is press published. And if I go back
3:09:03to the home page, I can see that I have
3:09:05the agentic workflow, which is the one
3:09:06that we have right here. If I go to
3:09:07apps, I can see that I have the agentic
3:09:09workflow, which is private, and this
3:09:11will now be able to run every 9:00 a.m.
3:09:13in the morning on a Monday. Replit is
3:09:15great if you want to code and deploy in
3:09:17the exact same place. It's also best to
3:09:19use if you're building a quick prototype
3:09:20or a proof of concept for something
3:09:22because it has the agentic feature which
3:09:23allows us to write code different
3:09:25softwares as well. And it's also great
3:09:26if you don't want to deal with GitHub
3:09:27itself or terminals or any setup at all.
3:09:30So, those right there were the four main
3:09:31ways that you can deploy and maintain
3:09:32your agentic workflows that you build on
3:09:34Cloud Code. Now, one important thing is
3:09:35that no matter what platform you choose,
3:09:37the debugging workflow is the exact
3:09:39same. So, here's the process. Your
3:09:40workflow breaks, you go to the
3:09:42platform's dashboard, you open the logs
3:09:44cuz every platform has logs, you find
3:09:45the error message, which is usually in
3:09:47red, You copy the error message, you go
3:09:49to Cloud Code, you paste the error
3:09:50message and say, "This is the error for
3:09:52my deployed agent. Fix it." Cloud Code
3:09:54fixes the code, you redeploy, which
3:09:56means that you push to GitHub or you run
3:09:58the deploy command, and you check if it
3:10:00works. If not, go back to step one,
3:10:01which makes it so much easier for you as
3:10:03you don't have to go in the code
3:10:04yourself. You can simply let Cloud Code
3:10:05take care of the whole thing. So, if
3:10:06you're still not sure as to which
3:10:07software you should go with when it
3:10:09comes to deploying, then here's the
3:10:11thought process. If you're someone who
3:10:12says, "I want the absolute easiest
3:10:14option to deploy," well, Modal is your
3:10:16best option. If you're someone who says,
3:10:17"My agent needs to run 24/7 like a
3:10:19background process," then Railway and
3:10:21Render will be good for you. If you're
3:10:23someone who says, "I want predictable
3:10:24monthly bills with no surprises," then
3:10:26Render is good. If you say, "I want to
3:10:27code and deploy without leaving my
3:10:29browser," then Replit is the way. And if
3:10:31you say, "I just want to test something
3:10:32quickly for free," then you can use
3:10:34Modal. Now, we do also have some other
3:10:35options like AWS, GCP, or Azure. The
3:10:38only reason why I did not include them
3:10:40is because they're more for
3:10:40enterprise-grade level uh deployment,
3:10:43and they're more advanced as well. And
3:10:44that right there is a complete cuz
3:10:45building the agent workflow was just the
3:10:47first half. Deploying is everything that
3:10:49comes after you build it, and it's
3:10:51arguably the most important step as
3:10:53building it is fun, right? But how do
3:10:54you actually give it to the client? How
3:10:56do you deploy it so it actually works
3:10:57without you being there all the time?
3:10:59And the good thing is that to deploy an
3:11:00agent workflow with all this code, you
3:11:02don't have to be a developer. All you
3:11:03have to do really is just follow step by
3:11:05step what I did in this video and let it
3:11:07run on its own.
3:11:12Right, congrats on finishing module
3:11:14four. At this point, you have
3:11:16everything. I mean, the foundations, the
3:11:17customizations, the prompting, the agent
3:11:19workflows, and the deployment itself.
3:11:22So, at this point, you're not just
3:11:23someone who uses Cloud Code, but you're
3:11:25someone who understands how it actually
3:11:27works, but also how to build with it.
3:11:29Now, it's my turn to prove it to you
3:11:30because in the final module, you're
3:11:31going to watch me build real products
3:11:33live from scratch using everything that
3:11:35we covered so far. There's no theory,
3:11:37just building, making it really, really
3:11:39practical for you. Let's dive in. All
3:11:41right, so when it comes to my YouTube
3:11:42content pipeline, there are four
3:11:43different stages. The first one is the
3:11:45ideation, which contains three different
3:11:47types of we call them skills or
3:11:48automations that I built inside of Cloud
3:11:50Code. The first one here is called the
3:11:51comment scraper, which is where the
3:11:53automation itself will go to x amount of
3:11:55videos posted by x amount of people in x
3:11:58amount of time, and then extract all the
3:12:01comments and then give me ideas for con-
3:12:02Twitter monitoring is where it's going
3:12:03to the actual Twitter profile, it's
3:12:04extracting all the tweets and gives me
3:12:05ideas. And now a lot of videos is the
3:12:07videos that have gone well for other
3:12:08channels, why have them gone well and
3:12:10how can you apply the same idea to my
3:12:12channel? Then we have packaging, which
3:12:13is title and thumbnail. So for example,
3:12:15this video right here, we have a title,
3:12:17there's a picture before that made you
3:12:18want to press on the actual video, and
3:12:20that's the packaging. [music]
3:12:21Then we have the video itself, which is
3:12:22the hook. So the first 30 seconds of
3:12:24this video, that was uh a hook, that was
3:12:26scripted. There's a script, which is
3:12:27what I say in throughout the video. Of
3:12:29course, a video like this, there's no
3:12:30script. I go off the rip, right? But
3:12:32something that's more like a head
3:12:34talking video where I go through a
3:12:36concept, right? Where it's me standing
3:12:37up. Well, that's scripted, so we have a
3:12:39script. And then, we have visual
3:12:41diagrams. So this right here, for
3:12:42example, is built by an automation. I
3:12:45didn't actually build it myself. That's
3:12:46diagrams. And then lastly, we have post
3:12:48video, which is looking at the
3:12:49performance of my videos, but also views
3:12:50to clicks. Views to clicks is looking at
3:12:53the amount of people who clicked to
3:12:54watch my video going through my program
3:12:57uh versus the amount of views, right?
3:12:58Cuz obviously, on the back of my videos,
3:13:00I want people to work with me
3:13:01one-to-one. I want to help them, I want
3:13:03to help them scale their agencies. So I
3:13:04want to check which videos do people
3:13:06like the most that make them want to
3:13:08kind of want to work with me one-to-one.
3:13:09That's what it is. So with that said,
3:13:10here is my Cloud Code workspace. It's
3:13:12called the YouTube workspace. It
3:13:13consists of different folders uh that we
3:13:15have. The first one is doc.cloud, then
3:13:17we have context, executions, outputs,
3:13:19references, scripts, .env, and then the
3:13:21main one here, which is cloud.md. Now, I
3:13:23actually separated the folders into four
3:13:26different categories. The first one is
3:13:27commands, then we have skills, then we
3:13:29have context, and then we have the
3:13:30cloud.md. I'll explain what these are as
3:13:32well, but not a main priority like they
3:13:34are in main folders that we use, okay?
