YT YouTube2Text

YouTube2Text → Transcripts → Claude Code for Non-Coders (6 Hour Course)

Claude Code for Non-Coders (6 Hour Course) — Transcript

by Nate Herk | AI Automation · 85,931 words · 11,730 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00In this course, I'm going to take you
0:01from a complete beginner to someone who
0:02is AI native and can build automations,
0:05AI agents, literally anything that you
0:07can describe, you're going to be able to
0:08build by the end of this course. These
0:10are the topics that I'm going to cover
0:11with you guys. So, feel free to skip
0:12around to whatever pies your interest,
0:14but I'm going to go through all of this
0:15in order, and I'm going to do real
0:17examples and step-by-step builds
0:18throughout. So, don't want to waste any
0:20time. Let's just jump straight in. So,
0:22this whole course is basically assuming
0:23that you know nothing and that you don't
0:24have a technical background, which if
0:26you guys don't know who I am, my name is
0:28Nate and I do not have a technical
0:29background. But, I've been able to do
0:31some pretty incredible things with AI in
0:32the past couple years. I've got multiple
0:34different businesses that are based on
0:35content, education, certifications,
0:37events, consulting, and all of that is
0:40powered by a small team who is really,
0:42really good at using AI. The whole idea
0:44is that one person can do the work that
0:46it used to take teams. And I felt that
0:49ROI myself. And if you guys haven't, by
0:50the end of this video, you will have a
0:52very clear path to feeling that ROI and
0:54having AI systems that help you do more
0:56and that will eventually become the
0:58baseline. So, it's great to be getting
0:59ahead of this stuff. Okay, so first
1:01things first, what is Claude Code?
1:03Because I know the word code probably
1:04intimidates a lot of people. If you've
1:06never coded before, you don't have to
1:08code. So, Anthropic, that's the company
1:10that is behind this ecosystem, right?
1:12Um, the cloud models, Claude Chat, that
1:15is probably what a lot of you guys are
1:16used to. So, there's basically these
1:18three products. The first one is Claude
1:20chat and that is where you can talk to
1:22Claude in the web or wherever you want
1:23to do it on your phone and you just talk
1:25to that large language model the AI
1:27model and you get an answer. So kind of
1:29like a a chatbot. Then we've got Cloud
1:31Co-work which is kind of like the next
1:33step up. Fun fact, Cloud Co-work the
1:35entire product was built using cloud
1:37code by a few engineers. It would have
1:39taken the team weeks and weeks and much
1:41larger team but they were able to ship
1:43this in like a week with a few
1:44engineers. So very very cool. And cloud
1:46co-work is much more for kind of like
1:48your knowledge workers, your managers.
1:50It's much more of a simple interface to
1:52do simple automations and stuff, but
1:54cloud code is just the most powerful.
1:57And the cool thing about cloud code is
1:58like I said, you don't need to know how
2:00to code at all. And so I really don't
2:02ever use cloud co-work even though you
2:04could argue a lot of the stuff that I do
2:06in cloud code could be used in co-work.
2:08code is just way more powerful and it's
2:09where I see the future as far as
2:11thinking about building your own agents
2:13and just having things that are a lot
2:14more agentic for you. So those are kind
2:17of the three products, but what's cool
2:18about it is they're not that different
2:20to actually use. So if you learn one
2:22really well, you're going to be able to
2:23transfer those skills wherever you go.
2:24So don't feel like you're locking in by
2:26choosing one of these products. That's
2:28why I choose the most powerful product.
2:30So if I open the Claude desktop app,
2:32what do we see right here? This is
2:33Claude chat. You know, I can come in
2:34here and I can say, "Hi, Claude." and it
2:37can kind of understand a little bit
2:38about me. It can see my preferences.
2:40Obviously, it knows my name, but this
2:42isn't where you're going to get real
2:44work done. This is where people are kind
2:46of treating this as like a Google search
2:47engine or they're connecting their Gmail
2:49and helping it, you know, having it help
2:51them write email drafts and stuff like
2:52that, but it's not true agentic power.
2:55And this is Claude Code. What do you
2:57see? It's a very similar interface. You
2:59just talk to it in chat, but now what
3:00happens is it's able to access your
3:02local files. As you can see here, I can
3:04choose a folder to actually work within,
3:06whether that's my desktop or my
3:07downloads or my Herk 2, which is kind of
3:09my AI operating system, which is a term
3:11I might throw out a little bit today.
3:13Don't worry about it too much. At the
3:14end of this video, I'm going to show you
3:16a full course around how to actually
3:17build your own AI operating system, but
3:19this comes first. The whole point here
3:21is that Claude Code is able to work
3:24inside of local files as well as touch
3:26anything out there online. It can touch
3:28your Gmail, your Slack, your CRM. It can
3:30touch anything online, but it can also
3:32touch your local files, which is really,
3:33really important. And the other cool
3:35thing to note is that when you're in
3:36cloud chat, you know how you guys
3:38probably are aware of models like Opus
3:40or Sonnet regardless of the model or
3:42even Fable. Now, Claude Code is powered
3:44by all of those same models. So inside
3:46of cloud code, you still use sonnet or
3:48haiku or opus or fable, but the code
3:51aspect is basically just enthropic
3:53saying, "Hey, because you're working
3:55with local files, here are some other
3:57things you can do like web searches and
3:58web fetches and searching through local
4:00files that make you a little bit more
4:01powerful than cloud chat." So
4:03fundamentally, it's not that different
4:05from cloud chat. I don't ever use cloud
4:07chat anymore. Everything that I would
4:08typically do in Cloud Chat, I'm doing in
4:10Cloud Code because Claude Code knows
4:13everything about me, my business. It can
4:15see all my YouTube videos. It can see
4:17all of my Slack threads. It can see
4:18everything that I'm doing, which makes
4:20this feel way more like an employee
4:22rather than just like Claude chat, which
4:24kind of feels like a virtual assistant,
4:25right? I'm not reexplaining myself to
4:27Claude Code because it knows everything
4:29about me. So, you're probably going to
4:31hear the term harness, magenta harness,
4:33AI harness. That's what Claude code is.
4:36And this is kind of the way that I think
4:37about how you build with AI. This three
4:40sort of layer circle. At the core, we
4:42have the AI model. So, Opus 4.8 or Fable
4:45or GPT, whatever model you're using,
4:46that's at the core. Then around that, we
4:48have the AI harness. So, cloud code is a
4:51harness that uses AI opus or other cloud
4:55models to power it. And then on top of
4:57that harness, you have you. So, your
5:00brain, your prompts, your context, your
5:02data, your business. That's how it
5:04works. So if we make this a little bit
5:06more practical with analogy here, we've
5:08got the model here. We've got a car and
5:10then we've got you. So opus is a model,
5:13right? But without this is basically the
5:15engine. So we put this inside. If I can
5:18make the layering right, we put this
5:20basically inside of the car. Wherever
5:22you think the engine in this car might
5:24be, let's just say it's here. Then this
5:27is still not going to really be that
5:28useful. It's not going to be able to
5:29drive or go anywhere without the human.
5:31So the human has to come in here and
5:33drive the car. So the human is steering
5:35the harness and the harness is powered
5:36by the AM model. That might make no
5:38sense to you at this point. That's okay.
5:40I just wanted to sort of lay out these
5:42terms because you will hear these terms
5:44the more you get into AI. So just a
5:46simple way to think about it because as
5:48you continue to evolve and as the tools
5:50continue to evolve, you might switch out
5:52the model. So maybe you'll be using
5:53cloud code, but eventually you might
5:54switch this out for a different model or
5:56maybe you'll be using the model, but
5:58you'll switch out a different harness.
5:59Maybe you want to get in a different car
6:00and just drive it in a bit of a
6:02different way. So they're
6:03interchangeable but at the core this is
6:05what we're looking at AI harness and you
6:08and really the most important thing in
6:09this whole picture is you. If you are
6:13not giving it the right context then
6:14it's not going to be able to do the
6:16right stuff for you. So that's what it
6:17looks like. Okay. So that is what cloud
6:19code is. Now before we start getting
6:20into some of the technical stuff of
6:22getting set up and you know building
6:23stuff I wanted to talk about the mindset
6:25because the mindset around this stuff is
6:27so important for many reasons. I think
6:29the first one is that you know the space
6:30changes so fast. new models, new tools,
6:33new headlines. The space changes fast.
6:35So, how do you learn things in a way
6:37where it's not outdated next week? And
6:39that's what I want to teach you guys
6:40today. That's so important, in fact,
6:42that I actually wrote a book about this.
6:44It's called Becoming AI Native. And in
6:46this book, I go over kind of like
6:48throughout the chapters, these 12 big
6:49mindset shifts. And so, throughout this
6:52course, I'll probably relate back to
6:53some of these mindset shifts a little
6:54bit. But, I wanted to talk a little bit
6:56about mindset before we jump into the
6:57weeds of everything today. But before we
6:59hop into those mindset shifts, I want to
7:01talk about real quick with you guys. Six
7:03AI skills that I think every single
7:04person needs to master just to basically
7:06future proof their careers. AI is real
7:08and it's not going away. Just like
7:10social media replace newspapers and
7:12billboards and Netflix replace cable, AI
7:14will change and replace millions of
7:15jobs, including yours, unless you learn
7:17six important skills. These six AI
7:19skills will futureproof your career so
7:21that you don't have to try and start a
7:22business if you don't want to or switch
7:23careers if [music] you don't want to.
7:24So, these six skills are going to apply
7:26no matter what job title you hold or
7:28whatever career you're pursuing. And I
7:29can guarantee you the last skill in this
7:31video will surprise you because it's the
7:32most unique skill, but it does work. So,
7:34let's just get started with skill number
7:36one, the AI person. So, becoming the AI
7:38person. I know that that sounds like
7:39super obvious because this is an AI
7:41video, but I think that a lot of people
7:43misunderstand what that actually means.
7:45If you're watching this, you probably
7:46feel like you're on the beginning side
7:47of AI. And honestly, you probably are
7:49because there's so much happening right
7:51now. There's always new models. There's
7:52always new tools. There's always new
7:53agents and workflows, new benchmarks,
7:55all of this kind of stuff coming out
7:56every single week. But I bet if you went
7:58and you talked to [music] your friends
8:00or your family or just a lot of the
8:01people that you work with, they probably
8:02already think of you as an AI person. Or
8:05if they don't, you want them to. And
8:07that's the point. Being the AI person is
8:08relative. It doesn't mean that you're
8:10the best AI engineer in the world or
8:11that you understand [music] every single
8:13AI model and the architecture behind it.
8:15You guys probably think of me as an AI
8:16person, but in reality, I feel
8:18overwhelmed every single day by how much
8:19I still don't know. But just remind
8:21yourself it's all relative. It just
8:23means that inside of your circle you
8:25know more than the other people and that
8:26matters way more than you might think.
8:28I've seen this happen over and over in
8:29my communities. People start picking up
8:31AI almost like a hobby. So they're
8:32playing with Claude, maybe they're
8:33testing out codecs, maybe they're
8:34messing around with Google V3, building
8:36little tools or little AI side projects,
8:38automating parts of their job, just
8:40running a bunch of experiments. And then
8:41what happens is they start showing
8:42people and that's usually all it takes.
8:44They show someone at work, hey, look at
8:46this thing I built this weekend. Or hey,
8:47I used Claw to clean up this process. or
8:49hey, I figured out a way to make this
8:50task, which normally takes me three
8:52hours, only take me 20 minutes because
8:53of AI. And all of a sudden, that person
8:55becomes known as the AI person. And the
8:57reason this is so powerful is because
8:59companies are about to have a ton of
9:01like AI moments. They're going to get
9:02access to a new model or they're going
9:04to spin up a new internal product or
9:05maybe they're going to want to build a
9:07little AI task force internally. [music]
9:08And when that happens, someone's going
9:10to say, "We need someone to lead this."
9:11You want someone in the meeting to say,
9:13"Actually, I know someone who's really
9:14into this stuff." Or something like, "We
9:16should ask Nate. you know, he's told me
9:17that he's been playing around with all
9:18these AI tools. That's how these
9:19opportunities open up before there's a
9:21formal job title. And that's how people
9:22get pulled into better projects. That's
9:24how people become more valuable without
9:25quitting their job and starting from
9:27[music] zero. And the data backs this
9:28up. So IBM's 2026 CEO study found that
9:3185% of CEOs said that all functional
9:33leaders have to become technology
9:35experts in their own domain. Not just
9:37the CTO, not just the engineers, not
9:39just the IT team, everybody. So whether
9:41you're in marketing, sales, finance,
9:43ops, legal, customer success, whatever
9:45you're in, this applies to you. And the
9:46thing to remember here is that it's not
9:48just in one department. It's not like
9:49cyber security where you kind of have
9:51like a new team and then you're done.
9:52It's going to seep into every single
9:54role and every vertical whether you like
9:55it or not. And I know some of you are
9:57probably thinking, "Okay, but my role
9:58doesn't really need AI and that's your
10:00mistake. Your role will need AI because
10:02every single role will." So instead of
10:04trying to switch careers or learn a
10:05completely new job, just take a look at
10:07your role. Take a look at what you
10:08already do and ask yourself, how do I
10:10become faster and better and more useful
10:13at this specific task and this specific
10:15role with AI? So, I mean, think about
10:16it. If you were an accountant when Excel
10:18came out and you basically just said,
10:19you know what? I'm good. I like the way
10:20I do this. I'm going to keep doing all
10:22of this on paper and with a calculator,
10:24you were done. Like, that was probably
10:25your last day at that company, if you
10:27said something like that, if you made
10:28kind of like that stubborn statement.
10:30The people who learned Excel first were
10:31just faster. They got through
10:33spreadsheets in a fraction of the time
10:34that it used to take them. Let's just
10:35say from two spreadsheets a week to 10.
10:37That's just random numbers, right? But
10:38that new level of output became the
10:40baseline. And AI is like that, but it's
10:42a lot bigger. Cuz right now, being the
10:43AI person feels like an edge, but in a
10:45few years, it's just going to be the new
10:46normal. So, the advantage isn't waiting
10:48until someone's forced to learn it. The
10:50advantage is becoming the new normal.
10:51While most people still think it's
10:53optional. So, practically, how do you
10:54actually do this? Well, I would say you
10:56pick one main AI tool and actually get
10:58pretty good with it. So, as of May 2026,
11:00recording this video, for me, that's
11:01Claude. And I use Claude for my general
11:03knowledge work and to build automations.
11:05But the exact tool isn't the point. The
11:06point is, you need one tool that you're
11:08not just messing around with. You're
11:09using a tool to actually deliver some
11:11sort of ROI. And then you can pick one
11:13workflow in your current job. You can
11:14take something that you already do every
11:16week. And then figure [music] out how
11:16you can use AI to make it better or
11:18faster. Document what changed, how long
11:20did it take before, how long after, what
11:22got better, what still needed human
11:23judgment. Now, obviously, be smart.
11:25Like, don't expose company data or break
11:26any regulations. Don't try to automate
11:28something at work without getting like
11:30permission about it. And if your company
11:31has guidelines and you can't use Claude,
11:32then that's actually another opportunity
11:34for you to [music] look at because you
11:35can go dig into what those regulations
11:37are and you can figure out which tools
11:38you actually could use. And that alone
11:40shows that you care. But I just want you
11:42guys to remember that you don't need to
11:43[music] change careers. You can just
11:44find the AI native version of the career
11:46that you already have. So that's skill
11:48number one, become the AI person. But
11:49skill number two is where a lot of
11:51people are going to mess this whole
11:52thing up because the more you use AI,
11:54the more tempting it gets to just trust
11:55the output and say like that's good
11:57enough. And that's where skill number
11:58two comes in, which is taste and
12:00judgment. As AI gets better, it gets
12:02easier and easier to just trust the
12:03first thing that it gives you. And I saw
12:05this joke the other day that was on one
12:06side, we had a person with a bullet
12:08point and they were using AI to turn
12:10that bullet point into like a super
12:11professional structured email to send to
12:13the team. And then on the other side, we
12:14had that team and they were using AI to
12:16turn that structured email into just one
12:18bullet point. Now, it's funny, but it
12:20does also kind of create the perfect
12:22image of where work is going if we're
12:23not careful. You know, everyone's
12:24transforming things, but is everyone
12:26reading it? And that's dangerous because
12:28when you first start using AI, you
12:29review everything, right? Like you read
12:31every word, you double check the claims,
12:32you make sure it sounds like you, but
12:34the outputs start getting pretty good
12:35and then you get more comfortable with
12:36it and you just kind of let your guard
12:38down and that's the trap. And sometimes
12:39the giveaway is really small like
12:41[music] m dashes. You know, AI is
12:42notorious for putting m dashes in
12:44everything because it's been, you know,
12:45trained on so many white papers and
12:47formal documents. And so like for me,
12:49I've basically never manually written or
12:51typed an M dash in my entire [music]
12:52life. So if something goes out for me
12:54with five m dashes in it, people who
12:55know me are going to look at that and be
12:57like, "Okay, Nate obviously didn't write
12:58this. This is AI." And the problem is
13:00the second they think that it changes
13:01the way that they interpret that entire
13:03message. They start wondering, "Did this
13:05person actually read this? Is any of
13:06this true? How much of this is actually
13:07them?" And that's where taste comes in.
13:09And just to be clear, I'm not saying
13:10that I don't use AI to write or that you
13:11shouldn't. I think everyone should. It's
13:13just about taste. But anyways, this
13:14issue comes up for me all the time with
13:16video. So AI helps me make motion
13:18graphics that are honestly way better
13:20than what I could do by hand. They're
13:21way faster. They're cleaner, they're
13:22more polished, but it doesn't always get
13:24right where the motion graphics should
13:26come in or how long they should stay on
13:27screen or the visuals and like what's
13:29explaining and if it's distracting or if
13:31it's helpful. And that is still my job
13:32to watch the whole video back and give
13:34feedback. And that's going to be true in
13:35every field. AI can write the sales
13:37email. You still need to know if it'll
13:38annoy the prospect. AI can draft you the
13:40HR memo, but you still need to know if
13:42it'll make the employees feel weird. So,
13:43how do you actually build this skill?
13:44First, you should study the best work in
13:46your field. If you're in sales, study
13:47great sales emails. If you're in
13:48marketing, study great landing pages.
13:50Next, start saving examples. So, build a
13:52library of stuff that you actually like
13:54and stuff that sounds like you. And when
13:55something's good, don't just say, "Cool,
13:56that's good." And just copy and paste it
13:58somewhere else. Ask it why. Ask what
13:59made it good and ask what makes it clear
14:01and ask what makes it trustworthy and
14:03tell it why you think it's good and tell
14:04it what you like about it. Third, every
14:06time you correct AI, you feed that
14:08correction back into the system. So, the
14:09feedback loop is for good things, but
14:11also for bad things. If AI writes
14:12something and you change five things,
14:14say, "Hey, here are five things that I
14:16changed. Here's why. Update your
14:17instructions so that next time it's
14:18closer." And that's how you actually
14:20train the system to better understand
14:21your [music] taste. Because at the end
14:22of the day, AI can generate the work.