3:13:36So the first one is called commands. So,
3:13:38commands are reusable slash commands,
3:13:40like slash competitor analysis or slash
3:13:42create plan or slash something else,
3:13:44that allow you to trigger, to start any
3:13:47workflow just by calling it using a
3:13:49slash, right? And when I say workflow, I
3:13:51mean it triggers a skill. So, if I want
3:13:53to run the Excalidraw diagram, for
3:13:55example, right? Then I would do slash
3:13:57Excalidraw diagram, and that would call
3:13:59that specific skill instantly. Now, if
3:14:01you're not sure what skills are, first I
3:14:03recommend you watch this video up here,
3:14:04um but in case you are aware, skills
3:14:06would be modular instructions that we
3:14:08give Claude to explain or to give it an
3:14:12overview of how to do a specific task
3:14:14step-by-step. So, for example,
3:14:15Excalidraw is a platform like Miro where
3:14:17you build kind of visual diagrams, like
3:14:19I showed you before. Well, inside the
3:14:20skill, there's step-by-step instructions
3:14:22as to how to do that specific thing. And
3:14:23so, all in all, our Claude is going to
3:14:25be our kind of company, right? And
3:14:29inside the company, there's different
3:14:30employees, which are responsible for
3:14:32different things. And those employees
3:14:34are skills. Then we have context, so
3:14:36everything that Claude needs to know
3:14:37about my business. So, first of all, who
3:14:39I am, my strategy when it comes to the
3:14:41content, my data, and my content
3:14:44guidelines. Cuz, for example, let's say
3:14:45we're using the skill to script my
3:14:47videos, right? I'm making a script. When
3:14:49I want to make my script, I also want
3:14:50Claude to use my YouTube content
3:14:52guidelines and business info to check,
3:14:54"Hey, this is the script, this is the
3:14:56guidelines. Are we sure we're following
3:14:58the right guidelines?" That's all it is.
3:15:00And then finally, we have the Claude.md,
3:15:01which is a master file here that ties
3:15:03the whole workspace together, telling
3:15:05Claude how all the pieces work in one
3:15:07system, okay? So, it tells Claude, "Hey,
3:15:08we have the folder here which is called
3:15:10context, the folder here which is called
3:15:12skills," and it gives it an overview of
3:15:14the whole organization. The reason that
3:15:15is is because it will look at the
3:15:16Claude.md every single time before it
3:15:19does anything. That way, it has context
3:15:20as to your workspace and how it all ties
3:15:23together. Now, first things first is the
3:15:24ideation. So, ideation consists of the
3:15:27comment analysis report, it consists of
3:15:29the competitor analysis, which is the
3:15:31outlier videos, the videos that have
3:15:33done really well for other channels, and
3:15:34then we have Twitter cloud monitoring.
3:15:35I'm going to run it first and then I'll
3:15:37walk you through how it's set up as
3:15:38well. So, dot comment analysis report.
3:15:41And then we can do bypass permissions
3:15:43cuz I don't want to tell it yes,
3:15:45approve, approve, approve. I just want
3:15:46CloudCode to do its own thing cuz it
3:15:48knows everything already. As you can see
3:15:49here, it says that it fetched 1,369
3:15:51comments across five different channels.
3:15:53Now it's reading through all the
3:15:54comments. It's looking at the YouTube
3:15:56content guideline, very important. All
3:15:57right, so now the skill finished
3:15:58running. It said it found 1,000 369
3:16:00comments. It made the PDF that I can see
3:16:03and it gave me the top five content
3:16:04opportunities that I can do. So, the PDF
3:16:06looks like this where we have my
3:16:07channel, top four content ideas
3:16:09requested by your audience. And by the
3:16:10way, these are the comments that you
3:16:11guys add to my videos. And then what I
3:16:13also did is I looked at Nate O'Brien,
3:16:15looked at the top four content ideas
3:16:16from his audience, looked at Liam
3:16:18Outlaw, looked at Evan Nichols Survive,
3:16:20and then Jack Roberts as well, right?
3:16:22And we have uh the ideas as well. And
3:16:24then it gave me the cross-channel
3:16:25themes. So, what are the common request
3:16:27or ideas that people want across all the
3:16:29channels? For example, how do I actually
3:16:31deliver AI work to clients? This came
3:16:33from my channel, Liam, and Nate. Or
3:16:35token management, cost optimization for
3:16:37CloudCode, Nate, Jack, and Liam, and so
3:16:40on, right? Plus, it also gave me content
3:16:42opportunities specific to me. So, how to
3:16:44deliver AI products to clients, complete
3:16:46guide. This is the biggest gap in the
3:16:47market. Nobody's making comprehensive
3:16:48content about the delivery side, account
3:16:50setup, hosting, handover documentation,
3:16:52maintenance agreements, access
3:16:54management. You've done this for real
3:16:55clients, make the definitive video,
3:16:58right? And we go step by step. Then we
3:16:59have the stats at the end with
3:17:00subscribers, videos, comments, and
3:17:02average comment per video. Guys, I
3:17:04believe you should be commenting more on
3:17:06my videos so we can get the numbers up.
3:17:07But hey, comparison is the thief of joy,
3:17:10am I right? Um anyways, we have the full
3:17:12PDF which I can use now as my way of
3:17:15generating ideas for content, which are
3:17:17really, really good. And inside the
3:17:18skill, we have the skill that I'm doing,
3:17:20which is the full instruction manual for
3:17:22the skill, which goes to step one, fetch
3:17:24comments. Step two, analyze the report.
3:17:26Step three, generate PDF. And that's it.
3:17:28And it gives you the PDF itself. Then we
3:17:30also have scripts, which are the
3:17:31different Python files, the code that
3:17:33basically allows the automation to run.
3:17:35Think of this like if you know n8n, if
3:17:37you've heard of n8n or make.com, think
3:17:39of the script, each script, as like a
3:17:41workflow, right? And so we are running
3:17:42workflows. The first one is fetching
3:17:44comments. So we're running the workflow
3:17:45to fetch comments, to extract comments
3:17:47from videos. Then we run the generate
3:17:48PDF as another workflow to basically
3:17:51make a PDF that we can use. And then we
3:17:53have references, which is the report
3:17:54format is giving the clock code a a way
3:17:57to understand what the format is of the
3:18:00report itself. All right, so we have raw
3:18:01stats, we have cross-channel themes,
3:18:03which is how we make the report look
3:18:04like this. How does it know the
3:18:05structure? And how is it consistent with
3:18:07the structure? Well, it has instructions
3:18:09for that. All right, so now that we're
3:18:10done with that, we can go to the other
3:18:11videos. So here I can press competitor
3:18:14analysis.
3:18:15Uh there we go.
3:18:16And then I can do bypass permissions.
3:18:19So I can let it run. So the first thing
3:18:21it does is that it looks at the CSV
3:18:22file, which is called competitors.csv.