14:24Taste is deciding what deserves your
14:26name. Because remember, if you produce
14:27something with AI, your name is signed
14:28to it. Whether that is something that's
14:30really, really good and the whole team
14:31loves it, you [music] will get credit
14:33for it. Or if it's something that's bad,
14:34you will take the blame. It doesn't
14:36matter if AI wrote it, doesn't matter if
14:37you wrote it for that piece of [music]
14:38work because your name is assigned to
14:39it. So now the question is, how do you
14:41actually get AI to produce better work
14:42in the first place? Because if you're
14:43just typing prompts and just [music]
14:45crossing your fingers, then you're
14:46leaving a lot on the table. So that's
14:47skill number three and it's something
14:48that you can apply the second you close
14:49out of this video. And that skill is
14:51becoming a context engineer. So you
14:52might have heard the term prompt
14:54engineering. This was a huge thing a
14:55couple of years ago. The whole idea of
14:56prompt engineering was that if you
14:58wanted a better output from an AI model,
14:59you had to give it a good prompt. You
15:01had to give it a role, clear
15:02instructions. You had to tell it the end
15:03state. You had to give it examples. You
15:05had to tell it what to do and what
15:06explicitly not to do. But prompt
15:08engineering is getting less important
15:09over time because the models are just
15:11getting so much better on their own.
15:13Even Andre Karpathy, who's one of the
15:14goats of AI and actually just joined
15:16Enthropic, called context engineering
15:18the delicate art and science of filling
15:20the context window with just the right
15:22information. So translation in layman's
15:24terms, prompts are how you ask. Context
15:26is what your AI actually knows. In
15:28context engineering is way more durable
15:29than prompting because no matter how
15:31good the models get, they still need to
15:32know what's actually in your brain. So
15:34what's going on in your business, what's
15:35on your calendar, what your priorities
15:36are, stuff like that. So here's my
15:38personal example. I've built what I call
15:39my AI operating system or my AI OS. And
15:42[music] basically, it has pretty much
15:43all the context that's in my head. It
15:45can see my meeting transcripts. It can
15:46see all my YouTube videos. It can read
15:47through my DMs and channels in ClickUp
15:49and Slack. It can pull my emails. It
15:51honestly knows what's going on in my
15:52world better than I do because it can
15:54recall everything instantly and
15:55perfectly. And I can't do that. So, it's
15:57kind of like this running joke that if
15:58someone couldn't get a hold of me, they
16:00should just message my AIOS and it would
16:01actually give them an answer that's
16:03better and faster than waiting for me to
16:05respond. And that's the point you want
16:06to get to where an AI has so much
16:07context about you that you can say
16:09something like that. So, how do you
16:10actually start that? Well, the simplest
16:11move, stop opening Claude or ChatGBT in
16:14a blank chat. Instead, spin up a custom
16:15GBT or spin up a Claude project and feed
16:17it real context from whatever you're
16:19working on. So, say you're running a
16:20marketing campaign for a new product
16:21launch. Don't just open up a fresh chat
16:23every time you need help with ideas.
16:24Spin up a project, drop in documents
16:26with your product details, your
16:28marketing calendar, add copy that's
16:29worked well in the past, add copy that's
16:30flopped in the past. Now, the AI is
16:32actually working with that context, not
16:34generic best practices. And the analogy
16:36I keep coming back to here is a summer
16:37intern. When a new intern shows up at
16:39the company, you have to sit them down
16:40and kind of onboard them, right? Like
16:42you have to explain what the business
16:43does, walk them through who's on the
16:45team and who does what. You have to tell
16:46them what current projects matter. And
16:47only after they have all that context
16:49can they actually contribute in a
16:50meaningful way. And AI is the exact
16:52same. And without the context, it's just
16:53a smart intern who's guessing. And
16:55remember, the context you're giving for
16:56the most part is data that's not
16:58publicly accessible. The context about
17:00your subject matter expertise, your
17:01brain, your IP, that's what makes the
17:03outputs unique. If everyone's using that
17:05same model and asking for the same
17:06things, then everyone's outputs will
17:07look the exact same. So your context is
17:09really, really important. So just
17:10remember, garbage in, garbage out. If
17:12you give your AI bad data and no
17:14context, then you're going to get a very
17:15generic output. So that's skill number
17:16three, become a context engineer. Now,
17:18skill number four is one of the most
17:19underrated skills on the entire list.
17:21And in the AI era, it might be the
17:23biggest separator between the people who
17:24win and the people who get left behind,
17:26and it's iteration speed. Now, if skill
17:28number two was about knowing what good
17:29looks like, this skill is about getting
17:31there as fast as possible. So, the two
17:33skills kind of work hand in hand, but
17:35this one stands on its own because in
17:36the era of AI, the people who iterate
17:38fastest are the ones who win. If you can
17:39move fast without sacrificing quality,
17:41you're just going to outperform
17:42everybody because every iteration is
17:43more data. Every iteration is a chance
17:45to learn what's working and what's not
17:46and a chance to make your skills and
17:48your agents and your prompts and your
17:49context, all of it better. The analogy I
17:51always go back to here [music] is like
17:52you're teaching a kid to ride a bike.
17:54You can't just chuck a kid on a bike and
17:56say, "Have fun." And expect them to go
17:57ride a mile. That's not how it works.
17:59You would put them on the bike. You'd
18:00maybe put one hand on their back. You'd
18:01hold the handle and you'd start walking
18:02with them. You would feel if they were
18:04leaning left and correct them. You'd
18:05say, "Hey, you know, shift your weight
18:06over to the right a little bit. You're
18:07helping them calibrate. And after each
18:09run up and down the driveway, you
18:10continue [music] to calibrate. You
18:11continue to iterate and adjust. And the
18:13more time that that kid spends on the
18:14bike with your guidance, the more that
18:16you can start to slowly let go and
18:17eventually you take off the training
18:18wheels and one day you give them a
18:20little push and they just ride and
18:21they're pedaling and they are doing
18:22great." And that's exactly how building
18:24with AI works. You very rarely can just
18:26oneshot something. You use the data and
18:28you feed it back in and you make it
18:30better. And the thing is once you've
18:31taught one kid to ride a bike, teaching
18:32the next kid is easier and teaching the
18:34third kid's easier. And by the time
18:35you're teaching your 15th kid how to
18:36ride a bike, you've pretty much got the
18:37process down [music] to a science. And
18:39now obviously every use case is
18:40different, right? Like some agents are
18:41more complex than others and and they
18:43don't always get built the same. But the
18:44idea of your process in building agents
18:47gets better every time. So hopefully you
18:48guys get the point that I'm trying to
18:49make here. Remember earlier that little
18:51example I said of like let's say people
18:52are typically producing two spreadsheets
18:54[music] a week and then after Excel they
18:56move that baseline up to 10. The faster
18:57you can iterate, the faster you're going
18:59to be able to produce things, which
19:00means your new baseline is going to be
19:02higher than everyone else's baseline as
19:03far as like units of output. So how do
19:05you actually train yourself to be able
19:06to iterate and move faster? This part
19:08may sound silly, but the first thing I
19:10think is to master keyboard shortcuts.
19:12Stop using your mouse for every little
19:13thing and honestly stop typing
19:14everything, right? Like use voice input.
19:16It's way faster than typing. And we
19:17actually have a voiceto text tool called
19:19Glido that I use literally every day. So
19:20if you want to check it out, links in
19:21the description. But the bigger move is
19:23rapid prototyping. Don't sit there
19:24trying to plan the perfect [music]
19:26version. Just build the ugly version
19:27fast. See what breaks, fix it, and
19:29iterate. That's the whole idea of
19:30getting out a P or a proof of concept.
19:32Now, there's another half to this skill
19:33that's just as important, which is
19:35knowing when to stop iterating. Because
19:36when you're building AI tools, it can
19:38feel like there's no such thing as a
19:39finished product. [music] I've been
19:40there. There's always a nice to have.
19:42There's always one more feature you
19:43could add. So what you have to do is
19:44give yourself a north star. You have to
19:46tie one automation to one very specific
19:48business metric and you have to define
19:50what done is. You have to define what
19:52done looks like before you even start
19:53building. So if it's a customer support
19:55automation, tickets resolved per day. If
19:56it's a sales automation, maybe it's
19:58qualified appointments set per week. If
20:00it's an ops automation, maybe it's
20:01refund percentage going down by X%. So
20:03pick the metric, build until you hit it,
20:05and once you hit it, move into
20:06maintenance mode. Obviously, over time,
20:07you can probably find ways to improve it
20:09and maybe improve the metrics even more,
20:11but the heavy lifting is done. So
20:12whether you're building automations for
20:13yourself or for a client, a clear
20:15definition of done is what keeps you
20:16from scope creeping on yourself. And
20:18that's skill number four, iteration
20:19speed. Now skill number five is going to
20:20feel a little bit different and it's
20:22inspired by Iron Man. So this skill is
20:24building your own Jarvis. So you guys
20:25have seen Iron Man, right? Tony Stark
20:27doesn't sit at his computer typing
20:28prompts into Jarvis all day. Jarvis is
20:30already always there. He runs in the
20:31background and he notices things and he
20:33pings Tony when something needs
20:34attention. He'll even kick off tasks
20:36before Tony even asks. Now, this is
20:38different from skill 3. Context
20:39engineering was about teaching your AI
20:41what you know. Skill number five is
20:42about teaching your AI to act on what it
20:44knows without you having to be the
20:46trigger. So here's the way I think about
20:47this. Imagine you build an automation
20:48that only runs when you explicitly fire
20:51it off. That's great. It's going to make
20:52you a lot more productive. But if you're
20:54not around to trigger it, nothing
20:55happens. Now imagine you build a system
20:57that fires on its own. While you're in a
20:59meeting, while you're on a walk, while
21:00you're taking a nap on the beach, that's
21:02real leverage. So the move here is to do
21:04an audit of your day. What things do you
21:06do every week that get triggered by
21:07something predictable, meaning maybe a
21:09specific type of email coming in or
21:11every Monday morning or every Wednesday
21:12evening or every time a new lead lands
21:14in your CRM? Every one of those triggers
21:16is something that you can actually hand
21:18to a system and tell it to do X, Y, and
21:19Z when A or B happens. But here's the
21:22catch. The second you take yourself out
21:23of the loop, the risk obviously goes up
21:25because you're not sitting there
21:26watching it and making sure nothing goes
21:27wrong. There's no catching the mistake
21:29before it reaches a customer or pulls
21:31the wrong data or sends the wrong
21:32message to the wrong list. So the moment
21:34you remove yourself from the process,
21:35the system has to be pretty airtight and
21:37pretty battle tested. Which is exactly
21:39why a lot of people screw this up. The
21:40second they hear Jarvis or an always
21:42personal AI assistant, their brain jumps
21:44straight to building an AI agent for
21:46basically every function. Whether that's
21:47a new email or an end of week report. A
21:49lot of people just jump straight to an
21:51AI agent. So the real skill here is
21:52knowing when something needs an AI agent
21:54versus when it just needs a simple
21:56workflow that doesn't even use AI at
21:57all. So I think about this like a
21:58vending machine versus a slot machine. A
22:00vending machine is deterministic. You
22:02put in a quarter, you hit E4, you get a
22:03Coke. Same input, same output every
22:05single time. A slot machine is
22:06nondeterministic. You pull the lever.
22:08Sometimes you win, sometimes you lose,
22:09sometimes nothing. So AI agents are slot
22:11machines. And essentially, like every
22:13time you talk to an AI, it's almost like
22:14you're gambling. Like not really when
22:16you put the right harness and context in
22:17place, but you never know what's going
22:18to come out the other side. Agents are
22:20really powerful when you need deep
22:21reasoning and you need, you know,
22:23variability, but they cost more. They
22:24fail more often in unexpected ways. So
22:27they introduce more risk. But if you
22:28have a simple, you know, if this, then
22:29that, that's a workflow. And that's just
22:31a vending machine. Predictable, it's
22:32really cheap, and it doesn't break. So
22:34if your task is something like every
22:35morning at 9:00 a.m., pull last week's
22:37revenue from Stripe and post that in
22:38Slack, that does not need an agent. A
22:40simple workflow could do that in 5
22:41minutes and basically never fail. But if
22:43your task was something like read these
22:44incoming customer emails and understand
22:45what they actually want and draft a
22:47tailored response, now you need some AI
22:49in there because the input is messy and
22:51there's reasoning and you have to
22:52generate some sort of content. And
22:53honestly, this is the elite version of
22:55being the AI person that we talked about
22:57at the start. Because in a world where
22:58everyone is shouting AI, AI, AI, the
23:00person who can actually step back and
23:02say, "Hey, we don't actually need AI
23:03here. We can solve this cheaper, faster,
23:05and with way less risk." That person
23:07stands out way more than the one who's
23:09cramming AI into every single task. So
23:11being the person that has that take
23:12signals that you actually understand the
23:14business problem, not just the AI hype.
23:16When you're building your Jarvis, ask
23:17yourself two questions for every task
23:18that you want to automate. First one, do
23:20I actually need to be the one triggering
23:22this? Or can the system fire this off on
23:24its own? And second, does the step
23:26actually need AI or could a simple
23:27Python script or no code workflow do it
23:30at a fraction of the cost with less
23:31risk? And what you want to do is default
23:33to the simplest thing that gets the job
23:35done. Because the people who win in the
23:36AI era aren't the ones who are building
23:37the fanciest agents with hundreds of
23:39tools and hundreds of sub aents. They're
23:41the ones building systems that run
23:42quietly in the background, costing them
23:43almost nothing and doing real work
23:45whether they're there or not. So that's
23:47skill number five. But the final skill,
23:48I can guarantee is something that you've
23:50never heard of. At least not in this
23:51context. I'm talking about unemployment
23:53insurance. And no, I don't literally
23:54mean taking out insurance. Rather, I
23:56mean creating your own insurance. This
23:58might be a bit of a hot take. Not
23:59everyone's going to agree with me on
24:01this, but I'm bringing it up because I'm
24:02really confident this is going to become
24:04way more normal over the next few years.
24:06And the skill is building multiple
24:07income streams using AI so that no
24:09single employer or client can take you
24:11out. The old career model was basically
24:13like one job, one income, a 401k, maybe
24:15a few investments. But basically, all
24:17your eggs were in one basket. And if you
24:18got fired, you were kind of back on the
24:20hunt. You were polishing your resume,
24:21applying to 100 jobs, hoping somebody
24:23bit. And this new model that I'm talking
24:24about that I see emerging is job
24:26stacking. Your day job plus a couple of
24:28AI powered side income streams. I've
24:30already seen a ton of people running
24:31multiple remote jobs, you know,
24:32part-time gigs, side projects, and
24:34stacking that all to equal way more
24:36income than they'd ever make at just one
24:38full-time job. I'm not saying that every
24:39one of you guys should just go quit your
24:40full-time job and do this. I'm saying
24:42that it's already happening and it's
24:43about to become way more common because
24:45AI lets one person do work that used to
24:47take a team of five. Now, the thing I
24:48want to hammer home here, you don't have
24:50to stack five income streams in
24:52completely different domains. That's how
24:53people end up burnt out and broke. The
24:55better version is one passion with
24:57multiple branches. I'm a really, really
24:58strong believer that to be successful at
25:00anything in life, you have to enjoy it.
25:02You have to have at least some kind of
25:03passion for it. If you're chasing AI for
25:05the wrong reasons or you're going after
25:07something because someone said there's a
25:09lot of money in it, then people are
25:10going to be able to see right through
25:10that and it's going to be really hard to
25:12be successful. So, what I want you to
25:13take out of this is to figure out what
25:15motivates you. Figure out what you're
25:16actually passionate about. And that's
25:18where your north star comes from. I've
25:19got a few different income streams
25:20myself. And what's cool about it is that
25:22they all stem off of my same north star.
25:24Same theme, same expertise just packaged
25:26in a few different ways. For example,
25:27you've got your career, that's your
25:29foundation. your expertise about that
25:30career packaged into maybe a course or a
25:32niche newsletter or blog or microsass or
25:35maybe even some consulting on the side.
25:36It's the same domain, but it just takes
25:37different [music] shapes and that's how
25:38you avoid the biggest trap with this
25:40whole idea which is distraction. When
25:41you're starting, just pick one and go
25:43hard until you have momentum under you
25:44and then you can sort of branch out.
25:46Now, a couple quick caveats to mention
25:47here is once again be smart, check your
25:49employment contract, watch out for
25:50non-competes, disclose whatever you're
25:52doing on the side if your company
25:53requires it, like don't do anything
25:54sketchy and don't burn your day job
25:56chasing the side thing and be safe. But
25:57how do you practically do this? Well,
25:59honestly, this really depends on who you
26:01are as a person, but if I had to give a
26:02default move, I would say building in
26:04public. Experiment with AI tools, build
26:06small things, and share what you're
26:06learning. Document the wins and losses.
26:08Build a tiny brand around the work
26:10you're already doing. Because the second
26:11you start posting, you become
26:12discoverable. That's how opportunities
26:14show up, clients show up, job offers
26:16show up, people want to work with the
26:17people actually doing the work. And this
26:19is also something interesting to think
26:20about. The world is shifting in a way
26:22where humans are using AI for almost
26:23everything, right? Which means when
26:25humans go to search the internet for
26:26something, they're probably going to do
26:28that through some sort of AI interface.
26:30Which means if you don't exist basically
26:32at all, somewhere where an AI can find
26:34you and find information about you, then
26:36it's going to be a lot tougher to be
26:37discovered. Now, if building in public
26:38isn't your thing, that's fine. You just
26:40have to find your own version. Maybe
26:41it's a quiet consulting practice, niche
26:43newsletter that doesn't require your
26:44face. Maybe it's a product that you
26:45build and sell without ever showing up
26:47on camera. Medium is completely up to
26:49you, but the point is you start building
26:50something that's actually yours. So,
26:52those are the six skills that I'd
26:53recommend learning and developing to
26:54futureproof yourself in the AI era. I'm
26:57a strong believer in adaptation and
26:58survival of the fittest. So, as long as
27:00you keep up with the changes and
27:01developments in the space that matter
27:03for your northstar, your ability to earn
27:05and live will always be protected. And
27:06now that we have discussed those six
27:07skills, keep those in mind as you work
27:09your way through the rest of the course
27:10and the rest of these mindset shifts.
27:12Everything loops back to those. So, now
27:14let's get back to those mindset shifts.
27:15The first thing is that AI native isn't
27:17what you know. It's not how much
27:19expertise you have or how many models
27:21you can name. It's about what your hand
27:23reaches for. So throughout this course,
27:25what I want you to think about is how
27:27would you do this manually? When you
27:29know that you need to respond to an
27:30email or that you need to analyze a
27:31report, what do you do? You probably
27:33open up the tab on your browser or maybe
27:35you even have it bookmarked and you're
27:37constantly switching between tabs and
27:39context switching. But being AI native
27:42means, okay, rather than doing this
27:44manually myself, let me default first to
27:47doing this with AI, doing this through
27:48cloud code, doing this through my other
27:50tools, using AI to do something like
27:52research and analysis before I give it
27:54my first pass. That's how someone
27:56becomes truly AI native and way more
27:58productive. I hardly ever leave cloud
28:01code. Most the day when I'm working, I'm
28:03working inside of this interface right
28:04here because it just makes me way more
28:06productive. And then the next one that I
28:07wanted to call out, and like I said,
28:08we'll probably revisit all these, is
28:10number four. Don't quit in the dip. The
28:12payoff is the climb. And here's what I
28:14mean by that. Let me just make a quick
28:16chart here. Whenever you decide to learn
28:19something new, there is typically a
28:21short-term cost that you have to bear.
28:22Whether that is because you're a bit
28:24overwhelmed or whether it's because you
28:25have to learn a new skill or set up a
28:27new system, whatever it is, there's
28:29usually a short-term cost you have to
28:30bear. And that discourages a lot of
28:32people. So let's say you start learning,
28:34you know, you start taking this course,
28:35you start getting a little overwhelmed.
28:37Do not quit. Do not click off the video.
28:39At least save it for later and come back
28:40to it because like I said, what happens
28:42is you start learning, right? And a lot
28:45of people expect that the learning is
28:47going to be linear like this. They
28:49expect this is how their progress is
28:50going to look. But typically what
28:52happens is their progress looks more
28:54like this and it becomes exponential.
28:57And so this gap right here, this is
29:00where people end up dropping out. they
29:02get overwhelmed and they quit here
29:03before they start to get all of the
29:05actual exponential benefits of learning
29:08the thing and implementing the new
29:10technology. And look at all this green
29:12that you're actually going to get when
29:13you keep learning and you keep building.
29:15Another way that I like to think about
29:16it is let's say on this range, right?
29:19The old way of doing something is maybe
29:22getting you results that are about here.
29:24And what happens is the dip that I'm
29:26talking about when you start to learn a
29:27new method. You may feel slower in that
29:29week and you may feel like you're doing
29:31it worse in that week because you're so
29:33used to the old manual way. But what
29:35happens is is that short-term dip, maybe
29:37you're dipping in 20% productivity. Is
29:40that dip worth, you know, maybe the 60%
29:44productivity that you ultimately will
29:46have? In most of the cases, the answer
29:48is yes. But this is where people drop
29:50off. And that's what I'm trying to
29:51prepare you guys for. Do not drop off in
29:53the dip. Do not drop off in this gap
29:55right here. And because I'm designing
29:57this course for knowledge workers,
29:58managers, regular people that don't have
30:00coding backgrounds. Think about it like
30:02this one overarching rule. You are just
30:05a manager. You're managing AI agents.
30:08What does that mean to me? Think about
30:09it like if you've ever managed an
30:11employee, which a lot of you guys
30:12probably have, but if you haven't, this
30:14is typically what it would look like.
30:15First of all, you get them onboarded.
30:17You let them get to know you. You get to
30:19know them. You let them get to know your
30:20business a little bit.
30:22You don't want to overwhelm them. You
30:23don't throw them 10 products on day one.
30:25You slowly phase them in until you start
30:27to feel a little bit more trust. But
30:29that doesn't mean you just let them run.
30:31Your job is to very, very clearly tell
30:33them, "Hey, this is what you need to do
30:35today. This is what good looks like.
30:36This is what bad looks like." And then
30:38when they finish their job and they give
30:40you something, you don't just pass it
30:41along or accept it as is. You review it.
30:44You look at it. You use your judgment.
30:46You use your taste. And then you say,
30:47"Hey, here's what you did good and
30:49here's what you did bad." Now take my
30:51feedback and iterate again and update
30:54your instructions. You know, remember
30:55that I told you this so that next time
30:57it's not bad and next time it's even
30:59better. And that doesn't mean it's going
31:01to be 100% on the first pass, but it
31:03means every single time that you get a
31:05deliverable from your AI agent, that's
31:06an opportunity to improve the system.
31:09And the cool thing is all of this
31:10improvement, all of this instructions,
31:12everything I'm explaining right now and
31:13everything for the rest of this course,
31:15it's just natural language. So if you
31:17can think clearly about what you want,
31:18which humans are pretty good at, and you
31:20can describe what you want pretty
31:21clearly, which once again, humans are
31:22pretty good at, then you will be good at
31:24managing AI agents. All right, so let's
31:26move on to number three, which is
31:27installing and signing in to cloud code.
31:31So you can just go ahead and Google
31:32cloud code install, right? You can click
31:34on the quick start docs, and then you
31:36can pretty much figure out right here
31:37how to get a cloud subscription because
31:39you do need a cloud subscription to use
31:40cloud code. And then how do you actually
31:42install it? And there's a few ways to do
31:44this. If you want to use one of these
31:45commands in your terminal to install
31:47cloud code on your device, that works
31:49fine. And by that I mean you would open
31:51up your command prompt or your
31:52PowerShell or, you know, whatever it is
31:54on your operating system and just run
31:56these commands. You could even have
31:58Cloud Chat help you out with this if
32:00you're having trouble for some reason.
32:01But it's super easy. And what I would
32:02recommend is getting the Claude Code
32:04desktop app or just cloud desktop app.
32:05So Google that, click on this right
32:07here, and then download for your
32:09operating system. So for me, I'm on
32:11Windows and I would download this, run
32:13the wizard and then when you open up
32:15Claude, it will look something like
32:16this. And you're just going to go ahead
32:17and get started. Now, this is where you
32:19are going to, like I said, have a paid
32:21plan for Claude Code. You can start on
32:23this pro plan. As you can see, you get
32:24Claude Code and then you can upgrade
32:26that later if you want to max the $100 a
32:29month plan or the $200 a month plan. I
32:31know that sounds expensive, but think
32:32about this. For 200 bucks a month, you
32:35can get basically a full AI employee,
32:36which is the cheapest employee you might
32:38ever get for the amount of work you can
32:40do. For reference, um, you know, a good
32:42project manager or a good software
32:43engineer could cost you way upwards of
32:45$100,000 a year, whereas this is only
32:47going to be 200 bucks a month, which is
32:49very cheap. So, start with Pro and
32:52upgrade later as you need. And once you
32:54have that account, then all you have to
32:56do is actually sign in on the Claude
32:58desktop app. All right. So, now that we
33:00are set up, let's talk about where to
33:01run Cloud Code. So, I just showed you
33:03guys how to use the cloud desktop app,
33:05right? You can come in here and we can
33:07talk to Claude chat right here. I can
33:09say something like hello or sorry, it's
33:11not cloud chat. This is cloud code and
33:12this is the cloud desktop app. So, I'm
33:15going to be using this throughout the
33:16course because I think the interface is
33:17super nice and we can manage all of our
33:19different projects and our chats on the
33:20lefth hand side. So, this is what I will
33:22be using today. However, I will say a
33:25lot of the times I do like to use this
33:26in VS Code. So, VS Code is just a simple
33:28IDE. It's completely free to download
33:30and it lets me use cloud code in the
33:31terminal like this or I can even use the
33:34uh cloud code extension which looks once
33:37again a little bit more userfriendly. So
33:39the point I'm trying to make here is
33:40there's a lot of different ways you can
33:41run it. Other YouTubers run it different
33:43ways, your friends might run it
33:44different ways, but under the hood it's
33:46basically all doing the exact same
33:47thing. So don't stress too much about
33:49where you use it. If later on in 2 weeks
33:51you want to switch to VS Code, you can
33:53do so and nothing will change. All of
33:55your files and projects will still be
33:56there. all of your sessions, all of your
33:58apps, whatever you built, it's still
33:59there. It's just the way that you
34:01actually interact with it. So, don't
34:02stress about it too much. Like I said in
34:04today's video, I am going to be using it
34:06right here with the Clawed Desktop app.
34:09Okay, so earlier I said that this can
34:11work with your local files, right? But
34:13what does that actually mean? So, if I
34:15open up my file explorer right here, you
34:17can see that I've got, you know, my
34:18downloads, my desktop, a bunch of other
34:20folders, pictures, whatever. And Claude
34:23Code is able to navigate through all of
34:25this, edit these things, move them
34:27around, find things for you. So, just as
34:29a super simple example, right here, I
34:30have AIS Live Black, which is just a
34:33logo, right, for our upcoming event, and
34:35this is in my downloads folder from a
34:37few weeks ago. Let's say I knew that
34:39that was there, but I forgot exactly
34:41what it was called or how to find it. I
34:43could real quick just say, "Hey, so I
34:46know that I have the AIS Live logo, the
34:49black version, somewhere in my downloads
34:50folder. could you real quick just find
34:52that for me and then you know just help
34:53me pull that up. And so what happens
34:55when you actually send off a message to
34:56claude code is it will tell you what
34:59it's doing. So right here it says
35:00searching AIS live. That means it's
35:02searching through my um local files and
35:06it's searching for these terms AIS live
35:08logo ais
35:10you know BB black AIS live. It found
35:13this right here. It found the AIS live
35:15PNG. Then it says, "Okay, there's it
35:17with the blue live and the red dot,
35:19which means it used its vision to look
35:21at it to verify that this actually is
35:23the real logo that I'm talking about."
35:25And now it's using something called
35:26PowerShell. So, it's basically just
35:28running commands. And it said, "Pulled
35:29it up. Here's the file. It's in your
35:31downloads." There's also an AIS Live
35:32White sitting right next to it. Do you
35:34want me to copy these into the brand
35:36assets so that it's in our project here?
35:38Or, you know, like what do you want to
35:39do with it? And basically, if I wanted
35:41to just find it again, I could just copy
35:42this file path right here. I could then
35:44go into my file explorer and I could
35:46just come in here, paste that, hit
35:47enter, and it will open up the picture.
35:49So now I was able to locate exactly
35:51where that was. But yeah, if I wanted it
35:53to organize my downloads folder or
35:54organize all of my documents, it could
35:56do so. Now, another quick example, it
35:59can also create things for you like
36:00Excel sheets, Google Sheets, Google
36:02Docs, HTML, um, documents, websites,
36:05apps, anything. So here's a quick
36:08example. I gave it a SLG goal prompt
36:10which I'll talk about later in this
36:11course, but that basically just means
36:12that I'm able to set a condition and
36:15Claude will keep working until that
36:17condition has been met. I mean, that's
36:19so that's just an employee, right? So, I
36:22basically said, I want a quarter 2
36:23assessment of my YouTube channel
36:24performance. This means that you have to
36:26go to YouTube, pull the data, and then
36:28put everything into an Excel sheet. And
36:30I don't only want you to display the
36:32stats. I want you to do deep analysis as
36:34if you are my master content strategist.
36:38an analyst. So then what happens is it
36:40reasons through and you can see here it
36:42even took screenshots to verify that
36:43everything looked good and to verify
36:45that everything was accurate and then it
36:47comes through and it gives me this Excel
36:48sheet and as you can see it put it here
36:50inside of my Herk 2 project in a folder
36:52called projects in a folder called
36:54YouTube Q2 2026 assessment and then it
36:56gave me this Excel sheet which if I pull
36:58this up this thing is pretty legit. We
37:01have a start here page with a bunch of
37:02you know just basically onboarding us to
37:04this doc. I can see the executive
37:06dashboard with my real-time stats,
37:07subscribers, my performance from the
37:09quarter, things that mattered. I can
37:11look at per video scorecard with length,
37:13views, views per day, watch hours, all
37:15these other stats, monthly trends,
37:17content pillars, format and length,
37:19audience, and traffic. So, this is
37:22insane. How long would this have taken
37:24you to manually go pull out, you know,
37:2775 videos, put all the analytics in
37:29here, color code it, design it, do all
37:32of this, and this literally took my AI
37:34agent about 10 minutes right here inside
37:38of Cloud Code. And once again, all I did
37:40was I used my completely natural
37:41language. All of you guys could have
37:42instructed Claude to do this. Okay, so
37:44we're flying through here, just getting
37:46through a lot of the beginning stuff,
37:47and hopefully you guys' mindsets are
37:48getting in the right spot and you're
37:49getting excited. Let's talk a little bit
37:51about prompting. What exactly is
37:53prompting? Prompting is the way that you
37:55talk to your AI in a way that actually
37:58helps it achieve the goal that you want.
38:01Now, there's a few levers to pull here
38:02when it comes to prompting. And it's
38:04really not that complicated. You know,
38:06we used to have this term, which it
38:07still exists, but there's this term
38:09called prompt engineering, which is
38:12basically the art of designing prompts
38:14to get your AI agents to actually do
38:16what you want them to do. Now, prompt
38:19engineering is becoming a little bit
38:20less important over the years. It's
38:22still very important, don't get me
38:22wrong, but the the models are getting so
38:24much better where the prompt is less
38:26important. It still is important, but
38:29over time, I think it will become less
38:30important because the models are getting
38:31smarter. But generally, the things that
38:33I like to tell my agent are the role. I
38:36like to define, hey, here's who you are.
38:39Here's what you're supposed to do. And,
38:41you know, context. So, you are a master
38:46content strategist like you just saw in
38:47that example. you are helping Nate who
38:50runs a business doing X Y andZ and here
38:52is what's important to Nate's business
38:53these X Y andZ metrics and here is the
38:55avatar for Nate's business so giving
38:57context on you know the background so
39:00context and background I should say
39:02because not only is the background
39:04important but specific context so I need
39:06you to help me run you know a Q2 an
39:08analysis on my YouTube channel so that I
39:11can look at all the stuff and I can make
39:13Q3 even better example instead of saying
39:16hey help me write this email to my boss
39:18say, "Hey, help me write this email to
39:20my boss. I need you to, you know, sort
39:22of tread lightly here because I've
39:23gotten in trouble twice in the past
39:24month, and I'm asking for more time off,
39:26and you know, he's a really great and
39:28understanding person, but I just feel
39:30like a little bit guilty for asking for
39:31this time off, so that's why I need help
39:33writing this email." Giving your agent
39:35that context is going to make it
39:37understand your actual desires much
39:39better. So, you're not sitting in this
39:41place where you're like, "Okay, AI
39:42sucks. It didn't give me what I wanted."
39:44it's probably because you didn't give it
39:46clear enough instructions of what you
39:48actually want. Now, the next thing I
39:49like to do is I like to negative prompt.
39:51Basically meaning what not to do. So, if
39:54you think about it like you're
39:55instructing a student or, you know, a
39:57kid who's trying to learn a new process,
40:00you have to tell them what not to do
40:02because they're curious and they're
40:04going to try different things unless you
40:05explicitly tell them not to. Now, if you
40:07were telling someone like, I don't know,
40:08a 45-year-old how to make scrambled
40:11eggs, you probably wouldn't tell them
40:12not to put their palm on the stove top
40:14because, you know, they've been around,
40:16they have experience, they know not to
40:17do that. But an innocent mind probably
40:20doesn't know that. So, you would say,
40:21"Hey, don't ever touch the pan. It's
40:23going to be very hot." So, negative
40:25prompting, I found, has really helped
40:27make sure I keep the agents on the guard
40:29rail that I'm actually, you know, trying
40:30to keep them on. And then another thing
40:32that I really, really like to do is add
40:34verification. Basically, the idea of,
40:37you know, and I'm actually going to type
40:38this out because it's so important in
40:39all caps. Make the AI prove its work. Is
40:45that the right its? That might not be. I
40:47think in this case, that's the correct
40:49it. But either way, you guys know the
40:50point I'm trying to make. Make the AI
40:52prove its work. Think about it like
40:54this. Let me make another one of my
40:57little axis charts here. Okay. So when
41:00you ask AI to do something, what are you
41:03ultimately looking for is you are
41:04looking for it to get you an output that
41:07is 100% perfect. Now realistically that
41:10doesn't happen very often. So what
41:12happens is on the first pass you're
41:14maybe getting somewhere around 60% of
41:16the way there and then what happens is
41:17you give feedback like we talked about
41:18earlier and then it tries again and now
41:20you get a little bit higher and you just
41:22keep doing these manual iterations of
41:23feedback until you eventually get to
41:25this point where you're satisfied with
41:26the work and maybe you take it home that
41:28last 2% or 1%. Now this is because it is
41:33not the one verifying its work. You're
41:35the one verifying its work. So all of
41:37these steps are keeping you in the loop.
41:39But what if you could actually make the
41:41AI check its own work so that on the
41:43first try now it's maybe getting and
41:45that's not completely aligned. Let's try
41:46that again. So that on the first try
41:48it's maybe getting you 80% of the way
41:50there. And then you the human have to
41:52iterate two or three times and then
41:53you're there. And this gap is what
41:55you're trying to close here by allowing
41:57the AI to prove its own work. So what
41:59could that look like? Let's say you are
42:01building um a website and you want to
42:04make sure that the form submission works
42:06and that it doesn't accept, you know,
42:08bad versions of an email. For example,
42:10if people are submitting their emails,
42:12it has to be, you know, an actual at
42:14something domain. Well, you can have the
42:17agent open up the website and test it a
42:20100 times and test that, you know, none
42:22of these edge cases sneak through and
42:24then it proves to you, hey, here's what
42:25I ran. Here's why I'm confident. And
42:27what what you'll find is when you do
42:28stuff like that, when it's trying to
42:30prove it's work and it finds bugs, it
42:32will fix the bug and then keep testing
42:33and then fix the bug again and then keep
42:35testing. So the way that you do
42:36verification is obviously different for
42:38whatever the task at hand is, but every
42:40task at hand has some sort of
42:42verification. Just think about it like
42:44this. If a human gave you this work,
42:46what would you do to approve it? Would
42:48you just read it through a bunch of
42:50times to make sure there's no
42:50grammatical errors? Would you actually
42:52open it up and use it? Would you make
42:54sure that there are no, you know,
42:56elements out of bounds or something like
42:58that? Whatever you would do to verify
42:59it, chances are you can tell cloud code
43:02to do that to verify it. So those are
43:05kind of like the main four things just
43:06to keep it simple that I'm always
43:08thinking about when I am prompting my
43:10cloud code to do something for me. All
43:12right, so let's talk a little bit about
43:13tokens and models. So what is a token?
43:17Well, a token is basically what we are
43:19being built for. So when you pay for
43:23credits on some account, you're paying
43:25for credits. When you are paying for
43:26your AI model usage, you're paying for
43:29tokens. A token is essentially how AI
43:32interprets text. So maybe four
43:35characters is a token or maybe a
43:37punctuation mark is a token. It's not a
43:39deterministic rule of what a token
43:40really is, but roughly 3/4 of a word or
43:43short words. So maybe in this scenario,
43:45the sentence cla code helps you is four
43:48tokens. Now you don't need to know
43:49exactly this like that's not important
43:51if you can identify hey that's one token
43:53that's 12 tokens. What's important is
43:55that you know the pricing. So different
43:58models cost different amounts and they
44:00all have different pros and cons. So if
44:02we look at for example right now in the
44:04current you know July of 2026 clawed
44:07models with their tokens. Haiku is fast
44:10and cheap. So for certain scenarios you
44:12only need Haiku because it's fast and
44:14cheap. It's $1 for a million input
44:17tokens and it's $5 for a million output
44:20tokens. Sonnet 5 is $3 for a million
44:23input tokens and $15 for a million
44:24output tokens. And this is a very
44:26balanced model. And Opus 4.8 is the most
44:28expensive model right now besides Fable,
44:30which is even more expensive, double the
44:31cost of Opus. But Opus 4.8 is $5 for a
44:35million input tokens and $25 for a
44:37million output tokens. Now, yes, you
44:39guys are noticing there's input and
44:40there's output. And the output tokens
44:42are more expensive. So, what is the
44:44difference there? Well, let me open up a
44:46new chat real quick. Or actually, let's
44:48just go to the one where I said hello.
44:49So, every time that I shoot off
44:51something, my prompt goes into the
44:54model. So, this is where we're being
44:56charged for Opus, for example. Actually,
44:58let me switch this model back to Opus.
45:00This is where we're being charged $5 per
45:02a million input tokens for this. Now,
45:04everything that it spits out, all of
45:06these things, all of these commands,
45:08every line of text that it gives us that
45:09it puts out, that's being charged for
45:12Opus at $25 per million output tokens.