3:18:24And this is a CSV file that I have down
3:18:26here, which is a list of 43 channels,
3:18:30including mine, which are going to be
3:18:32the channels that we use to extract the
3:18:34past 10 videos that they've done, and
3:18:36then we can start doing the math as to
3:18:38what videos have done really well for
3:18:39them compared to the other videos that
3:18:41they've done. All right, now it's saying
3:18:42that it just finished and exported
3:18:43everything to the Google Sheet. I'll
3:18:45show you exactly what that means. We
3:18:46found four all-around videos across 42
3:18:48channels. A lean week. All scored 10 out
3:18:51of 15. Cool. And then we have key
3:18:52patterns, most actionable for you. But
3:18:54if I go here to my views to clicks
3:18:56Google Sheet, where I host everything,
3:18:57we can see that we have the channel
3:18:59name, subscribers. I think it's a bit
3:19:01blurry, but it's fine. Video URL, the
3:19:03number of views, the upload date, the
3:19:05duration of the video, the medium views,
3:19:07so the average views that they got over
3:19:09the past 10 videos. And then we have the
3:19:11all-around multiple. So the all-around
3:19:12multiple is a mathematical equation that
3:19:15we use, which is how many views this
3:19:17video get over the average views of
3:19:19their channel to see if this exceeded
3:19:21that, right? And then also gives me a
3:19:22relevant score. So, for example, this
3:19:24video right here, which was about MCP is
3:19:26dead and nobody saw it coming, yeah,
3:19:29gave a zero relevant score because it's
3:19:31not aligned with my brand or audience.
3:19:33That's what I mean by Cloud Code using
3:19:35my brand guidelines and my the person
3:19:37that I'm targeting, what they care
3:19:38about, and what that looks like. And
3:19:40what we have relevant scores of three,
3:19:41three, and two. So, not only does this
3:19:43save me hours and hours and hours of
3:19:44time of me having to go to their videos,
3:19:46having to look at the views, having to
3:19:48average it out, having to make the
3:19:49equation, but it also gave me the
3:19:51relevance that I could use for my
3:19:53channel. And so, here I could do
3:19:55that to A. I can look at these two at
3:19:57the top ones. Let me look at the ideas,
3:19:59how to use Cloud Code work, how to get
3:20:00unlimited website clients. And so, I
3:20:01could do a video based on that in my
3:20:03kind of way. And this skill right here
3:20:04is consisted of references. So, this is
3:20:06the scoring and output, which is a kind
3:20:08of document that goes through how to
3:20:09score an outlier video, what does that
3:20:11look like? And then we have again the
3:20:12scripts, which are the different
3:20:14workflows, and then we have the
3:20:15skill.md, which is the step-by-step
3:20:17instructions that we go through to be
3:20:19able to actually make this work. And
3:20:20then we also have google.token.json,
3:20:22which is the way that we are able to
3:20:24access the YouTube API, but also my
3:20:26Google Sheet account. Then the last one
3:20:27we have is called the Twitter Cloud
3:20:29Monitor. So, Twitter Cloud Monitor is a
3:20:31workflow that goes to configuration
3:20:34profiles, yeah. These profiles in
3:20:36Twitter, and it might look a bit kind of
3:20:38weird, but uh by the way, Cloud Code,
3:20:40this not me, and it looks at these
3:20:42channels, not channels, these Twitter
3:20:44profiles, right? And what it does is
3:20:45that it goes through each profile and
3:20:47extracts the tweets from the past 24
3:20:49hours. Right? And we'd have this
3:20:51workflow running every single day, and
3:20:52it gives me the new updates on Cloud
3:20:54Code, cuz honestly, I can't keep them up
3:20:55myself, and I'm sure you can also do.
3:20:57So, that's exactly how a lot of
3:20:58YouTubers do get their ideas is through
3:21:00Twitter, right? So, why not make an
3:21:01automation that can automate that
3:21:03process of you having to find the tweets
3:21:05yourself. So, I can go here and put
3:21:06{slash} Twitter monitor Cloud Monitor. I
3:21:09can do bypass permissions, right? It
3:21:11just finished running. It said it
3:21:12scanned through 63 tweets. The relevant
3:21:14score above three is just two, and then
3:21:16we have the key findings, and it also
3:21:18made the PDF, which was sent to my
3:21:20email. If I go to my email, I can see
3:21:21that I have a new PDF.
3:21:23I can open it up, and it can give me now
3:21:26the new features that we have here, and
3:21:28then we have the content opportunities
3:21:30that we can use, right? Now, are all of
3:21:32them going to be uh great ideas? No,
3:21:34right? But, it filters them out. Cuz not
3:21:36every day Clock Coder releases
3:21:38something, maybe every 2 days or every 3
3:21:39days. So, it looks at the tweets, it
3:21:41checks the relevance, and then it gives
3:21:42me the report [music] as well. So, I
3:21:43basically keep myself updated without
3:21:44having to go to these 12 different
3:21:46profiles every single day and check what
3:21:48they've posted. That's how you get the
3:21:49YouTubers making videos about the new
3:21:51{slash} command or the new {slash} body
3:21:53or the new {slash} execution or the new
3:21:55{slash} whatever, right? Uh it's because
3:21:57they look at these exact profiles that I
3:21:59walked you through. Now that we're done
3:22:00with ideation, let's go to the next one,
3:22:01which is packaging. So, packaging is
3:22:03consisted of thumbnails and titles. So,
3:22:05let's go through thumbnails first. So,
3:22:06the first one is thumbnail creation. So,
3:22:07the way that this skill works is that we
3:22:09take a video. So, let's say we have a
3:22:10video uh like this. Let's copy the
3:22:13video. Copy the link. And then,
3:22:16we go back to Clock Coder, and I can put
3:22:18{slash} thumbnail
3:22:20creation.
3:22:22There you go. And then, what we do is I
3:22:23took that link of the video with Alex
3:22:25Hormozi. It will look at my face, right?
3:22:26That we have a picture of, which is
3:22:27going to be here, reference. It's right
3:22:29here.
3:22:31That's my face. And then, it will take
3:22:32that face, it will replace Alex
3:22:34Hormozi's face in that thumbnail with my
3:22:35face, and then give me back the
3:22:37thumbnail, basically. What source would
3:22:38you like to use as a thumbnail? I'm
3:22:39going to put the link here
3:22:41of that video that I copied. What it's
3:22:42going to do now is it's going to go
3:22:44inside this video, it's going to
3:22:45download the thumbnail, and then it's
3:22:46going to go to my picture, swap the
3:22:48face, and then give me back the image.
3:22:50And this is using a server called
3:22:52file.ai, because that's where we're
3:22:54going to use the nano banana model to be
3:22:57able to make this happen. All right, so
3:22:58it just finished the thumbnail. I'm
3:22:59going to say, "Can you open up the
3:23:02image?" And to be honest with you,
3:23:04it didn't do the best job, right?
3:23:09It replaced my face with his face.
3:23:12Obviously, that's who he looks like. I
3:23:13don't have a beard, as you can see. But,
3:23:15it copied the words. It made the words
3:23:16good. It made these massive traps that
3:23:19he has, and it put my face there. So,
3:23:21ideally, now we can go and say, "Hey,
3:23:23can you Can you remove the beard from my
3:23:25face?" But, it swapped the face
3:23:27initially. And by the way, this is all
3:23:28prompting. So, one thing I could say is
3:23:30like, "Hey, CollabCode, it actually
3:23:32added a beard. Can you remove the beard?