45:16So, when it is outputting stuff, that's
45:19more expensive than what you're feeding
45:20into the model. So, for example, if I go
45:22back to the YouTube Analytics one where
45:24it gave us this Excel sheet, let me ask
45:26a quick question. How many tokens did
45:30you have to output to actually generate
45:32that YouTube assessment Excel sheet? And
45:35because you were using Opus, how much
45:38real money would that have costed me?
45:40Okay, so keep in mind that this is an
45:42estimate. But here is what this looked
45:45like. So it had to run different
45:47scripts. It had to check them out, write
45:50up the brief, it had to build the Excel
45:52sheet with different actual characters
45:53inside, and then it had to output
45:55basically all those commands. So they're
45:58calling this an average of let's just
45:59say 25,000 output tokens. And at the
46:02pricing for Opus 4.8 of $25 per million,
46:05this basically costed us about 63 to run
46:09because of the output tokens. Now what's
46:11interesting though is because in this
46:12session you can see that we've used
46:14about 428,000
46:16tokens out of the a million context
46:18token window which I will explain in a
46:20bit if that makes no sense to you. But
46:22the point here is that we probably sent
46:25in and the AI model had to look at
46:27hundreds of thousands of tokens because
46:28it had to pull so much data from YouTube
46:31and analyze so many things and that was
46:33going into the model. So those input
46:35tokens because typically you use a ton
46:37more input tokens. That's why they're
46:39build lower. This will all start to
46:41click a little bit more when you really
46:42start to get your hands on. But I just
46:43wanted to show you a real quick example
46:45of kind of the difference and
46:46understanding that different models have
46:48different, you know, pros and cons. fast
46:50and cheap, balanced, most capable, but
46:51they also come with different prices.
46:53Now, keep in mind because you guys are
46:56on a clawed subscription, when you come
46:58in here and you go to your usage, this
47:00is how you're being build. You're not
47:01being build per token because you're
47:03using this inside of your subscription.
47:05In your subscription, you have different
47:07limits. You've got a current session
47:09limit, which is a 5 hour rolling window.
47:11You've got all models, which is a weekly
47:13limit. And then right now, we have a
47:15fable limit as well. So you can only
47:16use, you know, this much fable before
47:18all of this would then switch usage
47:20credits on top. It's really cool because
47:21I'm on the $200 a month max plan for
47:24cloud code. And if I filled up every
47:26single session, every single weekly
47:27limit, I would actually be getting
47:29around $8,000 of inference out of my
47:31$200 a month plan. So we're getting this
47:34on a huge discount right now. So if you
47:36are using cloud code or you're using the
47:38cloud models or any AI models through
47:40API billing, which is through token
47:42billing rather than through a
47:43subscription, you are paying a ton more
47:46for those tokens. So that is something
47:48else important to keep in mind here. But
47:50don't stress about this too much. You're
47:52on a subscription, you're getting a good
47:53deal. Just monitor your session limit,
47:55monitor your weekly limits. I've got
47:57videos coming later and I've got, you
47:59know, sections in this course later
48:00about context window management and
48:02session limit management. So just keep
48:04that in mind for later. But I just
48:06wanted to break that down a little bit
48:07for you guys so that these terms all
48:09sort of click and make sense. Okay. The
48:11next thing we have here to understand is
48:13called the claw.md.
48:15So if I just real quick open up this
48:18project, right? This is my Herk 2
48:19project which I'll refer to a lot. This
48:21is my AI operating system. This is the
48:23place where I have basically everything
48:24about my business. If I go over here to
48:27my files and I scroll until I find my
48:31claw.md, as you can see, there's a lot
48:32of folders and files inside of this
48:34project. There it is. I scrolled right
48:36past it. This is my claw.md. So, let me
48:38open this up full screen. This is
48:40basically the system prompt for my AI
48:43agent. So, before I read this, let me
48:44just show you a quick visual. So, this
48:46is our little cloud code agent, right?
48:48And let's say we open up a new chat. So,
48:50this thing is completely fresh. It just
48:52woke up. And I go ahead and I shoot off
48:54a message to this thing that says hi.
48:56Right? So I say hi. I don't know why
48:59this keeps going green, but I say hi to
49:01my AI agent. What it does before it
49:03processes this message is it's going to
49:05read the cloudmd
49:07so that it can basically get, you know,
49:09it can orient itself with where we are.
49:12So it will read this cloudmd which is
49:16bunch of lines of text and then it will
49:18read my message and then it knows how to
49:20respond more accurately. So, if I go
49:23back into this real quick and actually
49:26just real quick close out of this file,
49:27what you'll see is that when I said
49:29hello, it said, "Hey, Nate, ready to go
49:31whenever you are." A few things on the
49:33radar in case any next AIS live is this
49:35Friday and Saturday, July 11th and 12th.
49:37CIA opens July 28th. Open threads,
49:40highros, link tagging, keep the best of
49:42fable, what do you want to work on? The
49:44reason it knew this is because it was
49:46able to orient itself with my cloud.MD
49:48MD and read about my projects and you
49:50know what's going on in my business so
49:52that it's able to help me out way more
49:54specifically. So let's take a brief read
49:56through some of the stuff that's in my
49:57cloudmd here. So this is me setting up
50:00the role, right? You are Nate Herk's
50:02executive assistant. Your job is to help
50:03him spend less time on operations,
50:04people management, and admin so he can
50:06focus on learning AI tools and making
50:08YouTube videos. That is his number one
50:10priority. I've then given it a routing
50:12map. So this is where things live. If
50:14you want to find things about, you know,
50:16his business, his team, his OTAAS, his
50:18strategy, go here. Here's the path. If
50:20you want to find things about corporate
50:22structure, entities, tax, IP, then go
50:24here. Voice and style, go here. Course
50:27knowledge, go here. Projects, as you can
50:29see, I'm just telling it where
50:31everything lives. And now it can read
50:33through this and it can help me out way
50:34more specifically. So, this is something
50:36that you'll build over time. I'm going
50:38to build one with you guys. Don't worry,
50:39just a sec. But this is basically just
50:42what it is. It's a system prompt for
50:43your AI agent and it changes all the
50:45time because remember how earlier I
50:47talked about the loop of getting an
50:50output
50:52using judgment to assess it and then
50:54instructing Claude to change so that it
50:56doesn't happen again. This is where I
50:57would say you know like let's say it
50:58gives me this output right let's say it
51:00gives me some research brief that was
51:01just absolutely horrible. I would then
51:03say okay
51:05Mr. Claude I read this research brief
51:08and I don't like this because you didn't
51:10use enough sources. is you only gave me
51:12four. Um, you also put a ton of m dashes
51:15in here which I don't like. So update
51:16your instructions in the claw.mmd so
51:19that you don't do this again. So that
51:20next time I ask you to do research, you
51:22do it better. And then claude will write
51:24changes in its own claw.mmd and then
51:27next time you talk to it, it will be
51:28smarter. So that's how these things just
51:30get smarter and smarter as you use them.
51:31And by the way, when you guys see a file
51:33that ends in MD, all that means is
51:36markdown. Markdown is basically just
51:38like a a computer language and it really
51:41just lets computers understand like
51:43headers and bullet points and really
51:46just structure like that. So this is
51:49markdown when it's rendered nicely for
51:51us humans to read. But this is kind of
51:54what raw markdown looks like. You can
51:56see we use these pound signs to
51:58indicate, you know, status and layers
52:00and we use these dashes for bullets and
52:02everything like that. But MD just means
52:04markdown. The same way later in this
52:06course you might see a py file which
52:08would be a Python file or a txt which
52:11would be a text file. So the dot and
52:13then the you know whatever comes after
52:15that the suffix or whatever is just the
52:18type of file. All right. So here's what
52:20I'm going to do that I want you guys to
52:21do with me. I'm on my desktop. You can
52:23be wherever you want but I'm going to
52:24create a new folder. I'm going to open
52:26up a new folder right here. And I'm just
52:27going to say I'm just going to call this
52:29knowledge work since that's kind of what
52:30this course is about. So, I have a
52:32knowledge work folder on my local
52:34desktop that I just created. And what
52:36we're going to do now is I'm going to
52:37open up Claude and I am going to click
52:39on the plus um right here, new session.
52:42And instead of working in my Herk 2
52:44directory, I'm going to open a new
52:46folder. And this is where I'm going to
52:47go to my desktop, and I'm going to find
52:48that knowledge work folder, and I'm
52:50going to open that up, which basically
52:52means we're now working inside of this
52:53knowledge work folder. Where is it?
52:56There it is. Okay. Select the folder.
52:58Now, it's going to ask if we want to
52:59trust this workspace, which I'm going to
53:01do because it says Claude may read,
53:03write, or execute files in this
53:05directory. So, only proceed if you
53:06actually trust the space, which I do.
53:08So, I click trust. Now, what happens?
53:10Well, let's just start building this
53:12thing out. I'm going to say, "Hey,
53:14Claude Code, my name is Nate." We're
53:16just going to start off conversation
53:17like that. Now, there is one thing to
53:19keep in mind here is because we're in a
53:22new project, Claude still knows a little
53:25bit about me. And the reason is because
53:27there's a difference between your global
53:29cloudmd and your project level. So let
53:32me explain that real quick as well. If I
53:34go down here, I've got some stuff ready.
53:36Okay, so global versus project. Global
53:39is something that claude reads every
53:41single time and that is at the global
53:43level. Whereas project is only
53:45specifically in that project. So the
53:48example you guys just saw, we looked at
53:50this cloud.MD where you know it said,
53:51"Hey, you are Nate's executive
53:53assistant." This gets loaded in whenever
53:55I'm inside of my Herk 2 project. But
53:58whenever I'm in this new knowledgework
54:00project, if I go over here and I go to
54:01my files, this folder is empty. So
54:03there's no project level claw.mmd file
54:05being read yet. But what there is is a
54:08global level claw.mmd. And that's how it
54:10knows my name. Well, I just said it, but
54:12that's how it knows other things about
54:13me because there is a global file. So
54:16real quick, I'm just going to show you
54:17what the global claw.mmd looks like. And
54:19before you get a little bit overwhelmed
54:21or stressed about this, like I said,
54:23this is something that evolves over time
54:25and it really isn't a big deal because
54:27you're able to change it so frequently
54:29by asking Claude to just change it. And
54:31for me, the global cloudMD is just
54:33things that I like and things that I
54:34don't like. So, let me show you exactly
54:35what I mean by that. The way I get there
54:37is I go to my PC, I go to my drive, I go
54:40to users, and I click on my user. And
54:42then there's a cloud folder right here.
54:44And then from here, I can open up my
54:46claw.md, which is just a markdown file.
54:48So, this obviously looks a little bit
54:50more um ugly because it's opened up as
54:52as markdown in my notepad. But look at
54:55this. This is my global cloudmd. These
54:57rules apply to every product on your
54:59machine. They govern any writing meant
55:01for Nate or publishes Nate. So, LinkedIn
55:03posts, YouTube scripts, comments,
55:04emails, captions, docs, anything. So, I
55:07have this AI phrase kill list, which I
55:09said, hey, every time it outputs
55:11something and I say I don't like how
55:12that sounds because it sounds like AI, I
55:14say add this to my global claw.mmd. You
55:16know, I don't like this stuff. So, it
55:18won't say all of these phrases. It will
55:20never write this in my LinkedIn posts.
55:21It will never write this in my YouTube
55:22video scripts or anything like that
55:24because this is a global rule. And look
55:26at this. It's also just learned other
55:27things about me over time. He here's the
55:29things he doesn't like. Here's other
55:30things that I've learned. And this is
55:32just my global claim. It's very simple.
55:34So, if you've got specific things about
55:35your business or specific things about
55:37the way that you like to treat Claude or
55:38you like Claude to treat you, then you
55:40can put those in your global rules
55:42because you're okay with the fact that
55:44every single project ever that you do
55:45with Claude on this local machine, these
55:48rules will be in play there. So, that's
55:49kind of the difference between um as I
55:52said global cloud.MD and project
55:55cloud.MMD. And it's important to keep
55:56this in mind because later when we talk
55:58about things like skills, there's also
56:00the element of is this a global skill or
56:02is this a project level skill? Okay.
56:04Anyways, this is a new project, right?
56:06This is called knowledge work. So, I'm
56:07just going to say I'm building a brand
56:09new project right here and this is for a
56:12course that I'm teaching. So, this is
56:13kind of a demo project for me, but I
56:15want to show my audience how to set up
56:17cloudmds, how to build skills, how to
56:19build automations. And so we're going to
56:20do a lot of this together, but right now
56:22what I want you to do is just initialize
56:24with a cloud.mmd and acloud folder.
56:27There doesn't have to be anything inside
56:29of the cloud. Um, but just initialize
56:31with a cloud.md folder and just throw a
56:34little bit of baseline information in
56:35there for now. And by the way, if you
56:36guys are interested in the voicetoext
56:38tool that you see right up on screen
56:40right here, it's it's very pretty, then
56:42check out the link in the description.
56:44It's called Glido. It's our tool that we
56:46created and it is the fastest and the
56:47most private on the market. So check it
56:49out. Anyways, as you can see, it is
56:51going to start building that for us. So,
56:52in a sec, we'll see over here on the
56:53folder side, we will see a cloudmd and
56:56we will see acloud folder. There we go.
56:58They just popped up. So, the cloud
56:59folder has nothing, but the cloudmd. Let
57:01me open that up real quick. This file
57:03gives cloud code the context it needs to
57:05work in this repository. Keep it short
57:07and current. Here's the project
57:08overview, demo project, here's a
57:10structure, here's the conventions, and
57:12any notes. This cloudmd file will grow
57:14over time. So as we go through this
57:17course and as we build new things, we
57:18will all watch this cloudmd file grow
57:20together. So that is just showing you
57:22how easy it is to get that set up and
57:24that your agent can fix that, edit it,
57:26delete it, stuff like that at any time.
57:28Now you guys are probably wondering what
57:29is this.cloud folder. So let me go back
57:31up here and show you that that is
57:33exactly what we are going to cover next
57:35is the cloud folder. These are basically
57:37the first two things that I always set
57:39up when I'm working on a new project in
57:41cloud code. I have the cloudmd which
57:43once again is the rules in plain English
57:44and then we have the cloud folder which
57:46is basically just think about it almost
57:48like a settings folder. It's the config
57:50folder. Usually the things that go in
57:52here there sometimes are more but the
57:53three things that I want you guys to pay
57:54attention to the settings.json which has
57:57things like permissions preferences the
58:00agents folder which is where you can
58:01build your own custom sub aents and put
58:03them in there and then skills which is
58:05where you package up your skills and
58:07they live in here so that Claude can
58:09call on them automatically. Now, just as
58:11an example, let me show you inside of my
58:13Herk 2 project what this looks like. So,
58:15I'm going to come back over here and I'm
58:16going to open up my doclaude, which is
58:18right here. You can see that I've got
58:20my, you know, agents, the one I talked
58:22about. I've got my settings files, and
58:24I've got my skills. Now, obviously,
58:26there's some other things because this
58:27project is massive. And that's why I
58:29wanted to call out just those three. And
58:31really, the most important one, in my
58:32opinion, is skills. We're going to talk
58:34about skills later more in depth, but
58:36look at this. Every single skill that I
58:37have in here is just natural language.
58:39So, for example, if I open up this grill
58:41me skill, it's just a markdown file. So,
58:43if I open this up, it looks very similar
58:45to the claw.mmd. It's all natural
58:47language and it just tells Claude code
58:49what to do in this specific skill. So,
58:51that's the exact same way that it works
58:53in here with our sub agents. And our
58:55settings is basically just a JSON file
58:57that shows Claude what it's able to do
59:00and what it's not able to do. You can
59:02see we have allowed permissions. We have
59:03some environment variables up here.
59:05We've got things that we've denied. That
59:07is basically how this settings file
59:09works and you don't need to know how to
59:11build this. It will build it
59:12automatically and you can use natural
59:13language once again. So every single
59:15project will have these two things but
59:16these two things also live on a global
59:18level because once again if you wanted
59:21to have you know you have your global
59:22cloudmd but then you also have global
59:24skills potentially. I only have one
59:26right now that's global on this machine
59:28but you can have skills that apply to
59:30every project. You can have sub agents
59:31that apply to every project. So that's
59:33the difference between global and
59:35project. But these are two things that
59:36you always want to get set up when
59:38you're building a new project. Okay,
59:40anyways, let's keep on moving here. Um,
59:42one thing I do want to call out is you
59:44might see my claude here switch over to
59:46usage credits. And once again, that's
59:48because I'm about to go past my 5 hour
59:50limit, in which case, if I want to keep
59:52going, I will then be paying per token.
59:55So, because I'm on the 200 bucks a month
59:56plan, that doesn't happen too often. But
59:57when I am pushing this thing to its
59:59limits, I do hit that. And so if you
1:00:00guys see that pop up on my screen
1:00:02somewhere else in this course in a few
1:00:03minutes, don't be concerned. That's all
1:00:05that that means. So anyways, the next
1:00:07thing I want to talk about API keys and
1:00:10env. So first of all, what is an API? It
1:00:13stands for application programming
1:00:15interface. And all of that really means
1:00:17is it's a method to allow one software
1:00:19to talk to another software. So imagine
1:00:22Gmail sending a letter to claude code.
1:00:25That's via API. So the point I'm trying
1:00:27to make here is remember earlier in this
1:00:30course when I showed you guys this demo
1:00:32where I opened up, you know, this
1:00:34YouTube thing and I said, "Hey, go to
1:00:36YouTube and pull my data." The only
1:00:38reason that it was able to go pull my
1:00:40data from YouTube is because I gave it
1:00:42my API key, which is essentially a
1:00:45password. So if I said, "Hey, go pull
1:00:47Mr. Beast's YouTube data," it could only
1:00:49pull publicly accessible YouTube data.
1:00:51it couldn't pull some of those more, you
1:00:54know, detailed stats that I saw from my
1:00:56channel because I don't have his API
1:00:57key. So, what we had to do was we had to
1:01:00give it access to actually pull that
1:01:02information. And that's when it used
1:01:04this fetch YouTube data py, which py
1:01:06stands for a Python file, and it used
1:01:08that in combination with my API key. As
1:01:10you can see here, YouTube data API. So,
1:01:13that's what I'm going to show you guys
1:01:14now in our knowledge work project. We're
1:01:16going to get set up with an API. So, the
1:01:18example that I'm going to show you guys
1:01:19is Tavi. Tavi lets agents search the web
1:01:22better. You can see here they even have
1:01:24an official Tavi agent skills for Cloud
1:01:26Code, which I'm not going to dive into
1:01:27right now, but the reason I wanted to
1:01:29show you guys Tavi is because you can
1:01:30sign up here and you get a,000 free
1:01:32credits right away, no matter what. So,
1:01:34Cloud Code does have built-in tools for
1:01:36research. So, if I say, "Hey, can you
1:01:39please just research what is context
1:01:41engineering?" And just give me like a
1:01:42one-s sentence example or sorry, a one-s
1:01:44sentence definition. So it will
1:01:47basically because I'm prompting it to
1:01:49research this. What it's going to do is
1:01:50it will probably you see it says finding
1:01:52tools. It's looking through different
1:01:54tools and then it's going to actually
1:01:55search the web using a search the web
1:01:58tool. Query select web search. So web
1:02:01search is one of the tools that it can
1:02:02use. It also has something like a web
1:02:03fetch. But anyways the point I'm trying
1:02:06to make here is cloud code can natively
1:02:08search the web. But if we give it
1:02:09something like tavally it's always good
1:02:11to just get other sources right. So what
1:02:13I'm going to do here is go back into
1:02:14tavi. You can see right here I have an
1:02:16API key which I'm going to go ahead and
1:02:18copy. Now API keys because they're
1:02:20passwords once again do not share them
1:02:23with anybody because let's say I am on a
1:02:26paid plan of Tablety and let's say I
1:02:27paid for 5,000 credits. If I gave you
1:02:29guys my API key, you guys could all take
1:02:31that and spend my money and I don't want
1:02:34to, you know, I don't want you guys to
1:02:35spend my money. So I'm not going to show
1:02:36you my API key in this video. I will
1:02:38because, you know, it's a demo and I'm
1:02:40showing you and I'm going to delete it.
1:02:41But that's how you should treat API
1:02:43keys. You shouldn't be posting them
1:02:44online. and you shouldn't be sharing
1:02:45them across your team unless you know
1:02:47your team is supposed to. So anyways,
1:02:49what I'm going to do is say well not say
1:02:52I'm going to go into my files and you
1:02:53can see that once again we only have
1:02:55these two. So what we need is we need a
1:02:57file called aenv
1:02:59which basically just stands for like
1:03:01environment. So environment variables
1:03:03hey cloud code I want to connect you to
1:03:05tavi and in order to do that I need to
1:03:07give you my API key for tavi. So can you
1:03:09please create in this project aenv file
1:03:12so that I can upload my secret. Now the
1:03:14reason we want to create thisv file is
1:03:16it just it's a safe place to put secrets
1:03:19because they won't get pushed to GitHub.
1:03:21You can see here it said let me also add
1:03:23a dot getit ignore so the secret never
1:03:25gets committed. Now, I'll talk about
1:03:27GitHub more later on in this course, but
1:03:29basically what that is is think about
1:03:31this. When you make a Word doc on your
1:03:34computer, no one else can get that,
1:03:36right? But if you push that up to like a
1:03:38shared drive, then other people can
1:03:39collaborate on it and work on it if you
1:03:41make that public. At least you can still
1:03:42push something to the cloud or push
1:03:43something to GitHub and keep it private.
1:03:45So anyways, the point being if you put
1:03:47something in yourv that will never get
1:03:49pushed to GitHub ever. So basically
1:03:52things that are secret and sensitive put
1:03:53in yourv. So anyways, it went ahead and
1:03:56it created that env as you can see right
1:03:58here. And there's currently nothing in
1:04:00there, but it did give us this
1:04:01placeholder. It says Tableau API key.
1:04:04This is where you would put your Tableau
1:04:05API key. So I'm just going to go ahead
1:04:07and delete the placeholder. Go back into
1:04:09Tavi. Copy this value. Go back into
1:04:12Claude, paste that in there, and then
1:04:14hit save. So now that our API key is in
1:04:16there and it's saved, I can say, cool.
1:04:18So I just gave you my Tavly API key. Go
1:04:20ahead and make a test request to Tavi to
1:04:22make sure that that works.
1:04:25So, I'll go ahead and shoot that off.
1:04:26The dictation tool spelled Tavly wrong.
1:04:28So, let me just go into here and add a
1:04:31correction in the dictionary for the
1:04:32real way that you're going to spell
1:04:34Tavi.
1:04:36Okay, there we go. Anyways, it's going
1:04:39to make a test search request. As you
1:04:42can see here, it made a post request to
1:04:43this endpoint. It searched for what is
1:04:45Claude Code by Enanthropic, pulled the
1:04:47key from the ENV, and got a 200
1:04:50response. So, all of that means it
1:04:52worked. Here's what I want you guys to
1:04:53notice. I didn't tell it how to do any
1:04:55of that. It went off, it researched
1:04:56Tavi, it figured out the endpoint, which
1:04:58is this thing right here, and then it
1:05:00just made the request, which is awesome.
1:05:02Before, you know, a year ago, if you
1:05:03were still building with Naden or
1:05:05something like that, we had to basically
1:05:07by hand research the API endpoints and,
1:05:10you know, plug all this in by ourselves.
1:05:11But now, Agentic AI is getting so
1:05:13powerful that my simple request saying,
1:05:15"Hey, just use Tavi turns into the agent
1:05:18reasoning, researching, testing, and
1:05:21verifying." So super super cool. Now
1:05:23what would happen if I went into my
1:05:24files and I went into thev and then I
1:05:27basically just changed the API key. So
1:05:29this is no longer a valid API key.
1:05:30That's not correct. I'm going to go
1:05:32ahead and clear out the conversation. So
1:05:34I can do a slashclear command right
1:05:36there. And now the conversation's reset.
1:05:38And now if I say, hey, can you try to
1:05:40use Tavi to search for Lenol Messi?
1:05:44When I shoot this off, we're going to
1:05:46see that it's going to try to use Tavi,
1:05:48but it should come back with like a 400
1:05:50error. and it's gonna say, hey, you
1:05:51know, like something's wrong with the
1:05:52API key or something like that. Okay, so
1:05:54now it just defaulted to the search, you
1:05:56know, the the default cloud code web
1:05:57search tool, which I didn't want it to
1:05:58do. So, let's try this again. No, we do
1:06:02have a Tavi API key that I've given you.
1:06:04Try to use that.
1:06:06Don't use the clawed native web search
1:06:09tool. Okay, so let's analyze what just
1:06:11happened here because there's a lot of
1:06:12lessons inside. So, let me close this
1:06:14stuff out.
1:06:15We used Tavi, right? We used it with the
1:06:18correct API key and it worked. And then
1:06:19I cleared the session. When I cleared
1:06:21the session and asked for it to use
1:06:23Tavi, it said there's no Tavly tool
1:06:25connected in this session, so I can't
1:06:27use it specifically. Now, what does that
1:06:29tell us? That tells us that Cloud Code
1:06:31just forgot what we just did. It forgot
1:06:33that we just successfully used Tavi. We
1:06:36also see that again when it said, "Okay,
1:06:38I found the key, but there's no MCP tool
1:06:40wired up, but let me try to call Tavi's
1:06:42REST API."
1:06:44It tried. It didn't work. So, it said,
1:06:46"Okay, let me try a different method."
1:06:48that also didn't work and then what it
1:06:49did is it said okay confirm that the you
1:06:52know the API key is correct so that just
1:06:54shows you that it didn't work because
1:06:55the API key was wrong but the other
1:06:57lesson here is that we just used Tavly
1:07:00successfully so why did it have to do
1:07:02all this research once again you know
1:07:04what we would do now is save that
1:07:06somewhere in our project so what I'm
1:07:08going to do is I'm going to fix the API
1:07:10key real quick so I'm going to go back
1:07:11into tablet copy this go back into
1:07:13claude we're going to open up our files
1:07:15go to thev and we're going to replace
1:07:17this once again with the true API key,
1:07:19which is correct. And then I'm going to
1:07:20go back in here and say, "Okay, thanks.
1:07:23You were right. Our API key was wrong. I
1:07:25just fixed it. So, go ahead and test it
1:07:27again." But then the other thing I want
1:07:29you to do is somewhere in this project,
1:07:32save this as, you know, a skill or save
1:07:36this as something because we will
1:07:37probably be using Tavi frequently. And
1:07:40when I want you to do research, I want
1:07:41you to default first to Tavali and then
1:07:44to the default clawed web search tool.
1:07:47So actually don't create a skill. Just
1:07:49put this in the cloud. Denmd
1:07:51save the endpoint so you understand like
1:07:52we've done this before. I don't want you
1:07:54to research it every time and that's how
1:07:56you can keep getting smarter every time
1:07:57I interact with you. So that was
1:08:00obviously kind of a a very messy casual
1:08:03version of my prompt, [snorts] but
1:08:05that's the way I do it, right? I
1:08:06identified something. I found myself
1:08:08repeating something that I didn't want
1:08:10to and then I feed it back in because
1:08:12here's the thing that may not seem like
1:08:13a big deal allowing it to search and
1:08:15remake the request. But what happens is
1:08:18this every time costs you tokens and
1:08:21tokens cost money, right? Cuz they go
1:08:22against your session limit. So the more
1:08:24efficient you can be with memory and
1:08:26with things like that, the more you're
1:08:28also going to be saving tokens. So it
1:08:31tried it again. It worked now that the
1:08:32API is correct. And then it said, "Let
1:08:35me save this setup to cloud.MD MD so I
1:08:37don't have to figure out again and now
1:08:38that's all saved. If I go back into our
1:08:40files and we open up our cloudmd we can
1:08:42now see here that it says research and
1:08:45web search default tavly first the API
1:08:48key lives in thev here is the off method
1:08:50here is the endpoint and here is how all
1:08:52that works. So if we were to clear the
1:08:54memory right now and then ask it to
1:08:56search the web it would actually use
1:08:57tabi and that is how you iterate on your
1:08:59system. Now, the other thing I wanted to
1:09:01talk about real quick is that we right
1:09:03here saw the term MCP, which stands for
1:09:06model context protocol. If you guys have
1:09:08never heard of that, it's basically it's
1:09:10very similar to an API. APIs have lots
1:09:12of different endpoints that can be hit.
1:09:13And MCPs have a lot of different tools
1:09:15that can be hit, but essentially it's
1:09:16the same theory of how can we connect
1:09:19cloud code to QuickBooks or Google
1:09:21Sheets or Gmail or SharePoint. It's all
1:09:24about connecting to other tools via MCP
1:09:26or API or you might even hear later on
1:09:29in this video something called a CLI.
1:09:31They're all basically just methods of
1:09:32connecting to different tools. They all
1:09:34have little bit of different pros and
1:09:35cons, but I don't want to get into the
1:09:36weeds of that right now. There's no
1:09:37reason getting overwhelmed about that.
1:09:39We're just focused on let's connect to
1:09:41the tools that we use every day to make
1:09:43ourselves faster. So, that is going to
1:09:45do it right now for number 10. Let's
1:09:48move on to permissions and settings. So
1:09:51back in this example, as you guys saw,
1:09:53we have a couple things to think about.
1:09:54The first thing is the actual permission
1:09:56mode that we use inside of cloud code.
1:09:59So if I go into here, you can see that
1:10:01we have different modes. We have auto
1:10:03mode, which is the default. We have
1:10:05manual permissions, we have accept
1:10:06edits, we have plan mode, and then we
1:10:09have bypass permissions. And if you
1:10:10don't see bypass permissions, you would
1:10:12go into here, you'd go to your settings,
1:10:14and then you would go to cloud code. And
1:10:17if you keep scrolling down, there's
1:10:18going to be something right here, allow
1:10:19bypass permissions mode. This allows
1:10:23Claude to just do everything. It will
1:10:25never stop and say, "Hey, is this okay?
1:10:26Can I do this?" It will just do
1:10:28everything. And um
1:10:31hence the name bypass permissions,
1:10:33dangerously skip permissions. So,
1:10:35typically being on auto mode works just
1:10:37fine, but there might be some times
1:10:38where you do want it to be able to
1:10:39bypass, but there are some things that
1:10:41you want to be explicitly very careful
1:10:43of. We've all heard those horror stories
1:10:45of agents deleting databases. We
1:10:47actually had something internally where
1:10:48an agent accidentally sent out an email
1:10:50to like 150,000 people with a discount
1:10:52code that wasn't supposed to go out. The
1:10:54reason that happened though wasn't
1:10:55because of a permissions mode like down
1:10:57here. That was because the agent had
1:11:00access to so many tools. And really what
1:11:03you should be doing is like let's say
1:11:04you connect your agent to your CRM and
1:11:07your agent only needs to be able to read
1:11:09it. It doesn't need to be able to delete
1:11:10records or update records. So there's no
1:11:12point in giving the agent that actual
1:11:13tool to be able to do so. So that's
1:11:16where you might want to think about like
1:11:17API keys with scoped permissions and
1:11:19things like that. And what's nice about
1:11:21thinking about scope permissions is it's
1:11:22usually pretty userfriendly on the third
1:11:25party tool side. So here's 11 Labs for
1:11:28example. If you guys don't know what
1:11:29this is, it basically is just like AI
1:11:31voice really. And you can also do like
1:11:33you know you can create voice agents and
1:11:35phone agents and things like that. So,
1:11:36if I come in here to developers and I go
1:11:38to my API keys, you can see I've got a
1:11:40bunch of API keys here. Let's say I'm
1:11:41going to create a new one. Let's say I'm
1:11:43creating this API key and I just need it
1:11:46for sound effects. That's it. And for
1:11:48some reason, it's really high risk where
1:11:50if this API key does other things, it
1:11:52would be bad for the business. So, I
1:11:54would name this key so I am aware of
1:11:56like what it does. And it's always good
1:11:57to name your keys specifically because
1:11:59if you're giving keys to different
1:12:00people on your team or for different
1:12:01agents, you want to see what agents are
1:12:03using what keys, how often, and how much
1:12:05they're spending and how much people are
1:12:06using it. So anyways, that's what we're
1:12:09doing here. Demo sound effect. Right
1:12:11here, you can see that there's an option
1:12:12to restrict this key, which means I can
1:12:14restrict it on credits. So it can only
1:12:16spend a certain amount. So let's say I
1:12:18wanted to only let it spend 10 credits
1:12:21um on a certain, you know, time period,
1:12:23then I could do that. We can also
1:12:25restrict the endpoints. Endpoints is
1:12:28just a fancy word for like what it can
1:12:30do, capabilities, tools. So let's say
1:12:32for this key, I only wanted it to be
1:12:35able to do sound effects. So I would
1:12:36say, okay, you can access that. But for
1:12:38everything else, no access. You know,
1:12:41maybe I want it to be able to read all
1:12:43of this so it can read dubbing, read
1:12:44agents, read projects, read all this.
1:12:47But besides that, the only thing it can
1:12:49actually physically touch and manipulate
1:12:51and do is sound effects. And that's how
1:12:53you give your agents a key that you can
1:12:56actually sleep at night and feel 100%
1:12:57confident that nothing wrong will
1:12:59happen. Because if those horror stories
1:13:01where people delete a database with an
1:13:03agent, if they would have just
1:13:04restricted the ability to delete, then
1:13:06that never would have happened. So
1:13:08that's what I want you guys to be
1:13:09thinking about like this. Would you give
1:13:11your a new hire a credit card and say,
1:13:13"Hey, you can like don't spend anything,
1:13:15but with this card you could." No. That
1:13:18just makes no sense. So treat this once
1:13:19again, what did I say earlier in this
1:13:21video? You are a manager. Just think
1:13:23about this as this is a human. How would
1:13:26I interact with them? What access would
1:13:27I give them? What permissions would I
1:13:29give them? And a lot of the overwhelm
1:13:31about how you think about these agents
1:13:32might just disappear when you shift your
1:13:34mind to think like that. Okay. Now, the
1:13:36other thing we have are some other
1:13:37settings like local settings. So, I'll
1:13:39show you guys a quick example of that.
1:13:41If I go to my back to my Herku project,
1:13:43which is like I said is my main one that
1:13:45I kind of operate in and I go to my
1:13:48umcloud and we talked about our settings
1:13:50living inside of this.cloud folder. So
1:13:52I'm going to open up my settings. What
1:13:53we see here is some environment
1:13:55variables. So cloud code agent teams
1:13:58one. I'll talk about this later in the
1:13:59video during the agent team section, but
1:14:00that's an environment variable. I've
1:14:02also got some other things here that I'm
1:14:04going to blur out. But then I have
1:14:05permissions. And this is me saying
1:14:07here's what you're allowed to do cloud
1:14:09code. You're allowed to do bashes. So
1:14:11just like kind of running these
1:14:12commands. Web search is allowed. Web
1:14:14fetch, edit, write, mcps, mcps, glob,
1:14:17skills, all of this is allowed. But
1:14:19here's what's important. Here are the
1:14:21things you can never do. I I have this
1:14:23project set up so it can never remove
1:14:24anything or delete anything or change
1:14:26these directories or anything here that
1:14:28I would consider risky. And so what I
1:14:30would do is I would put my cloud code on
1:14:32bypass permissions mode. But I felt
1:14:34comfortable about it because it could
1:14:36never do anything that was actually
1:14:37risky. So, that's how I played with the
1:14:39settings in order to actually help me
1:14:40out a little bit with, you know, access.
1:14:42Now, the cool thing about that is I
1:14:44don't understand like a lot of those
1:14:46words, but I just said, "Hey, I'm trying
1:14:47to change my permissions and my settings
1:14:49inside of this project. Here's X, Y, and
1:14:52Z things I want to never happen. Can you
1:14:53help me update the settings file so that
1:14:55you physically cannot do those things?"
1:14:57And then it just worked for me.
1:14:59Obviously, I wanted to test it out a
1:15:00little bit cuz you don't always just
1:15:01take the output, like I said, and just
1:15:03fully blindly trust it, but that's what
1:15:05I did. Now, I will say though, pretty
1:15:07much if you come in here and you go to
1:15:09auto mode, it's going to be really
1:15:10solid. Auto mode was new. So, when I
1:15:13designed that settings file, auto mode
1:15:15didn't yet exist, but auto mode's pretty
1:15:16solid for you guys. When you've opened
1:15:18up Cloud Code and you've been following
1:15:19along with this video so far, you've
1:15:21probably been using auto mode, and
1:15:22that's going to be just fine. So, the
1:15:24settings thing is something to just be
1:15:25aware of, and later you might want to
1:15:27tweak that as you get a little more
1:15:28advanced, but right now, you're probably
1:15:29fine just sticking on auto mode. And
1:15:31then another setting that I want to
1:15:33bring to your guys' attention is down
1:15:35here. So obviously we have the model,
1:15:37right? And you can enable fast mode,
1:15:38which I basically never touch. It costs
1:15:40more and it's faster and I typically
1:15:42don't care about speed too much. We have
1:15:44the different models to choose between,
1:15:45right? So sonnet, haiku, opus, fable.