3:23:34Hey, CollabCode, can you remove the
3:23:35beard from my face in that thumbnail,
3:23:38and give me back the thumbnail?" And
3:23:39also, replace place the body with like a
3:23:40black shirt as well. Keep everything
3:23:41else the same. All right, it just
3:23:42changed it, and here we are, much better
3:23:44thumbnail here. It had no beard, and it
3:23:46put me in a better place as well. So,
3:23:47that's the thumbnail creation, which I
3:23:48think is nuts, cuz you can keep
3:23:50iterating it. You can keep saying, "Hey,
3:23:51can you change the text? Can you change
3:23:53this? Can you change that?" But,
3:23:54essentially, we're taking inspiration
3:23:56from other channels, and then using it
3:23:58for our own. Now, I don't recommend
3:24:00copying the exact same as someone else's
3:24:02thumbnail, but you can always copy the
3:24:04the way of thinking, or taking
3:24:05inspiration on the idea of the
3:24:07packaging, right? Now, with that said,
3:24:09we go to the next one, which is the
3:24:11title generation. Let's say I was going
3:24:12to do the title generation.
3:24:15I'm going to wait for it to ask me any
3:24:16questions. So, now it's saying, "What's
3:24:18the video content you'd like to title
3:24:19for?
3:24:20Any specific angle or hook you want? How
3:24:21many titles?" You can do 20 titles,
3:24:23that's fine. So, the video is about how
3:24:25I'm using CollabCode to automate my
3:24:27YouTube channel that I've grown to
3:24:2951,000 subscribers, about 2 million
3:24:31views in the past 7 months. That's what
3:24:34the video is about. I'm going to go to
3:24:35bypass permissions, just to make sure
3:24:37that it does everything. As you can see
3:24:38here, it's looking [music] at the title
3:24:39frameworks, because we have a list of
3:24:41title frameworks that work really well.
3:24:42So, 300 that it's using.
3:24:44It's then looking at the YouTube content
3:24:46guidelines, which is my MD document,
3:24:48which goes through my guidelines and my
3:24:50kind of avatar that I'm looking for. And
3:24:52then, it's going to give me 20 different
3:24:53titles. So, how I automated my entire
3:24:55YouTube channel with CollabCode. I use
3:24:57CollabCode to grow to 51 subscribers.
3:24:59Here's how. How CollabCode runs my
3:25:00YouTube channel, 51 51K subs in 7
3:25:03months. I let AI run my YouTube channel,
3:25:052 million views later. And it does a lot
3:25:07more, which I think are really, really,
3:25:09really, really good. And then it gives
3:25:10me the top titles based on curiosity,
3:25:12specificity, but also emotion. All
3:25:14right, with that said, we're done with
3:25:15packaging. Now we can go to the video
3:25:17itself. All right, so for the video
3:25:18itself, we have script writing, which is
3:25:20the hooks plus the script of the video,
3:25:22and then we have two different [music]
3:25:23Excalidraw diagrams and visuals that I'm
3:25:25going to walk you through. So the first
3:25:26one is script. So script writing,
3:25:29let's go here. Let's wait for it to ask
3:25:30me any questions about the video. Target
3:25:32length, any specific key points to
3:25:34cover, hook style and preferences, CTA
3:25:36goal. I'm going to start writing. There
3:25:38we go. So with all this said, I can
3:25:40press go, and then I can do bypass
3:25:42permissions to let it do its own thing.
3:25:44And while this is running, I can show
3:25:46you the actual references which we have
3:25:48here. So here we have the two different
3:25:51videos that I've done, so head talking
3:25:52videos, so previous examples of my
3:25:54scripts that have worked pretty well.
3:25:56Then the intro framework. So So what's a
3:25:58good hook? What does that look like? So
3:25:59we give it references. Script writing
3:26:01framework. So this is a YouTube video
3:26:03which talked about the best storytelling
3:26:06or the best ways to write scripts.
3:26:07>> [music]
3:26:07>> And then we have the SkilldonMD, which
3:26:09again is a step-by-step instructions as
3:26:12to what it needs to do. So step one,
3:26:14gather information. Step two, load
3:26:16context. Step three, build the audience
3:26:18avatar. Step four, write the hook. Step
3:26:19five, write the body. And then step six,
3:26:22write the mid-roll CTA. And step seven,
3:26:24write the closing CTA. Step eight, add
3:26:27production notes. Step nine, quality
3:26:29check. And then we have the output.
3:26:30Which is insane for AI to do all of
3:26:32this, cuz previously we had to spend so
3:26:34much time on everything, but now it does
3:26:35it in seconds. There we go. So now we
3:26:37have the script. I can go here and I can
3:26:39see that this is the full script. What I
3:26:41typically do, because this looks very
3:26:43confusing, is I copy this and I paste it
3:26:45in Notion
3:26:46like this. And now we have the hook. It
3:26:48took me 5 months to get my first AI
3:26:49agency client, 5 months of building
3:26:51systems that nobody asked for, sending
3:26:52messages that nobody replied to, and
3:26:53honestly questioning whether the whole
3:26:55thing was even going to work. And then
3:26:57in 14 months, I scaled the same agency
3:26:59to six figures working with over 50
3:27:01businesses and teaching over 22,000
3:27:03people in the process. All starting with
3:27:04zero business or technical knowledge. I
3:27:06can read the whole thing, obviously I'm
3:27:06not going to, in case you were
3:27:08wondering. But we have section one,
3:27:09which gives me the script itself. And
3:27:11I'm reading this, it's very, very good.
3:27:13And you can see it also gives the editor
3:27:15a kind of a reference of what edits
3:27:17would need to be here. Text on the
3:27:18screen, b-roll, pause, section two,
3:27:22section three, closing CTA, and that's
3:27:24it. Now bear in mind that you can use a
3:27:26general AI to make your script. There's
3:27:28a difference between a general slop AI
3:27:30script, and actually making a good
3:27:31script. And that's typically using
3:27:33references, using what's a good script
3:27:35look like, using human language, using
3:27:37good transitions. And so that's what we
3:27:39feed a lot of information to Claude Code
3:27:41to tell it, "Hey, here's the video that
3:27:43worked really well in the past. Here's a
3:27:45framework of how to make good scripts.
3:27:46Here's the idea, now give it back to
3:27:48us." And this also looks at the context.
3:27:51So the personal info, the strategy, the
3:27:52content guidelines, like I mentioned
3:27:54before. All right, so now that we've
3:27:55made the script, there are times in the
3:27:56video where we need supporting visuals.
3:27:58That would be a diagram, that would be
3:28:00some sort of uh yeah, a diagram or an
3:28:02illustration of a concept. So for
3:28:04example, here in the Claude Code Skills
3:28:06video, we have diagrams that look like
3:28:07these. We have this, all of these, where
3:28:10skills are stored, this, and it keeps on
3:28:12going, right? More and more diagrams.
3:28:14And so here I can say, "Hey, can you
3:28:15make an Excalidraw diagram based on a
3:28:17video that I'm doing about sub agents?"
3:28:19So make a diagram as to like what sub
3:28:22agents are, why they're important, how
3:28:24they're different from agent teams, and
3:28:27that's it. Make them really good,
3:28:28sophisticated, and give it back to me.
3:28:30I'm going to go on about this missions.