1:15:48And then what we can also do is the
1:15:50effort, which is pretty interesting. So
1:15:52effort lets you choose between faster
1:15:54and smarter. So, if we move the effort
1:15:56down to medium or low or extra high or
1:15:59max or ultra code, we can play with the
1:16:01effort levels. Now, I will be honest
1:16:03with you guys, I pretty much keep all my
1:16:06models just on high. I don't really like
1:16:08to play with it too much. Typically,
1:16:09what I like to do is switch between
1:16:11models rather than switch between
1:16:12effort. I think that for the most of the
1:16:14knowledge work, you really don't need to
1:16:16tweak it too much. But I will say the
1:16:19unit economics of understanding what is
1:16:22the right model for the specific task in
1:16:24front of you because you're also
1:16:25optimizing for cost is a very important
1:16:27thing to be thinking about as you get
1:16:29more advanced. Right now we're learning
1:16:30the fundamentals. It's not a big deal
1:16:32but as you get more advanced it is
1:16:33interesting to think about this kind of
1:16:34stuff. Look at this for example. We have
1:16:36this chart with GBT 5.5 which is
1:16:38currently OpenAI's best model but you
1:16:40know things move quick. We have Opus 4.8
1:16:42and we have Fable 5. We're showing all
1:16:44of these models on different effort
1:16:46levels. And on the y- axis, we see the
1:16:49score. On the x-axis, we see the average
1:16:51cost per task. Now, what I want you to
1:16:53pay attention to here is look at GBT
1:16:555.5. As effort increases, quality
1:16:59doesn't increase. Score doesn't
1:17:00increase, but cost certainly does. So,
1:17:02that's one where it's like, okay, why
1:17:03would you ever increase the effort? But
1:17:05here on Fable, as you increase the
1:17:07effort, the benchmarks show that your
1:17:10score meaning improves. Now, me
1:17:12personally, when I use Fable on high or
1:17:14max, what I've actually found is that
1:17:16it's slower and it overthinks and it
1:17:18overreasons for no reason. So, I like
1:17:20Fable on high. Opus 4.8 on extra high
1:17:23does significantly feel better, but it's
1:17:25slower and more expensive. So, it's it's
1:17:27kind of a balancing act to play, but
1:17:29looking at these benchmarks is
1:17:30interesting and it's good to know that
1:17:31you have that lever to tweak. But, like
1:17:34I said, for the majority of my knowledge
1:17:35work that I'm doing, I'm just leaving
1:17:37the models on high and that has been
1:17:39working pretty well for me. Okay, now
1:17:40let's talk about something really fun,
1:17:42which is privacy and your data. So, hey
1:17:46Claude Code, can you just do some quick
1:17:47research for me? What does Enthropic say
1:17:49about our privacy? You know, when I'm
1:17:51talking to Claude and the data goes to
1:17:53their servers, like what are they doing
1:17:54with it? Are they training on me? Should
1:17:56I be what should I be careful of? You
1:17:58know, how do I stay safe, especially
1:18:00inside of my own organization and at
1:18:01work? So, the reason I wanted to bring
1:18:04this up is because it is a big question,
1:18:05right? because these closed source
1:18:09models so basically meaning you know
1:18:11OpenAI Enthropic Google those are closed
1:18:14source models because we as consumers
1:18:16don't actually get to own the weights of
1:18:19those models you know we can't install
1:18:21them locally and we can't tune them or
1:18:23anything like that. Now open source
1:18:24models are the ones that you can
1:18:26download locally and tune them but
1:18:28they're not nearly as good as the closed
1:18:30source models. Now the benefit with open
1:18:32source models is that you can own them
1:18:34locally. So your data, your conversation
1:18:38never leaves your home because it stays
1:18:40on your computer. But every single time
1:18:41that I send a message to Claude or
1:18:43OpenAI or Google, it goes to their
1:18:46servers wherever they're running all
1:18:47their compute and then it gets processed
1:18:49and then an answer comes back. And
1:18:51that's why they have these massive, you
1:18:52know, data centers. But the point I'm
1:18:54trying to make here is for the company
1:18:55that you work for or your own business,
1:18:57you should not be sending over private
1:18:59sensitive data. For my case, it's really
1:19:02not that much, right? because I'm making
1:19:03content and I am doing stuff like that.
1:19:05I'm not directly handling people's
1:19:07credit card numbers or their personal
1:19:08identifiable information. But I wanted
1:19:11to bring this up because for some of the
1:19:13clients that I've worked with, they had
1:19:14very strict requirements. You know, like
1:19:16we couldn't do anthropic stuff for them.
1:19:19We had to do on-remise deployments and
1:19:20we had to think about security in a much
1:19:22different way as far as like encryption
1:19:23and stuff like that. But that's why I
1:19:26wanted to bring this up. I'm not going
1:19:27to dive into deep deep deep deep about
1:19:29like encryption and on-rem right now in
1:19:30this course at least, but I wanted to
1:19:32bring this up so that nobody here gets
1:19:34in trouble because they watch this
1:19:35course and then they're all of a sudden
1:19:37putting a bunch of company documents and
1:19:38legal contracts into AI when their
1:19:41company would get them in big-time
1:19:43trouble for doing something like that.
1:19:44So, be safe. Think about what's allowed
1:19:47inside of your industry, inside of your
1:19:49organization before you start chucking
1:19:50in documents into something like Claude.
1:19:53Aha. And look at this. I'll research
1:19:55this with Tavly first per hour clawmd
1:19:57setup. But anyways, I'll just go over
1:19:59real quick what this says. It depends
1:20:00entirely on which cloud you're using on
1:20:02consumer plans, which is, you know, the
1:20:04free pro max, not a team plan. As of the
1:20:07policy change effective October 8,
1:20:09training is opt-in, but it's a forced
1:20:11choice. So, you basically had to toggle
1:20:12this on or off. You can check this right
1:20:14now by going to your settings, going to
1:20:16privacy, going to help improve Claude,
1:20:18and then you can toggle it on or off.
1:20:19Now, just because they claim that
1:20:21they're not going to train models on
1:20:22your data, they still have that data.
1:20:25So, you know, that's why even if you
1:20:27have that toggled off, check with your
1:20:29own organization. Cool. So, now that
1:20:31we've covered probably, of course, the
1:20:33most fun and exciting thing of this
1:20:34course, let's just talk more about
1:20:36connecting your own tools because that's
1:20:37how this stuff gets super super super
1:20:40powerful. Okay, so in my Herk 2 project,
1:20:43which I know I've mentioned a lot, this
1:20:45is my AI operating system. So, it's
1:20:47connected to every tool that I actually
1:20:48use and it has information about my
1:20:50entire business. It knows my business
1:20:52better than I know it honestly because
1:20:53it has a better memory. But the idea is
1:20:55I got here because I've connected all my
1:20:58tools and because I use this for
1:21:00everything and every new memory, every
1:21:02new skill, every new project gets
1:21:04indexed into my wiki here, which I will
1:21:06have sections on exactly how you do this
1:21:08later in the course, but just bear with
1:21:09me here. The point I'm trying to make is
1:21:11that's my AI operating system with my
1:21:13second brain inside. And if you guys
1:21:15want to get the full course, which I
1:21:16would recommend you do after this
1:21:18course, that will be my free school
1:21:19community right here. The link is in the
1:21:21description. Like I said, completely
1:21:23free. So the point I'm trying to make is
1:21:26this
1:21:28idea of connecting your tools is a core
1:21:30piece of building your AI operating
1:21:31system. Context, connections,
1:21:33capabilities, and cadence. And so what
1:21:35I'm talking about right now is kind of
1:21:36like the context and connections thing.
1:21:39So this is where I like to start. I
1:21:41think of my tier one connections which
1:21:43are revenue, customers, calendar, coms,
1:21:46tasks, meetings, and knowledge. So,
1:21:48right now, you know, you've created your
1:21:50folder in Claude, which is going to
1:21:51evolve into your AI operating system.
1:21:53But you're thinking to yourself, maybe
1:21:54okay, Gmail, um, you know, Slack, what
1:21:58else do I connect? I don't know what
1:21:59else to connect. The answer is go look
1:22:02at the tabs you have open right now. Go
1:22:03look at your bookmarks. Go look at your
1:22:05history. What tools are you visiting
1:22:07frequently? what tools do you need to go
1:22:09to pull data or to talk to people?
1:22:12That's what you want to connect. So,
1:22:13it'll be things like where is your
1:22:15revenue living? Can you connect your
1:22:16bank accounts? Can you connect um any
1:22:18reporting tools? Can you connect Stripe?
1:22:20Can you connect your, you know, school?
1:22:22Where do your customers live? You
1:22:24probably want to also have some customer
1:22:25data in there. You probably want to have
1:22:27your calendar in there. You want to have
1:22:28communications, so Google Workspace, but
1:22:30also for me it's ClickUp and Slack. I
1:22:32have tasks in there. So, ClickUp and
1:22:34Notion. I've got my meeting recordings.
1:22:36my transcripts get pulled in
1:22:37automatically from Fireflies and then
1:22:39just other knowledge. So, I've got a ton
1:22:40of stuff locally. I've got a ton of
1:22:42stuff in my Google Drive and of course
1:22:44on my YouTube channel. So, that's
1:22:46obviously not all of it. I think every
1:22:48single week you're going to discover, oh
1:22:49shoot, I need to connect this. And then
1:22:51when you do, it's as simple as googling
1:22:53that tool and then you Google API
1:22:55documentation. So, let's say you use
1:22:57Fireflies and you want to be able to
1:22:58connect Fireflies to cloud code. You
1:23:00would just go Fireflies API
1:23:03documentation. Okay, cool. Fireflies has
1:23:05an API. I will go ahead and say, "Hey,
1:23:07Cloud Code, you know, take this link,
1:23:09figure out how to use this." And then
1:23:10I'm going to go get my API keys from
1:23:11Fireflies. And once I get that, I'm
1:23:13going to chuck that in the ENV. And
1:23:15boom. Congratulations. You just
1:23:16connected another tool to your AI
1:23:19operating system to your cloud code.
1:23:20That's all it takes. The hardest part,
1:23:23honestly, is just like identifying what
1:23:25tools you need. And the best way to do
1:23:27this is to actually test it. Like prove
1:23:30to yourself that you [snorts] can do
1:23:32this. And what I mean by that is you
1:23:34have to make the habit switch. So what
1:23:36it looks like right now is you've
1:23:37probably got all these different tools
1:23:39and I'm not going to um like label all
1:23:42these, but let's just say you've got all
1:23:44these different tools and you open up
1:23:46your Chrome or whatever browser you use
1:23:48and all day you're just switching
1:23:49between a bunch of different tools and
1:23:51you're copying and pasting things. What
1:23:53you need to do is just transition to the
1:23:55fact that I can just use Cloud Code and
1:23:58just talk to Claude Code and that's it.
1:24:00because Claude can be the one to go use
1:24:03all these tools. So, all I have to do as
1:24:05a human, you know, let me just draw a
1:24:07nice little picture of Oops, that's not
1:24:09what I wanted. As myself sitting here
1:24:12on, you know, my computer just talking
1:24:14all day. Then all I have to do, that's
1:24:17supposed to be a computer by the way.
1:24:18All I have to do as the human is just
1:24:20talk to Claude. And now this thing is
1:24:21like my assistant. It does everything
1:24:23for me. And I try to challenge myself.
1:24:26One of the metrics I challenge myself to
1:24:27is what percentage of my day, what
1:24:30percentage of my work can I do from this
1:24:33window right here? And I know that might
1:24:35sound weird. It might sound like, oh,
1:24:36I'm going to be less productive. But no,
1:24:38trust me, not only will you be more
1:24:40productive, but you're going to start to
1:24:41build more skills. You're going to make
1:24:42your cloudmd better. You're going to
1:24:44improve your entire system. And that's
1:24:45how you end up with something like this
1:24:46with hundreds of skills, millions of
1:24:49context files and projects, and
1:24:51something that feels like a co-founder.
1:24:52It's really a great feeling to be to
1:24:54have. So, you're going to go through
1:24:56these tiers. You're going to list out
1:24:58the tools that you want to connect and
1:24:59then just go through find if they have
1:25:01an API, find if they have an MCP, maybe
1:25:03they have a CLI, and start to get
1:25:05connected. What I want to show you guys
1:25:06right now, because I'm hoping a lot of
1:25:07you guys are on the Google Workspace
1:25:09ecosystem, is how you can connect to the
1:25:11Google Workspace CLI because that allows
1:25:14you to talk to everything in the Google
1:25:15Workspace environment. Mail, calendar,
1:25:18sheets, docs, slides, all of it. So,
1:25:22that's what I'm about to show you guys
1:25:23how to set up. Just a quick warning
1:25:25before this next video starts playing.
1:25:26Some of the clips that I'm inserting
1:25:28into this course were recorded a few
1:25:30months back, meaning they might be shown
1:25:31in VS Code extension or the terminal
1:25:34instead of the Cloud Desktop app that
1:25:36we've been using so far. I just wanted
1:25:38to give you guys a warning.
1:25:39Functionally, exact same. So, don't
1:25:41worry about it too much. It just might
1:25:42look a little bit differently, but all
1:25:44you have to do is listen to what I'm
1:25:45saying and follow along with what I'm
1:25:46actually doing and you will be just
1:25:47fine. All of this stuff is still
1:25:49relevant. Otherwise, I wouldn't be
1:25:50putting it in this course. So, hopefully
1:25:52that makes sense. See you guys in the
1:25:54video. Google just dropped what some are
1:25:56already calling the most powerful
1:25:57workspace CLI on the internet. So, if
1:26:00you've got a ton of stuff that lives in
1:26:01the Google environment just like I do,
1:26:03then you're going to love this because
1:26:04now any of my cloud code projects can
1:26:06access everything. And all I had to do
1:26:08was install one simple thing. So, here
1:26:11you can see I said, what can you do with
1:26:12GWS, which is Google Workspace CLI? So,
1:26:16it can search, list, upload, download,
1:26:18move, copy, share, anything in my Google
1:26:20Drive. It can do anything in my Gmail.
1:26:23It can do anything in my calendar. It
1:26:24can do anything with Google Docs. Same
1:26:26thing with Sheets. Same thing with
1:26:28Slides. And it also has multi-step
1:26:30workflow recipes. So these are basically
1:26:32like skills. These are chain command
1:26:34patterns for common tasks like creating
1:26:36docs from templates, reading sheet data,
1:26:38and creating a report doc, finding free
1:26:40time, and scheduling a meeting. And
1:26:41there are over a hundred of these that
1:26:42it actually has. So out of the box, when
1:26:44you give Claude Code the GWS CLI, you
1:26:47can do anything across any of the tools.
1:26:48And you also have access to over a 100
1:26:50skills. So, I don't know how many times
1:26:52you guys have tried to use something
1:26:53like Claude or Naden to build you a
1:26:55Google doc and you do this over API and
1:26:57it ends up just looking like something
1:26:59like this. It literally just looks like
1:27:00raw markdown and it's obviously
1:27:02horrible. And sometimes to go along with
1:27:04a YouTube video, I make resource guides
1:27:05that look like this, but obviously they
1:27:07have to be formatted. I've got like a
1:27:08header up here and I've got links and
1:27:10different things in this format. But now
1:27:13I can literally just take the link to a
1:27:14YouTube video. I can drop that into
1:27:16Cloud Code and say, "Create me a YouTube
1:27:18resource guide." It's going to go ahead
1:27:19and download that transcript from the
1:27:21video. And now what it's doing is it's
1:27:23creating the Google doc, not via API
1:27:25call, not via MCP, but via a bash
1:27:27command, meaning it's literally running
1:27:29a terminal command in order to talk to
1:27:31Google and make this. So, it just
1:27:33actually created the doc. Here's the ID.
1:27:35And now it's going to populate it with
1:27:36what I need. And now it finished this
1:27:38up. It gave me the link. I'll click into
1:27:39this. And we can see, boom, we have an
1:27:41actual resource guide. It's got the
1:27:43image inserted up here as a header. It's
1:27:44got a link that goes right back to my
1:27:46YouTube channel. It breaks down the
1:27:47market traditional automation. It goes
1:27:49through all this stuff and then even has
1:27:50my CTA at the bottom as you can see
1:27:52after all these horizontal lines to join
1:27:54the plus group. So that was obviously
1:27:56just one quick example, but there's so
1:27:58many different benefits here using this
1:27:59workspace CLI. The first one is that you
1:28:01have one interface. So basically, like I
1:28:03said, it was one GWS CLI that Cloud Code
1:28:05now has access to and it can access my
1:28:07Gmail, my drive, docs, sheets, calendar,
1:28:09admin, and more. It's also JSON first
1:28:11with structured responses. So our AI
1:28:13agent is really, really good at working
1:28:14with it. We have auto discovery, meaning
1:28:16the CLI is pretty much always going to
1:28:17stay up to date automatically. Pretty
1:28:19much zero maintenance because we
1:28:20authenticate and then we're going to be
1:28:22good to go. It has built-in skills for
1:28:24triage, for prep, for generations. Like
1:28:26I said, there's 100 others. And it's not
1:28:28much overhead because it's basically
1:28:29just one tool. It's not the same as like
1:28:31having all these different API endpoints
1:28:33or all of these different MCP configs
1:28:35and tools that would take up more
1:28:36context. But I know you're probably
1:28:38wondering, what is a CLI? It stands for
1:28:40command line interface. And what we're
1:28:42typically used to is a GUI or a
1:28:44graphical user interface where we can
1:28:45see buttons, we can see form fields, and
1:28:47we can click on things and that's how we
1:28:49navigate, but computers are more
1:28:51navigating by text and by commands and
1:28:53by typing. So that's really all that a
1:28:55CLI is. And this is an open- source
1:28:57Google Workspace product, and obviously
1:28:59it's completely free. So I'll leave a
1:29:01link to this GitHub repository down in
1:29:02the description if you want to check it
1:29:03out, read more about it. But I'm also
1:29:05going to walk through some of the key
1:29:05details right here. The first thing that
1:29:07I wanted to show you is if you go down
1:29:08here to the skills, this is where we can
1:29:10actually see all of the different kind
1:29:12of recipes they call them for pre-made
1:29:14multi-step workflows that it has. As you
1:29:16can see, creating events from sheets,
1:29:18creating presentations, creating meat
1:29:20space, label and archiving emails.
1:29:22There's so many different patterns that
1:29:23you might use from this pre-built
1:29:25library. Now, if we keep scrolling down,
1:29:27what you'll also notice is that right
1:29:28here it says this is not an officially
1:29:30supported Google product. Now, that
1:29:32doesn't mean that it's unsafe. This is
1:29:33an actual Google product, but the reason
1:29:35why it's not officially supported is
1:29:37because right now it's more of like an
1:29:38open- source beta. It's kind of a
1:29:40developer playground rather than like an
1:29:42enterprisebacked software. And you can
1:29:44see right here that it also says this
1:29:46project is under active development.
1:29:47Expect breaking changes as we march
1:29:49towards v 1.0. So this thing's already
1:29:51really good out of the box and it's only
1:29:52going to get better. And you can see,
1:29:54like I said, when Google Workspace adds
1:29:55an API endpoints or method, GWS picks it
1:29:58up automatically. So you might as well
1:29:59chuck it into cloud code right now and
1:30:01start getting used to it. Okay, so I
1:30:02just uninstalled this so I can walk you
1:30:04guys through step by step how this
1:30:05actually works. It's super easy. What I
1:30:07do is I basically copy the link to this
1:30:09GitHub repository as you can see. And
1:30:11I'm going to basically just give it to
1:30:13Cloud Code and say, "Hey, I want to
1:30:15install this GWS CLI, read through the
1:30:19documentation, and help me install
1:30:20everything that I need to install, and
1:30:22then we're going to get set up." So,
1:30:23this is basically going to do all the
1:30:24research for me, and then all I have to
1:30:26do is follow its instructions. So, it
1:30:28read the docs. It's looking at what we
1:30:30already have installed. It basically saw
1:30:32that I already had some of the
1:30:33prerequisites. So if you don't have
1:30:34those, you'll have to install those. And
1:30:36then it told me that we needed to
1:30:37install the CLI. So it did that. And now
1:30:39we have two options. So the first one is
1:30:41to install G-Cloud CLI so that we have
1:30:43automatic setup and off. Or we could do
1:30:45it manually by creating our own project
1:30:47and whatnot. So let's just go ahead and
1:30:48try option A. Okay. I thought this was
1:30:50going to be just like a simple command
1:30:52that it ran and then we were good. But
1:30:54it's actually like some other thing to
1:30:55install. So let's actually go back and
1:30:56try manual and I'll just show you guys I
1:30:58guess the harder way. Okay. So I'm going
1:31:00to go to this link. go to our Google
1:31:01Cloud Console and make sure you're
1:31:03signed in with the right account up in
1:31:04the top right. And I'm just going to go
1:31:05ahead and create a new project just to
1:31:07show you guys what this would look like.
1:31:08So, new project. I'm going to call this
1:31:10one Claude Code GWS. And we're just
1:31:15going to go ahead and create this
1:31:16project. So, this is spinning up right
1:31:18now as you can see. And now that it has
1:31:20been created, I'm going to select it so
1:31:21we're inside of it. And then I'm going
1:31:22to go up here and type in APIs and
1:31:25services. Click on that. And we have to
1:31:26set up our OOTH consent screen. So, I'll
1:31:28click on this. and it's going to say get
1:31:30started. Click on that. We have to give
1:31:33our app a name. And then we have to
1:31:35choose an audience. So I'm just going to
1:31:36do internal because I only need this
1:31:38right now for my own organization. If
1:31:40you want to do external, it'll basically
1:31:42have you do testing or published. And if
1:31:44you do testing, just make sure that you
1:31:46add your email as a test user. And then
1:31:49all you have to do after you put in your
1:31:50contact information is hit I agree. And
1:31:53then you go ahead and create that. Now
1:31:54once that has been done, you're going to
1:31:56go to create a client ID. So, I'm going
1:31:58to go back into APIs and services. I'm
1:32:00going to go to credentials and then I'm
1:32:02going to go ahead and do a create
1:32:03credential oath client ID. Now, in here,
1:32:05we're going to choose a desktop app. I'm
1:32:07going to just call this GWS and go ahead
1:32:10and hit create. And now we have our
1:32:12client ID and our client secret. And so,
1:32:14what you're going to do is download this
1:32:15as a JSON file. Now, you can see here
1:32:17that it says to download that file and
1:32:19save it to your global.config/GWS.
1:32:23So, basically, if you can't find this,
1:32:25just say, "Hey, can you give that to me
1:32:26in a full path?" And then you can paste
1:32:28that into your finder or your file
1:32:30explorer and it will take you there. It
1:32:31will probably look something like this.
1:32:33And then you just drag in that
1:32:34credential thing. I called mine client
1:32:36secret. And cloud code will be able to
1:32:38look at this globally now. And so what
1:32:40you'll notice is that we didn't in this
1:32:41project yet enable these APIs. So let me
1:32:44just show you what happens without that.
1:32:45So it says the last step is to run GWS
1:32:48off login. So I just said, hey, I
1:32:50finished option B. The credentials are
1:32:51called client secret. And then I told it
1:32:54to run the O login. So that should
1:32:56basically open up a tab for you, but if
1:32:58it doesn't, then you can ask for it to
1:33:00give you that URL so that you can
1:33:01actually authenticate in. So you would
1:33:03basically choose your account that you
1:33:04want to use. And then you just have to
1:33:06basically confirm that it can access all
1:33:07of these different things. As you can
1:33:08see, and then when you hit allow, you
1:33:11should be properly authenticated. After
1:33:12that, it's going to come back and say,
1:33:13"Okay, cool. Let me see if everything
1:33:15works." Now, this hasn't been perfect on
1:33:17the first try every time, but if you
1:33:19just go back and forth a little bit,
1:33:20say, "Hey, that didn't work. Hey, this
1:33:21is what I'm seeing." It will be able to
1:33:23get you there. It's going to be your
1:33:24best friend for something like this
1:33:26because remember it can read all of the
1:33:27actual documentation. And now it says
1:33:29that the author is working, but we have
1:33:30to enable these APIs in our Google Cloud
1:33:33projects. So basically just clicking
1:33:34open these one at a time and all you
1:33:36have to do is hit enable. So it's super
1:33:38simple. You just have to do this like I
1:33:40said for all of these different services
1:33:42that you actually want to be able to
1:33:43use. So that's why I did this on a new
1:33:44project cuz I just wanted you guys to
1:33:45see that. But if you already have one
1:33:47that has all these enabled, then you can
1:33:48just use that project and generate that
1:33:50OOTH client ID. So there you go. You can
1:33:52see that this works. I said, "Can you
1:33:54find my Google doc that I made in April
1:33:55of 2025?" And I went ahead and pulled
1:33:57links to all five of these because
1:33:59obviously that was a very vague request.
1:34:01And now we could take action pretty much
1:34:02anywhere in Google Workspace super
1:34:04simply with this CLI. But like I said, I
1:34:07just got this set up today and I've been
1:34:08playing around with it a ton in my
1:34:10executive assistant project and it's
1:34:11been awesome. It can literally do
1:34:13anything. So here I'm asking it to grab
1:34:14my unread emails from today and based on
1:34:16what it knows about my business and my
1:34:18priorities, give them a score and if the
1:34:20priority score is below five, just mark
1:34:21it as unread automatically. All right.
1:34:22So, here you can see it said, "Got 30
1:34:24unready emails. Here's my priority score
1:34:26based on your business context." And as
1:34:28I scroll down, you can see that it's
1:34:29getting different ratings. And based on
1:34:31what I'm seeing right now, this actually
1:34:33looks pretty good. So, then I started
1:34:34playing around with Google Slides
1:34:35because I use Gamma right now, but at
1:34:37some point I could imagine that if this
1:34:39gets good enough, then I wouldn't need
1:34:40Gamma anymore. And this is a free option
1:34:42compared to Gamma subscription. So, I
1:34:44had it create me a slide deck and it was
1:34:46okay. I threw in my brand guidelines. I
1:34:48threw in my logo and I said, "Hey, can
1:34:49you see this? you created this using the
1:34:51Google Slides and it's okay, but there's
1:34:53some weird things that I need you to
1:34:54fix. So then it came back and said, I
1:34:56cannot see the slides. I just know how
1:34:58to build them programmatically. So
1:34:59that's why there may be some errors with
1:35:00spacing and stuff. So then I basically
1:35:02just gave it access to ChromeDev tools
1:35:04so that it could open the page,
1:35:05screenshot it, look at it, and then we
1:35:07built a plan to add visual validation to
1:35:09this Google Slide Creator skill. So now
1:35:11you can see as it's going through, it
1:35:13actually takes screenshots and then it
1:35:14can make fixes based on that. So then
1:35:16after it fixes everything, it says,
1:35:17"Okay, cool. Updated the skill. take a
1:35:19look at it now. So, I'll open up this
1:35:21link. Brings me to Google Slides where I
1:35:23have this slide deck. It has kind of my
1:35:26brand colors. It's got the logo up top
1:35:27right. And then as we go through, we can
1:35:29also see that the spacing is a little
1:35:30bit better. It's still not perfect,
1:35:31obviously, but we have custom images
1:35:33here that were generated with Nano
1:35:35Banana 2. And even the images are kind
1:35:37of on brand with the sort of orange and
1:35:39blue color scheme. As you can see, we've
1:35:41got this one with the WAT framework.
1:35:43We've got this slide. And it even ends
1:35:45with the CTA for the free school
1:35:47community. So, just to see what else
1:35:48happens, I'm going to say, "Take a look
1:35:49at the slide deck and do another audit.
1:35:51How could you improve the skill in the
1:35:52future?" So, it's going to go ahead open
1:35:55up a tab as you guys just saw. It's
1:35:57going to take images. It's going to
1:35:58flick through the different slides and
1:36:00capture them. And as you can see over
1:36:02here, it now says take screenshot. And
1:36:04now, it's reading that screenshot right
1:36:06there.
1:36:07Now, it just moved on to the next slide.
1:36:09And it's going to go through and look at
1:36:10every single slide. And then, it's going
1:36:11to come back with a plan. And we could
1:36:13probably do a similar visual and
1:36:14validate flow with creating Google Docs
1:36:17as well. So now you can see it's almost
1:36:19on to that last slide. And I hope it
1:36:20fixes this last slide because what you
1:36:22can see here is that the spacing is
1:36:24really off down here. So you can see it
1:36:26came back with an audit. It came back
1:36:27with some future improvements. And one
1:36:29thing that I did notice is that because
1:36:31I made the window smaller, its
1:36:32screenshots were probably worse quality.
1:36:34So it said presentation mode screenshots
1:36:36would probably be better. But anyways, I
1:36:38just wanted to give you guys a little
1:36:39taste of how you can use the GWS CLI.
1:36:42but also use it with other tools to make
1:36:44the functionality even more powerful.
1:36:46And there is one more important thing I
1:36:47need you guys to understand about
1:36:49connecting your tools. So what you might
1:36:51have noticed is that in cloud code on
1:36:53the desktop app, if you go to customize,
1:36:55you can see some skills in here, but
1:36:56then you can go to connectors. And what
1:36:59are connectors? They're basically a
1:37:00really simple way for you to connect to
1:37:03different tools. If I click on browse,
1:37:04you can see we've got Reend, Tableau,
1:37:06Canva, Figma, Gmail, notion. And this is
1:37:09really easy because you basically just
1:37:11connect and then it prompts you to sign
1:37:12in. You know, like you kind of ooth in.
1:37:15Oh, that one's already installed. So, if
1:37:17I did Notion, for example, it would just
1:37:18bring me to Notion and it would want me
1:37:20to sign in here and then my Claude Code
1:37:22desktop app would automatically be
1:37:24synced up to notion and I could just
1:37:25interact with it. And it's very easy
1:37:27because you don't really deal with, you
1:37:29know, potentially the API keys in the
1:37:30same way. As you can see, here's where I
1:37:32would choose which workspace I would
1:37:33request connection to. Now, the reason I
1:37:35wanted to bring this up is because it's
1:37:38easier and a lot of people might just
1:37:39tell you to do that. But here's what I
1:37:41want you to think about. If you rely on
1:37:43these connections, that is not great
1:37:46because what's going to happen is if you
1:37:48switch off to a different desktop app or
1:37:50a different harness, you lose
1:37:51everything. So by going through the
1:37:54method that we've talked about where you
1:37:55open up your files and you go to yourv
1:37:57and you put your API keys in there and
1:37:59you build connections through that that
1:38:01means later if you want to switch to
1:38:02codeex which is chatgbt's coding agent
1:38:05or if you want to switch to Hermes agent
1:38:08or openclaw or any other new tool that
1:38:09comes out then you'd have to reconnect
1:38:11every single thing which is a huge pain
1:38:13or even if later you decided you wanted
1:38:15to go from cloud code desktop app to VS
1:38:17code which is something that I I like to
1:38:18use a lot as well you would have to
1:38:20reset all those connections again that
1:38:22is why I'm teaching teaching you guys to
1:38:23do this all a little bit more manually,
1:38:25but it's not too difficult at all, but
1:38:26more manually by setting up yourv um API
1:38:29keys in there because it's just way more
1:38:31transferable. You're you're way more
1:38:33tool agnostic and becoming tool agnostic
1:38:35is the most important thing you can do
1:38:37because these tools evolve so quick. New
1:38:39models, new tools every day, every week.
1:38:41So that is why I want you guys to think
1:38:42about the fact that all we're doing here
1:38:44is we're building a bunch of folders and
1:38:46files. any agent, any AI can sit on top
1:38:49of folders and files and look at them
1:38:51and be just fine and use them. So, that
1:38:53is super important to me and I just
1:38:55wanted to make sure that I communicated
1:38:56that to you guys because I knew some
1:38:58people might be thinking, okay, well,
1:38:59why would we not just use the connectors
1:39:00inside of the desktop app that are so
1:39:02much easier to connect? That's why if
1:39:04you want to go for it, but just think
1:39:06about when you have to switch or if you
1:39:07want to switch later, that's going to be
1:39:08a real pain for you to reconnect
1:39:10potentially like 20 or 30 or 40 tools.
1:39:13All right, awesome. Let's move on to
1:39:15skills, which is my favorite Cloud Code
1:39:17feature ever. Skills exist regardless of
1:39:21really whatever AI model you end up
1:39:23using. So, when you're building these
1:39:24skills, these are also tool agnostic,
1:39:26which is great, but skills are so so
1:39:27important. So, let me find some blank
1:39:30space over here so I can draw out some
1:39:31stuff about skills. We're going to grab
1:39:33this little cloud code crab, place them
1:39:35down here. So, when you're talking to
1:39:37Claude Code, the idea is that when you
1:39:40ask it to do something, it just
1:39:42understands you, right? knows your
1:39:44processes and it knows what to do and
1:39:46specifically if it does a process once
1:39:48good or well then you would hope that if
1:39:51you ask it to do it again next week it
1:39:53would do it just as well if not better.
1:39:56So the idea is that we have a bunch of
1:39:58skills, right?
1:40:00Let's just call Yeah, we'll just say
1:40:02skill. And these are also going to be MD
1:40:03files. So let's say skill.md as we know
1:40:05that that stands for markdown.
1:40:08Think about it like this. These are just
1:40:10recipes. If you wanted to make chocolate
1:40:13chip pancakes, I don't know why I always
1:40:14use this example, but it just makes
1:40:15sense. You would open up a recipe for
1:40:18chocolate chip pancakes. If this was
1:40:20your first time ever doing it, you would
1:40:21follow that recipe because you don't
1:40:23know off the top of your head the
1:40:24measurements. um how long to cook it.
1:40:27You know, you don't know exactly what to
1:40:28do, but what you would do is you'd open
1:40:29up the cookbook or you'd Google on our
1:40:32phone the recipe and then you'd follow
1:40:34it to a tea. You would see step one,
1:40:36step two, step three, step four, and you
1:40:37would just do that. After you finished
1:40:39making, you know, you got the
1:40:41deliverable from that recipe. You would
1:40:42say to yourself, "Okay, did I like this
1:40:44or did I not?" You know, maybe I wanted
1:40:45to add more chocolate chips. So, I'm
1:40:47going to change the recipe and add more
1:40:48chocolate chips. So, then you would
1:40:50change the recipe and then next time you
1:40:52want to make chocolate chips, you'd open
1:40:53this back up. You'd run it again and
1:40:55you'd see if you liked it and eventually
1:40:56you get to a spot where you like the
1:40:58skill or you like the recipe and now
1:41:00whenever anyone asks you, hey, you know,
1:41:02can you make me some chocolate chip
1:41:03pancakes? You know exactly how to do it
1:41:04because you reach for the skill and you
1:41:06just follow the instructions. So that's
1:41:08what our agents do. What ends up
1:41:09happening is we have tons of different
1:41:11skills and when we ask our agent to do
1:41:13something, it is able to think to
1:41:15itself, oh, okay, cool. So, this person
1:41:17wanted me to do a, you know, a morning
1:41:19briefing. And all I have to do now is I
1:41:22have to go grab this morning briefing
1:41:23skill, read it, and then I'm going to go
1:41:26execute, you know, the actual morning
1:41:27briefing on my workspace, you know, so
1:41:31on my laptop. Anyways, so it would read
1:41:33this and then it would execute. And then
1:41:36your job as the human is to say, "Okay,
1:41:38cool. This is pretty good." But once
1:41:40again, here's my feedback. And then you
1:41:42just get into this place where the agent
1:41:44would then loop. it would update the
1:41:45skill and then you run it again and then
1:41:47you judge the output, you loop, you run
1:41:49it again. So take a look at this
1:41:50example. This is my grill me skill. What
1:41:53you'll notice is there's this little
1:41:55front matter which is called YAML front
1:41:57matter. It it's yl front matter. I don't
1:42:00know exactly what the yl stands for but
1:42:02it's front matter. And this is important
1:42:04because this is what the agent reads in
1:42:08order to decide should I use the skill
1:42:10or not. So this skill is called grill
1:42:12me. It says, "Enter the user
1:42:13relentlessly about a plan, design, or
1:42:15topic, checkpointing every answer to a
1:42:17brainstorm file so nothing is lost. Use
1:42:19when the user wants to stress test a
1:42:21plan, get grilled on a design, run a
1:42:23brainstorm or discovery session, extract
1:42:24what's in their head into a doc, or
1:42:26says, "Krill me." I'm going to do a full
1:42:28section later on the skill. It's it's a
1:42:29it's a great skill. But this is what it
1:42:32looks like. So, let's say I say, "Hey,
1:42:34you know, Claude Code, I want to um
1:42:36build out a new course on cloud code for
1:42:40knowledge work." for example, which is
1:42:42actually I did this exact thing when I
1:42:44was planning out this course. I said,
1:42:46can you please grill me about the way I
1:42:47use cloud code, the way I think about
1:42:49it, and what I need to know. And then
1:42:51what it did is it just asked me tons and
1:42:52tons of questions. But before it did
1:42:54that, it had to read the skill and
1:42:55invoke it. And it and the reason I
1:42:58wanted to tell you guys about this front
1:42:59matter is because that is how it
1:43:00decides. Look at this, right? If I go to
1:43:02my Herk 2 project and if I go into
1:43:04mycloud and I go to the skills folder, I
1:43:07have tons of skills. There's like 40ome
1:43:09in here. So, because there's so many
1:43:12skills, what this looks like more like
1:43:14in reality is our agent has to look
1:43:17through
1:43:18so many skills to be able to pick out
1:43:20the right one, which is overwhelming and
1:43:23it starts to cost tokens. So that's why
1:43:26we use this front matter which it's a
1:43:28process called progressive disclosure
1:43:30which basically means the agent is able
1:43:32to quickly scan all of the front matter
1:43:34for every single skill and based on the
1:43:36description decide ah this is the one
1:43:39that I need to use based on the request
1:43:41that the user just requested from me. So
1:43:44that is why it would then pick this one
1:43:46then it could read everything and then
1:43:47it would invoke it. So that's how sort
1:43:50of the skills look. That's how they
1:43:52work. And that is how progressive
1:43:54disclosure works. So let us go ahead and
1:43:58build our own skill. The cool thing
1:43:59about skills is they can be invoked by
1:44:01natural language. Like if I said, "Hey,
1:44:02can you just grill me?" Or, "Hey, can
1:44:04you run like my storm research?" They
1:44:06would be invoked. But you can also
1:44:07invoke them from slash commands. So if I
1:44:09do a slash, you can see that I've got a
1:44:11command called session handoff, which I
1:44:12could invoke by doing slash. Or I could
1:44:15use, you know, slashgrill me, for
1:44:17example, which is the one that you guys
1:44:18just saw. So you can use slash commands.
1:44:21And what else is cool about that is
1:44:22there is a slash command from Enthropic
1:44:24called skill creator. Create new skills,
1:44:25modify and improve, blah blah blah. So
1:44:27I'm going to call on that skill and then
1:44:29I'm just going to in my natural language
1:44:31tell it what I want to build a skill
1:44:33around. Before I do this, there are two
1:44:35different ways that you can actually
1:44:37write skills or make skills. So let me
1:44:40once again come over here and show you
1:44:41guys visually. There are two ways. One
1:44:44of them is you kind of proactively are
1:44:46building skills and the other one is you
1:44:48build a skill after the fact. So what I
1:44:50mean by that is let's say we decide okay
1:44:53I want a skill for my morning brief. So
1:44:55what I'm going to do is I'm going to say
1:44:57hey Claude code help me build a skill to
1:45:00help me with my morning briefings.