3:28:31There you go. Boom. So now, what it's
3:28:33going to do is it's going to read the
3:28:35skill first. So let me read the
3:28:36Excalidraw diagram first. As you can
3:28:38see, Claude Skills Excalidraw diagram,
3:28:40Excalidraw diagram.md. It's looking at
3:28:42the instructions, and now it's going to
3:28:44give us back the code, which is
3:28:45essentially the diagram itself. All
3:28:46right, [music] so it just finished
3:28:47making the diagram. I can go here, I can
3:28:49copy this. So command A, and then bring
3:28:51it back, and paste it here. God damn.
3:28:54Look at this.
3:28:55Beautiful. So sub agents architecture
3:28:56and purpose. What are sub-agents? Why
3:28:59they matter? And sub-agents versus agent
3:29:00teams. I mean, guys, come on. This is
3:29:02insane for it to be made in a few
3:29:04minutes. It made some errors here, but
3:29:06nonetheless, this is crazy. And you can
3:29:08instantly make a video about sub-agents
3:29:10as soon as you've made the diagram. Now,
3:29:12this saves a lot a lot of time because
3:29:14sometimes when I'm explaining a concept,
3:29:16I want to show a diagram. I want to show
3:29:18some visual representation so you can
3:29:20hear but also see at the same time cuz
3:29:22you retain more by doing both. So,
3:29:23that's where I use this. Now, let's say,
3:29:24"Can we also add some visuals using the
3:29:27Excalidraw visual skill?" Yeah, add some
3:29:30visuals there as well. There we go. So,
3:29:31now it's going to take the diagram. It's
3:29:32going to then add visuals and then we'll
3:29:34see the [music] difference as well. All
3:29:35right, so it just finished making the
3:29:36three different images. So, I can go to
3:29:38the Excalidraw and I can drag them
3:29:40across from here to here. And here we
3:29:42have the three different visuals. So,
3:29:44you see how this is a diagram,
3:29:47but this are images.
3:29:49That's exactly what the Excalidraw
3:29:51visual skill does. Which one's better?
3:29:53It kind of depends on you. Sometimes I
3:29:55like to have a mixture of both, but
3:29:56nonetheless, they saved me a ton of time
3:29:59whenever I have to make a video and make
3:30:00diagrams for that video. All right, so
3:30:02now that we're done with the video
3:30:03assets, let's go to the last one which
3:30:05is post-video data. So, here on Google
3:30:07Sheets, I have a list called the
3:30:09views-to-clicks, [music] which is out of
3:30:11this many views, I get this many clicks.
3:30:13And this is a CTR. And then we have the
3:30:14type of video. Now, the reason we do
3:30:16this is because we want to see out of
3:30:18this many views, how many people click
3:30:20to see the video of my one-to-one
3:30:22program to then be able to work with me,
3:30:23right? And we do this because typically
3:30:25the more value that you provide, the
3:30:27higher the CTR is, okay? So, in this
3:30:29case, if I have more clicks and I have
3:30:32more data as to what videos people like
3:30:34and people want, then I can provide a
3:30:35ton of value to you guys and then
3:30:37be able to help you guys even further,
3:30:38right? So, now you might be asking
3:30:40yourself, "How do you do this?" So,
3:30:42manually, the way that this works is
3:30:44that if I go to one of my videos, you
3:30:45can press this link. Now, this link, you
3:30:46can see Bitly. Bitly is a software that
3:30:48pretty much all the YouTubers
3:30:50that I know kind of use to be able to
3:30:52track how many people click to the next
3:30:54thing and then obviously you click to
3:30:56kind of watch the video of like the full
3:30:58one-to-one program how I help you to be
3:30:59able to start and scale your own AI
3:31:01agency, right? And so here I see how
3:31:03many people clicked out of this many
3:31:04views. So out of this many views, how
3:31:06many people press this button? And it's
3:31:08very tedious because I have to go to
3:31:10Bitly. I have to see the views and it
3:31:11takes me a lot of time especially if I
3:31:13have to do this weekly for this many
3:31:15videos and the videos just keep on
3:31:17adding and adding and adding, right? And
3:31:18so what we do here is I have a
3:31:20automation that I can do I can put
3:31:22{slash} views to clicks analysis.
3:31:25And what this will do is that it will
3:31:27look at my videos. So all my videos in
3:31:29my channel since a specific date where I
3:31:31started tracking. So now it's running
3:31:32the automation and now it's said that it
3:31:34updated the sheet successfully. So if I
3:31:36go back I can see that this is now all
3:31:38in different places and I can delete
3:31:40these two columns. I don't know why
3:31:41they're here. There you go. And I can do
3:31:43sort Z to A and I can see the CTR here.
3:31:45So these are the videos where you had a
3:31:47higher CTR rate, right? There's actually
3:31:49BTC at top. Looking at BTC at top. There
3:31:51you go. And I can start to see, hey,
3:31:52these many videos on the bottom they
3:31:53don't do too well in terms of the CTR
3:31:55but these videos do. And so let me make
3:31:58more of those videos and less of these
3:32:00videos. And that's one of the ways where
3:32:01we strategize what content idea works
3:32:03better and what content ideas don't work
3:32:05better based on my audience, my offer
3:32:07and who I'm targeting.
3:32:11Now let's take it one step further cuz
3:32:12in the final lesson I'm going to show
3:32:14you how you can build and deploy $10,000
3:32:17websites using no code. Let's dive in.
3:32:20So to give you like a full recap of high
3:32:21level what we did, we made an image that
3:32:23looks like this of this Ferrari F1. Then
3:32:25we made a video that looks like this
3:32:27where we have the components of the car
3:32:28breaking up to then make a website that
3:32:30looks like this where we have Born to
3:32:32Perform, Seven Decades of Fury, Force of
3:32:34Maranello. Then we go down
3:32:36you can see the car breaking up and
3:32:38going back in which is super super super
3:32:40sick. And by the way, this right here is
3:32:41a video. So we have the video
3:32:42incorporated in the websites. You can
3:32:44also see here
3:32:46prancing which is the the horse of the
3:32:48Ferrari. So below that we can see
3:32:49maximum speed. We can see 0 to 100 km/h,
3:32:53top speed here. We also have the final
3:32:54page, which is experience a legend,
3:32:56reserve now. So reserve the spot for the
3:32:58for the actual car. Okay, so that's a
3:32:59full website that we're going to build.
3:33:00So we're going to go through everything
3:33:01step by step. Now there's a lot of
3:33:02moving components for this exact
3:33:03websites cuz there's different places
3:33:04where we get data, make the image, make
3:33:06the video, which is why we're going to
3:33:07break it all down into a four-step
3:33:09framework. The four-step framework is
3:33:11called plan, execute, review, and
3:33:13deploy.
3:33:14All right, so the first one here is
3:33:16plan. Okay, so here's the product that I
3:33:17used to be able to create the website
3:33:19that I showed you. As you can see here I
3:33:20have the video of the Ferrari that we
3:33:22used inside the website. And
3:33:23essentially, to be able to start
3:33:25everything, we need two different
3:33:26things. We need something called the
3:33:27cloud.md file, and we need the second
3:33:29thing, which is two different skills. So
3:33:31first thing is the cloud.md file. Now
3:33:32what is it? This right here represents
3:33:34the definition quite in a in a very
3:33:36quick way, which is are you tired to
3:33:38repeat yourself? Tell Claude to remember
3:33:39what you've told it using the cloud.md
3:33:41file. So essentially, it's a way for us
3:33:43to give Claude memory of the project.
3:33:45Cuz every single time that we're talking
3:33:46to Claude, it needs to have context as
3:33:48what it is that we're doing. So this
3:33:49right here describes the product, which
3:33:51is a landing page for Ferrari. Then we
3:33:52have the file structure. So when it
3:33:54makes the landing page, where does it
3:33:56store everything? What folders? Where do
3:33:57the files go inside the folders? And it
3:33:59gives structure to the whole project.
3:34:00Then we have more things specific to the
3:34:02actual landing page, the design. So
3:34:04these are the different designs that we
3:34:05use, animations, and external
3:34:08dependencies, coding conventions, and
3:34:09constraints. Now a lot of the things
3:34:11that are here, I don't fully understand,
3:34:12right? Because I'm not super technical.
3:34:14But we use AI to craft the cloud.md
3:34:16file, which makes it easier for us to be
3:34:17able to do it. So once you have the
3:34:18cloud.md file, which gives Claude code
3:34:20the memory, then now we have the two
3:34:22different skills. So skills are
3:34:24basically ways for us to give Claude
3:34:26code a series of instructions to make
3:34:29something better. So for example, we
3:34:30need to make a landing page, and we need
3:34:32to make a video that fits well inside
3:34:34the landing page. So the first skill
3:34:35that we need is called front-end design.
3:34:37Now a skill, again, is just a list of
3:34:38instructions. So this skill right here
3:34:40tells Claude code, "Hey, here's how to
3:34:43create a nice-looking UI, so user
3:34:45interface, basically a landing page,
3:34:47which increases the output for the
3:34:49landing page that we're going to make."
3:34:50And so, every single time it makes a
3:34:51landing page, it looks at the skill, it
3:34:52looks at all the components, how it
3:34:54needs to think about design, the
3:34:55front-end guidelines, and so on. And so,
3:34:57it will use the skill when making the
3:34:59landing page. Now, your question might
3:35:00be, "How do I even get the skill?" But
3:35:02if you go to {slash} manage plugins,
3:35:05you'll be able to see the skill right
3:35:06here [music] called front-end design.
3:35:08You can press install, and you have
3:35:10three options. Either you install it
3:35:12available for all your products, which
3:35:13means that you don't have to install it
3:35:15again for any other project, or install
3:35:16it for this project in specific. But
3:35:18bear in mind that if you do that, every
3:35:20single time that you start a new
3:35:21project, you will not be able to see the
3:35:22skill, and you'll have to install it
3:35:24again. Once you've installed it, you
3:35:25should be able to see either here or
3:35:26here. Make sure you refresh your
3:35:28CloudCode, and you'll have the skill.md,
3:35:30which is the instructions itself. The
3:35:31second skill is called the
3:35:32video-to-website skill. This right here
3:35:34is not something you'll be able to find
3:35:36in the plugins. So, you can find it in
3:35:38the resources that I left below. But
3:35:40essentially, this gives CloudCode
3:35:42instructions as to how it should take
3:35:43the video and turn that into uh
3:35:46something that you can use within the
3:35:47website. And it's very, very long, as
3:35:49you can see here, list of instructions.
3:35:50So, once you have these two things, we
3:35:52are good to go to the next steps. Before
3:35:54we do that, I made a new project with
3:35:55the two different skills and a cloud.md
3:35:57to keep it very, very simple as you can
3:35:58follow along with me when I do this. Now
3:36:00that we're done with planning, we can go
3:36:01to step two, which is execute. Okay, so
3:36:03when it comes to execute, we have to ask
3:36:05ourselves, "What do we need to be able
3:36:07to make this landing page happen?" So, I
3:36:09firstly need an image, which then turns
3:36:10into a video. And then, also, one more
3:36:12thing that I want to use is a reference
3:36:13landing page so that we can feed it all
3:36:15into CloudCode, and it uses the
3:36:17reference landing page, it uses the
3:36:18video, and it uses the skills to be able
3:36:20to craft us the beautiful landing page
3:36:22that we're going to use. Now, to be able
3:36:23to make the image and the video, we can
3:36:24use a software called Key API, which is
3:36:27one of the best softwares that I've used
3:36:28cuz it gives us access to all the video
3:36:30models, image models in the fastest way,
3:36:31and it pretty much has all the options
3:36:33that you need. So, first thing is to
3:36:34make an account, and then you can go to
3:36:35market to be able to add your API
3:36:37credits cuz this is not free. It costs
3:36:39credits to run these images and videos.
3:36:41So, first thing is to make an image and
3:36:42I'm going to be using the Nanobots 2
3:36:44model, which is a new model that Google
3:36:46Gemini made because it's [music] fast
3:36:48and it's cost-efficient. And down here
3:36:50we can write the prompt that we need.
3:36:51So, I'm going to be copying a prompt
3:36:52that I already made. Here we go. Which
3:36:54says that I need a high-resolution
3:36:56studio-grade image of a Formula 1 racing
3:36:57car positioned against a completely
3:36:59plain black background. The scene should
3:37:01be a minimal and isolated with no
3:37:02shadows, reflections, hands, or
3:37:04additional objects visible. The car
3:37:06should appear sharply detailed and
3:37:07professionally lit as if photographed
3:37:08[music] for a premium automotive studio
3:37:10shoot. And then we can remove this.
3:37:13We can go down to 4K.
3:37:15Aspect ratio, we can do auto. Automatic
3:37:17is fine. And I can press run. Bear in
3:37:19mind that if you're wondering how much
3:37:20does this cost, well, you paid $5 for
3:37:221,000 credits, so it costs you 18 to be
3:37:25able to do it, which means that you're
3:37:26paying about $0.09
3:37:29to make the image. All right, so it just
3:37:30finished making the image. As you can
3:37:31see here, very very clean. And now we
3:37:33can download this and we can go to the
3:37:35next step, which is making the video.
3:37:36So, I can go to AI video API. I can go
3:37:38to Kling 3.0. There's also other image
3:37:41or video models that you can use, but
3:37:42this one is by far the best that I've
3:37:44seen. I can take this out and I can drag
3:37:47the image across.
3:37:48Boom.
3:37:49So, this right here is the input image.
3:37:51I can take out the sound. We're not
3:37:53going to add any end frame. So, let's
3:37:54say you're making a a website about a
3:37:56cup, right? Just some sort of cup. You
3:37:58would have a cup with no water and then
3:37:59a cup with water. That's a start frame
3:38:01and then end frame. So, you can
3:38:03basically see the animation of the water
3:38:04going inside the cup. But in this case,
3:38:06we don't need it. And now when we get to
3:38:07the prompt, we can go to ChatGPT. I can
3:38:10paste this prompt right here, which says
3:38:11that it's an AI video prompt engineer.
3:38:13You can stop this video and read it, but
3:38:15I'm going to then give it the image of
3:38:17the Formula 1 racing car as an input.
3:38:20And then I'm going to press run. And now
3:38:21it's going to give me the full prompt
3:38:22that I'll need to be able to make the
3:38:24video.
3:38:25Now, one thing, I do like this prompt. I
3:38:27just think it's a bit long, so I can
3:38:28say, "Can you shorten this
3:38:31a little bit and make make it more
3:38:33concise.
3:38:34But sometimes when you provide too much
3:38:36context or when the prompt is too long
3:38:37for an AI, it does a worse job. Okay,
3:38:40cool. That's fine. So, I can go back
3:38:41here.
3:38:42We have the clean 3.0. I can paste this
3:38:44prompt, and now the duration will be 5
3:38:47seconds. That's fine. And I want to make
3:38:48sure that the mode, so the actual
3:38:50resolution, is higher. So, I can press
3:38:52pro cuz the resolution is basically how
3:38:54clean or how nice does the video look.
3:38:56We can press run now. As you can see,
3:38:57this is taking 135 credits, which is
3:39:00about $0.625 uh
3:39:02that we're spending on the video. All
3:39:03right, so I just finished making the
3:39:04video. Let's see the video here.
3:39:08Ooh, it's even better than the one I had
3:39:10before. Cool. So, now we can download
3:39:11this.
3:39:12Save this to your desktop. I can go back
3:39:14to Cloud Code.
3:39:15I can then give it the video
3:39:18as the input. Okay?
3:39:19And I can rename this
3:39:22Ferrari
3:39:23video.
3:39:24There we go. Now that we have the video,
3:39:26there's one more thing that we have to
3:39:27do, which is getting a reference landing
3:39:29page. Because yes, we can make a landing
3:39:31page, but if we give it stricter
3:39:32guidelines, it does a better job. So, if
3:39:34you go to dribbble.com, you can find the
3:39:36Charger product landing page right here,
3:39:38which is a very, very nice landing page
3:39:39that we can use as reference. So, you
3:39:40can go here, you can save video as
3:39:43reference.
3:39:45I can press save, and then I can drag it
3:39:47across the Cloud Code as well. So, now
3:39:48we also have the video
3:39:50of the landing page as a reference.
3:39:52Finally, I can go to Cloud Code. Make
3:39:54sure this right here is in plan mode,
3:39:55and then you can paste this here, and
3:39:56you will say, "I just dropped in a
3:39:58video, which is the Ferrari
3:40:02What is Ferrari video?
3:40:04There we go.
3:40:04Create a one-product landing page for
3:40:06this product. The site must feel modern,
3:40:07premium, and highly professional with
3:40:09smooth animations and polished UI design
3:40:10throughout. All text should be clean,
3:40:12minimal, and easy to read using
3:40:14high-contrast typography on a pure black
3:40:16background dark mode. The layout and
3:40:18visual styling should be seamlessly
3:40:19blend with the Ferrari image, making the
3:40:21entire page appear continuous and fluid
3:40:23with the background. And then, I will be
3:40:24able to say, "I also added a video,
3:40:27which is a reference video that I want
3:40:28you to look at in terms of the actual
3:40:30design of the landing page itself. And
3:40:32now I can press go. All right, so as
3:40:33Clock Code is actually making the plan,
3:40:34you can see that it's reading the
3:40:36skills. So, skill two, skill three when
3:40:39making the plan itself. That's where
3:40:41we're using skills. All right, so it
3:40:42just finished making a plan. Context
3:40:43streaming extraction, uh file structure
3:40:45design tokens, page architecture. And
3:40:47you're looking at this and you're like,
3:40:48"What the hell even is this?" The same
3:40:49here. So, don't worry. JavaScript
3:40:51initialization order. I'm wondering why
3:40:53they have to make such complex words for
3:40:54all of these. So, [music]
3:40:55let's go and press yes and auto accept.
3:40:58And then we can turn this to bypass
3:41:00permissions because I don't want to be
3:41:02approving every single thing. Bear in
3:41:03mind that bypass permissions is a bit
3:41:04more dangerous because it takes over and
3:41:06does everything, but I usually use it
3:41:08when I'm actually looking at Clock Code
3:41:09myself. I never go to sleep and then let
3:41:11Clock Code do its thing and then come
3:41:13back, right? All right, so it just
3:41:14finished making version one of the
3:41:15landing page. And [music] bear in mind
3:41:16that on the left-hand side, you will be
3:41:18able to see new folders that were
3:41:19created. And if I go to the Clock Code
3:41:21file that we had before, we had CSS
3:41:23folder, we had JS folder, and we had
3:41:25frames folder. And here you can see CSS
3:41:27folder, frames folder, and JS folder.
3:41:29So, that's what we mean by file
3:41:30structure. Where are the files going to
3:41:31be stored? CSS is going to be all the
3:41:33different styles
3:41:34>> [music]
3:41:34>> uh that we're going to use, so all the
3:41:35colors and so on. I'm only saying this
3:41:36because a lot of you guys are beginners
3:41:38and are getting into this and don't
3:41:39really know what all this code stuff is.
3:41:41On a fundamental level, that's
3:41:42essentially what it is. Putting files in
3:41:43the right folders and each one doing a
3:41:45specific thing. You don't have to know
3:41:47what the code actually is. So, I can
3:41:49copy the link and I can paste it in my
3:41:50browser.
3:41:55That's nice. Born to perform. So, we can
3:41:57see we have the the little page before
3:41:58we get into the website. We have this
3:42:00here, the animation, the car being
3:42:02broken into different pieces,
3:42:05and then coming back in.
3:42:06Form follows function. Okay.
3:42:09The speed, maximum power.
3:42:12Just sick. So, a bit of writing giving
3:42:14us some context. And that's it. To be
3:42:16fair, it did actually a very good job
3:42:18for it to be like a first time uh that
3:42:20of thing. All right, so now we can go to
3:42:21the next step, which is called review.
3:42:23Now, reviewing it means that we have to
3:42:25look at the website and say, "Okay, do
3:42:27we want any changes with the way that
3:42:28it's laid out?" I can already see a few
3:42:30changes that I want. So, for example, I
3:42:31want some something happening in the
3:42:32background here. So, some sort of lines
3:42:34or or dots. I have a vision of it, but I
3:42:36have to see. Then, right down here, I
3:42:38would want the card to be positioned on
3:42:40the right-hand side, and then the
3:42:41writing to come only from the left-hand
3:42:43side, not the right-hand side. Cuz when
3:42:44it goes over the card, it doesn't look
3:42:45nice. So, we can take a change by
3:42:47change. The first change is having
3:42:49something within the background here.
3:42:51So, I can go to 21st.dev, which is a
3:42:52software that we use to pull bits and
3:42:54pieces of different website animations,
3:42:56motion graphics, and so on, that we can
3:42:57use for our website. So, I said I wanted
3:42:59something in the background. So, what we
3:43:00could do, can look here, shaders.
3:43:03We can try this one, background paper
3:43:05shaders.
3:43:06We can copy the prompt.
3:43:07And I can go back to Clock Code. So, we
3:43:09can see we're using 54% of the context.
3:43:11And by the way, the more context you
3:43:12use, the worse the quality gets. So, I
3:43:15can put clear conversation, go to bypass
3:43:17permissions, and I can say, "Okay, so
3:43:18I'm looking at the website, and I want
3:43:20you to implement the following thing at
3:43:23the start of the website." And I can
3:43:24paste the prompt. So, essentially,
3:43:2621st.dev, to be able to get this, we
3:43:28have the prompt that allows Clock Code
3:43:30to understand exactly what it needs to
3:43:31do and how it needs to update the code.
3:43:33And as you can see here, it's a full
3:43:34full thing. I can press
3:43:36bypass permissions, [music] and we can
3:43:37see. Because, honestly, I just wanted to
3:43:39update, and that's it. It just finished
3:43:41implementing the change. So, if I go
3:43:42here to the website, I can refresh, and
3:43:45you can see now in the background is
3:43:46very, very subtle, which I really like.
3:43:48But, if you zoom in, you can see that we
3:43:50have different things changing in the
3:43:52background, which makes it a lot nicer
3:43:54to to look at visually. Now, that's the
3:43:55first thing. The second thing is putting
3:43:57the card to the right-hand side, and
3:43:59having the text go on the left-hand
3:44:00side. Okay, so, second thing that I want
3:44:02you to change is on the website itself,
3:44:04the card takes up the whole space. What
3:44:07I want to happen instead is I want the
3:44:09card to be on the right-hand side. So, I
3:44:11want it to take like, basically, 75% of
3:44:14the whole screen from the right to left,
3:44:1675%. And in the other 25% that's left on
3:44:19the left-hand side, then I want that to
3:44:21be the space where text comes through
3:44:22because I don't want text to overlay
3:44:24over the car. Does that make sense?
3:44:26>> [music]
3:44:26>> Keep the animation, keep the video, but
3:44:27make the car smaller and position it on
3:44:29the right-hand side of the screen. I'm
3:44:30going to go on plan mode for this one
3:44:32just because I want to make sure that it
3:44:33understands
3:44:34>> [music]
3:44:34>> what I'm talking about and a lot of this
3:44:35is just talking to Claude going back and
3:44:37forth as to what it is that you want.
3:44:38All right, so here it made it the plan
3:44:40for the change, the context
3:44:42implementation plan, 25% to the left,
3:44:44okay. Good. Moving everything to the
3:44:45left, the car to the right, everything
3:44:48looks good. So now I can press yes and
3:44:49no to accept and we can go back to
3:44:52bypass permissions to have Claude code
3:44:53make the changes. All right, so now it
3:44:54just finished making the change. I can
3:44:56refresh and we can see that all the text
3:44:58has been on the left-hand side and then
3:45:00this also is fine. Um, I can go down and
3:45:02as I scroll through we can see that all
3:45:04the text is on the left-hand side, which
3:45:06looks a lot better than us having to
3:45:07have the text over the car and that
3:45:09covers the whole website step-by-step.
3:45:11Now that we're done with step three, we
3:45:12can go to the last step, which is
3:45:13deployment. So the way that we deploy
3:45:15Claude code websites is through a
3:45:17two-step process. The first thing that
3:45:18we do [music] is we push all the Claude
3:45:20code into GitHub and then from GitHub we
3:45:22deploy it to Vercel.
3:45:24So the first thing we want to do is go
3:45:25to github.com
3:45:27and want to make an account. Once you
3:45:29made an account, you can press new,
3:45:30we're making a new repository and we can
3:45:32name this website Claude code.
3:45:36We can press create repo. And so
3:45:38basically what's happening is that we're
3:45:39going to be using Claude code and we're
3:45:40going to be giving all the code here
3:45:42into GitHub and from GitHub it's going
3:45:44to go to Vercel to then deploy it there.
3:45:46All I have to do here is press copy and
3:45:48I can go, I can clear the conversation.
3:45:51I can say, [music] "Okay, now I want to
3:45:53push this to GitHub. Here's the GitHub
3:45:55link repo that I've just made." I can
3:45:57paste this and I can go bypass
3:45:59permissions cuz I know it will do a good
3:46:00job cuz it's a pretty fast task. It
3:46:01should tell us to log in into GitHub cuz
3:46:03I didn't actually give it any
3:46:04credentials here. So you can see here
3:46:05GitHub needs your identity configured.
3:46:07Run these two commands in your terminal
3:46:08replace with your actual name and email.
3:46:10I don't know how to do that. Can you
3:46:11tell me how I should do it, What I
3:46:12should give you to make that happen? I
3:46:13do this just cuz a lot of you guys don't
3:46:15know how to use a terminal, so that's
3:46:16fine. But in this case, what I'm going
3:46:18to do is give my name and then the email
3:46:21that I'm using.
3:46:25All right, now it's saying that it's all
3:46:26live. So, if I go here, I can refresh
3:46:28and I'll have all the different files
3:46:30and folders of this product. They are
3:46:32automatically added here. And the last
3:46:34step here is to deploy on Vercel. I can
3:46:37press start deploying. I can continue
3:46:39with GitHub. And once you connect to
3:46:40GitHub, make sure to choose the correct
3:46:42product that you're working with. In
3:46:43this case for me it was website clock
3:46:44code. I can import and I can press
3:46:46deploy. And as you can see here, we now
3:46:48we have our product, so our website
3:46:51deployed. And now it says
3:46:52congratulations, you've just deployed a
3:46:53new product to make a lot of past
3:46:54projects. I can press continue to
3:46:56dashboard. And if I use this link right
3:46:57here, you can see that I have the
3:46:58website itself, which is sick.
3:47:01Boom.
3:47:06There we go. And this literally took 2
3:47:07minutes to deploy, which is absolutely
3:47:09insane. And now for the final thing is
3:47:11answering the question, what if you want
3:47:13changes? What if you're here and you
3:47:14want to change the website? Do you have
3:47:16to go to Vercel? Do you have to go to
3:47:17GitHub? Well, the beautiful thing about
3:47:19GitHub and Vercel and the whole process
3:47:21is that you just have to go to Clock
3:47:23Code. You have to tell Clock Code what
3:47:25you want to change and that will
3:47:27automatically update the GitHub, which
3:47:29will automatically update Vercel. And
3:47:30that's the whole flow of things. To make
3:47:32sure that you don't have to go to two
3:47:33different platforms at the same time and
3:47:35do any extra stuff. You can just talk to
3:47:36Clock Code simply and it will
3:47:37automatically do everything for you.
3:47:39All right guys, so congrats on finishing
3:47:41the full Clock Code masterclass where
3:47:43you went from a complete beginner in
3:47:45Clock Code all the way to building
3:47:47magenta workflows, deploying them and
3:47:50using Clock Code to build actual
3:47:51products from scratch. Now, everything,
3:47:53all the resources and templates of this
3:47:55video are going to be stored in the
3:47:56second link down below in my private
3:47:58community. You can go to the classroom
3:47:59section, go to the templates vault and
3:48:01you'll find everything there. All right,
3:48:02so now you've watched how to build with
3:48:04Clock Code. Your Your question might be,
3:48:06how do I build a business with it? Which
3:48:07is why I put together a full AI agency
3:48:09masterclass that you can check out on
3:48:10the screen, which goes through
3:48:12everything that you have to know as a
3:48:13complete beginner to make money with AI.
3:48:15With that being said, I really hope you
3:48:17found value from this video, and I'll
3:48:19see you in the next one.

About this transcript

This page contains the full transcript of CLAUDE CODE MASTERCLASS 4 HOURS: Build & Sell (2026) by Michele Torti, generated from the public captions YouTube serves with the video. The transcript has 55,900 words across 7,917 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.