1:45:01Here's what you're going to do. Step one
1:45:03is you go to my calendar and you see
1:45:04what's on my my day. Step two is you go
1:45:06to my ClickUp. You see my tasks that are
1:45:08due today. Step three is you read
1:45:10through my conversations and you see if
1:45:11there's anything that I committed to
1:45:13blah blah blah. That is how you build
1:45:14the morning briefing skill. Or what you
1:45:16could do on the other side is you just
1:45:19do that. So instead of planning out and
1:45:21saying, "Hey, help me build the skill,"
1:45:23you just do the thing. So you would do
1:45:25step one, you would do step two, you do
1:45:27step three, you do step four. You do all
1:45:28of that manually with Claude. And then
1:45:30you would say, "Hey, look back at all
1:45:33those four steps that we just did. That
1:45:35is a process that I do every morning.
1:45:37So, can you turn that into a skill? So,
1:45:39you basically have those two paths. You
1:45:41can either say, "Hey, help me write a
1:45:43skill." And then explain 1 2 3 4. Here's
1:45:45what we do. Or you actually do the
1:45:47action and then you say, "You know what?
1:45:48That would be a great skill. Turn that
1:45:50into a skill for me." So, let me show
1:45:53you. Um, we're going to proactively
1:45:54build one just to keep this simpler. And
1:45:56also, what I'm going to do is I'm going
1:45:57to clear out this conversation because
1:45:58we don't need all this in here. So, I
1:46:00can do a slash command to clear
1:46:02slashcle. And then we have a fresh chat.
1:46:04So, I'm going to invoke the skill
1:46:06creator skill. And now, let's think of a
1:46:08good use case to build a skill around.
1:46:11Hey Claude, I would like you to help me
1:46:13build a skill. So, I want something
1:46:14that's going to help me sort of just
1:46:16manage my email inbox a little bit
1:46:17better. You should be able to touch my
1:46:19inbox because we set up the Google
1:46:21Workspace CLI. So, that's the first step
1:46:24is see if you can reach it. And then
1:46:26what I want the skill to do is basically
1:46:27give me a rundown of all of my unread
1:46:30emails and then help me sort of just
1:46:32like triage them. Tell me if anything's
1:46:33important. tell me if you know there's
1:46:36anything that I committed to that I need
1:46:37to respond to and if I need to reach out
1:46:39to anyone else on my team basically just
1:46:41help me triage these emails and label
1:46:43them for me high priority you know which
1:46:45ones needs actions which ones can be
1:46:46completely ignored and then after that
1:46:49happens and after you've given me a
1:46:51quick brief we'll actually take action.
1:46:52So I'll say hey you know all of those
1:46:54down there you can just mark them as
1:46:55unread for these three emails I want you
1:46:57to help me create a draft blah blah
1:46:58blah. So you're basically going to help
1:47:00me triage my inbox and clean it up.
1:47:03Okay, so that was a very messy prompt,
1:47:05right? And that's why sometimes it's
1:47:07better to build a skill after you've
1:47:09done something. So, what we're going to
1:47:10do here is it's going to start to build
1:47:11out the skill thing. And then what it's
1:47:13going to do is we're going to run
1:47:14through this example and then we're
1:47:16going to go back and improve the skill.
1:47:17The first step is that it's checking the
1:47:19Google Workspace to see if it can
1:47:20actually hit my inboxes. There we go.
1:47:22Reachable. And there are 201 unread
1:47:24emails. So, I certainly do need the
1:47:26skill. And now what it's doing is it's
1:47:28saying, "Okay, so I have everything set
1:47:29up, but before I write the skill,
1:47:31there's a few questions that I have for
1:47:32you." So, it's asking me some questions
1:47:34to make this skill better. I have 201
1:47:36unready emails. How much should we
1:47:38triage in one run? I'm just going to say
1:47:40the most recent 25 to 30 just to keep
1:47:42this simple. When the skill labels an
1:47:45email, what actually should happen? Read
1:47:47Gmail labels. Brief only, no labels,
1:47:50stars plus brief. I'm going to go ahead
1:47:52and say brief only, no labels, just so
1:47:54it's not doing anything that I'm not
1:47:56explicitly approving, at least to start.
1:47:58Once we've ran the skill 10, 20, 30
1:48:00times and we've kind of like battle
1:48:02tested it and we feel more confident in
1:48:04it, then we can maybe make it a little
1:48:05bit more autonomous. But to start, I
1:48:07like to be in full control of
1:48:08everything. What categories should the
1:48:10brief sort emails into? Let's just go on
1:48:12high, action, ignore. That's totally
1:48:13fine for now. And it's going to keep
1:48:15building out this skill. So, while this
1:48:16is finishing running up, let me just
1:48:18talk about some of the common questions
1:48:19I get around skills because skills can
1:48:22be as simple as like a two- sentence
1:48:24prompt, but they can also be pretty
1:48:25complex. So, let me show you guys a
1:48:27quick example. If I go into my skills
1:48:30here, let me show you an example of one
1:48:32that is kind of more on the complex
1:48:34side. Okay, so this is one called
1:48:36packaging, which helps me out with
1:48:37sometimes packaging up YouTube videos
1:48:39and stuff like that and brainstorming.
1:48:40So, it's called packaging. Here's the
1:48:42description of when to use it. And then
1:48:44what you'll notice inside of this skill
1:48:46is we're routing to other files. So,
1:48:48sometimes it needs to read this full
1:48:49packaging playbook, which is a massive
1:48:52um markdown file of packaging rules. So,
1:48:54it says read this every time. It has
1:48:55Nate's actual decision logic. Then it
1:48:57went to the decision logic source. So
1:48:59another massive markdown file. Then it
1:49:02can also read my channel data. And then
1:49:03it can also look at thumbnail analysis.
1:49:05It can also look at a script to generate
1:49:06thumbnails. And it also has to find my
1:49:08API key. So this is an example where we
1:49:11have tons of brand assets like all of my
1:49:13thumbnail assets, the core model. And
1:49:16look how long this is. There's so many
1:49:17different instructions. There's so many
1:49:18different modes. There's so many
1:49:20different decisions to be made. And
1:49:21every time that I use the skill, it gets
1:49:23better and better, right? because, you
1:49:25know, I give it that feedback loop.
1:49:27There's lots of files that it
1:49:28references. Here's one called idea
1:49:30mining. So, we use this when someone
1:49:31asks for content ideas, video ideas,
1:49:33what to make next, or just to run idea
1:49:34mining. So, once again, this actually
1:49:36needs to be updated cuz I have more than
1:49:37this many subscribers now, but context
1:49:40files, you know, execution. But what's
1:49:42cool about this is this skill calls on
1:49:44sub agents. So, I know we haven't talked
1:49:46about sub aents yet, but skills can do
1:49:48so many things. It can reference full
1:49:50workflows. It can reference full
1:49:51scripts. It can reference sub aents. So
1:49:53the first agent that's supposed to run
1:49:55this skill will call on the YouTube
1:49:57analyzer sub agent and then later it
1:49:59will call on my researcher agent.
1:50:01Anyways, the point I'm trying to make is
1:50:03skills can be really simple. It could be
1:50:04a two sentence prompt. Skills can call
1:50:06other skills. Skills can call sub aents.
1:50:08It's basically whatever process you
1:50:10have. It's basically just an SOP.
1:50:12However you decide you want to do
1:50:13something, a skill just packages it up
1:50:16so that you can consistently do that
1:50:17again every time. Another cool thing
1:50:19about skills is because they're just
1:50:20markdown files, you can take skills from
1:50:22other people. You know, they will have
1:50:24open source GitHub repos like this
1:50:25superpowers plugin that has skills in
1:50:27here. And this gives us skills for
1:50:29brainstorming and for other things like
1:50:31that. If I go to the skills, you can see
1:50:32there's a brainstorming one, there's an
1:50:34executing plans, there is a dispatching
1:50:36parallel agents, there's writing plans.
1:50:38So, you can leverage other people's
1:50:40skills and other people's subject matter
1:50:41expertise by taking the markdown file
1:50:44that they give you for a skill. For
1:50:45example, this brainstorming one, like I
1:50:47said, just a markdown file. It's just a
1:50:49simple prompt and then you can put this
1:50:50into your own project and you can use
1:50:52it. So, if you go on X or if you go on
1:50:54LinkedIn, you'll see people talking
1:50:55about these really cool skills and
1:50:56plugins and it's kind of just like an
1:50:58open marketplace where people are
1:50:59sharing stuff. If you guys remember the
1:51:01grill me skill that I've been talking
1:51:02about, I was inspired to create that by
1:51:04Matt PCO who made this skill called
1:51:06grilling. And look how simple it is.
1:51:08This is literally one of those examples
1:51:09I talked about where this is like, you
1:51:11know, five or six sentences and it's
1:51:13just a prompt, but it's still very, very
1:51:14effective. And once again, I could
1:51:16download this, put it into my own
1:51:18project, and then use it. Okay, so the
1:51:20skill is done. Obviously, it's not
1:51:22perfect yet because we haven't even
1:51:23really tested it. But if I go into
1:51:24mycloud and I go to skills, you can see
1:51:26we now have one called inbox triage. And
1:51:29inside of this inbox triage skill, we
1:51:31have some references, we have some
1:51:33scripts, and then we also have the main
1:51:35markdown file. And what I'm assuming is
1:51:37in this main markdown file, it calls on
1:51:39the references right here. Um, yes,
1:51:41inbox triage scripts. So, it calls on
1:51:43the script and then it also calls on
1:51:45this reference which is the cookbook MD
1:51:47which it's still being written out it
1:51:48looks like. But that's the whole point
1:51:49is that inside of the skill, the
1:51:51markdown file is the master instruction.
1:51:53And sometimes you include other assets
1:51:55like in my YouTube thumbnail skill, I
1:51:58have brand assets in here that needs a
1:51:59call on because it needs to use my, you
1:52:01know, my head shot to create some of
1:52:04those thumbnails. So that is what this
1:52:05looks like. You can see that it's being
1:52:06built once again completely by Claude
1:52:08and it already has like the front
1:52:10matter. It already has the description.
1:52:12So you really don't have to worry about
1:52:13the technical details of how do you
1:52:15actually physically build one. You just
1:52:17have to worry once again about speaking
1:52:19clearly about what you want. Okay. So
1:52:20I'm going to blur out all of these
1:52:22emails obviously, but here are some
1:52:23things that it says high priority. And I
1:52:25would agree that those are high
1:52:26priority. We've got some things that
1:52:27need action. Um this one's something I
1:52:29just need to approve. This is something
1:52:31I need to respond to. And this actually
1:52:33this one could be ignored. And so what
1:52:35I'm going to do is say, "Okay, cool.
1:52:37This looks pretty good for a first pass.
1:52:39The only thing I would say here is the
1:52:42third email in the needs action bucket,
1:52:45you can move that to ignore. So if you
1:52:47ever see an email like that again in the
1:52:49future, just make a note that from this
1:52:52particular sender and for this
1:52:53particular reason, those can be ignored
1:52:55because I don't actually need to action
1:52:57those at all." You can see it says
1:52:58that's the skill. Nothing was touched.
1:52:59All 16 are still on red. Now tell me
1:53:01what you do and then after it makes this
1:53:04change I'm going to say cool go mark all
1:53:06of those that are in the can ignore
1:53:08bucket I'm going to say to mark those as
1:53:10unread in Gmail and then maybe you could
1:53:11say like hey so for those couple that
1:53:13were high priority can you create a task
1:53:15for me in ClickUp or can you create a
1:53:17task for me here and then I will
1:53:19remember to do that by the end of today.
1:53:21So when I talk about the idea that we
1:53:22want to make this thing understand your
1:53:24preferences and your workflows and your
1:53:26business so well skills are the core of
1:53:28that skills are the instructions and the
1:53:30preferences that you can save. So right
1:53:32here you can see it added this rule into
1:53:33the skill.mmd. So the skill.mmd got
1:53:35changed. You can see right here it said
1:53:36edited the skill.md and it added 13
1:53:39lines right there as you can see.
1:53:40Awesome. So for now this skill is
1:53:42complete. What I want you to do and add
1:53:44this to the end of the skills is that
1:53:46once the user has confirmed that all of
1:53:48those emails that are in the can ignore
1:53:50bucket, you can just go ahead and mark
1:53:52those off as unread or sorry, you mark
1:53:54them off as red once the user has
1:53:56confirmed. So go ahead and mark those as
1:53:58red and then update the skill to say
1:54:00that. Awesome. So it has marked those as
1:54:02red. I will check that in a sec. It's
1:54:04also edited the skill. So if I open this
1:54:06up, you can see it added this line right
1:54:08here that said, "How do you clear the
1:54:10ignore bucket?" Once Nate confirms that
1:54:11he's fine with it, you mark them as red,
1:54:13blah, blah, blah. So, that is how we see
1:54:14that it edited the skill. And now, let
1:54:16me check the email real quick. Awesome.
1:54:18So, blurring this out, of course, but
1:54:20you guys can see, well, you probably
1:54:21can't, but the ones that I said to
1:54:24ignore, it went ahead and it marked
1:54:26those off as red. Sweet. So, that is how
1:54:29we built our first skill. Just remember
1:54:31that all of the skills that I use on the
1:54:33dayto-day and on the week to week,
1:54:35they're not done. Every single time I
1:54:37use it, I'm able to give it some sort of
1:54:38feedback. Sometimes you don't, but you
1:54:40always want to look for areas to improve
1:54:43those skills, especially as different
1:54:46models come out. So like Opus 4.7 might
1:54:48behave differently on a skill than Opus
1:54:504.8. So every time a model drops, just
1:54:52run your skills and make sure you still
1:54:54like them. If you ever end up switching
1:54:55your harness, so if you switch from
1:54:57Claude Code to something else like
1:54:58OpenClaw or Codeex, they can still use
1:55:02those skills. You just have to make sure
1:55:03they're in the right folder for them to
1:55:04see. You know how this one is called a
1:55:06Claude? So, for example, um, codeex
1:55:09looks for skills in a codex folder, I
1:55:12believe, or maybe it's a agents folder.
1:55:14So, it's just a little bit about
1:55:15learning what's the terminology for this
1:55:16harness. But because the fact that all
1:55:18of them are just markdown files and all
1:55:20pi python scripts and files and folders,
1:55:22they all transfer over. So, that was
1:55:24your first skill. Now, what you'd want
1:55:25to do is you'd probably want to think
1:55:26about, let me think about my week. Let
1:55:29me think about things that I do that
1:55:30happen based on an event trigger. So,
1:55:32maybe every time a new lead processes a
1:55:34form, what do I do? That's a great
1:55:36opportunity for a skill or maybe even an
1:55:38automation. And then maybe everything
1:55:40like every Monday if I do something or
1:55:42every Friday if I do something, what is
1:55:43that process? And let me turn that into
1:55:45a skill. So you're just going to start
1:55:47stacking up your library of skills. All
1:55:49right. So moving on to number 15 here.
1:55:51We're going to talk about context
1:55:52windows, which is so so so important. So
1:55:56let me find my little graphic here for
1:55:58context windows. This is the one we're
1:56:00going to start with. So what is a
1:56:01context window? Remember how if I open
1:56:04up Cloud Code and I click on this little
1:56:06button in the bottom right, we see right
1:56:08here context window and it says 110,000
1:56:11out of a million, which is about 11% of
1:56:14our context window. I can open this up
1:56:16and see what lives in the context
1:56:17window. So messages are taking up 72,000
1:56:20tokens. System tools are taking up
1:56:2218,000 tokens. Skills take up this many.
1:56:24System prompts take up this many.
1:56:25Whatever. If I go ahead and I do a
1:56:28slashcle, that would wipe everything
1:56:30clean. That's how we get a fresh
1:56:32session. Right now, the idea is that we
1:56:36have a million tokens to play with until
1:56:38the model has to automatically like
1:56:39reset because it can't take all that
1:56:41context. It's too much, you know,
1:56:43information. But the thing is, the
1:56:47models reach something called a dumb
1:56:49zone. So basically, answer quality
1:56:51drifts as the context window fills. So
1:56:55at the beginning, you're super super
1:56:57sharp. And if you guys have ever talked
1:56:59to Claude in chat, you don't see the
1:57:01context window. You basically just keep
1:57:03talking. So if you've been having a
1:57:04conversation with chat with Claude chat
1:57:06for a full day or for multiple days and
1:57:09you ever think to yourself, "This is
1:57:10getting really, really dumb." It's
1:57:12because all of that context is being
1:57:14loaded up. So what that does is it
1:57:15causes Claude to get confused, forget
1:57:17things. It just gets dumber and dumber.
1:57:19And that is called the dumb zone. And
1:57:21it's also called context rot. So a big
1:57:23part of your job as a manager is to
1:57:26manage the context rot. Like you know
1:57:27how truck drivers are only allowed to
1:57:29drive a certain amount of time before
1:57:30they have to like pull over and rest?
1:57:33That's because the longer they're awake,
1:57:35the worse their cognitive function
1:57:36becomes, right? Like they their reaction
1:57:38time gets slower, their judgment gets
1:57:39worse. Same thing with models. So your
1:57:41job is to basically when you get to the
1:57:43point where the model is starting to hit
1:57:46that context rot territory, what you
1:57:48need to do is you need to be able to
1:57:49wipe the context clean without getting
1:57:52rid of the knowledge and like losing
1:57:53progress because that's obviously worst
1:57:55case scenario. So right now at the time
1:57:56of filming this video, the models
1:57:59basically have a context window if we're
1:58:00using cloud code of 1 million. So 1
1:58:03million is the max, right? On this end
1:58:05we have 1 million and on this end we
1:58:06have zero. Now what I like to do is I've
1:58:09typically found that when I get to 250K
1:58:13I like to do a reset. So this isn't
1:58:16drawn to scale, but let's pretend that
1:58:18this green for me is 250,000 about 25%
1:58:21of the full window. And that's when I
1:58:22will reset my context. Now, how do I do
1:58:25that? I built my own custom skill, which
1:58:27I'll give you guys for free. It's very
1:58:28simple, called session handoff. So, I'm
1:58:30going to do session handoff. Go ahead
1:58:32and execute that skill. Now, what that
1:58:35does is it's going to look at everything
1:58:36that we've done and it's going to look
1:58:38at anything that might be open like open
1:58:40decisions, any files that we were
1:58:42editing. Basically, a quick summary of
1:58:44everything that we've done, right? So,
1:58:45you can see we have where it started. We
1:58:48have decisions locked and what shipped.
1:58:50We have key files for the next session
1:58:51running state verification deferred and
1:58:53open questions and pick up here. So then
1:58:55what I do is I copy this message.
1:58:57Remember right now we're at 112,000.
1:59:00And then I do my slash clear. Now our
1:59:03context window is at zero. Then what I
1:59:06do is I paste in that session handoff
1:59:08message. I hit enter. And then the
1:59:10context window is going to fill up a
1:59:11little bit again, but it's going to
1:59:15actually pick up right where we left
1:59:16off. So the whole idea is like let's say
1:59:18you know every single you've got shifts
1:59:21you've got workers that come in and they
1:59:22work on code or they work on projects
1:59:24for five hours at a time when the person
1:59:27is basically closing out their shift
1:59:29they're going to hand over a document to
1:59:31the next engineer and say hey here's
1:59:32what I did here's you know some bugs
1:59:34here's some things to keep in mind here
1:59:35are the files I was working on and then
1:59:37that that new engineer can open up their
1:59:39shift right on the same page so that's
1:59:42basically what the session handoff skill
1:59:43does you can see I'm completely picked
1:59:44up it knows everything because it
1:59:46wouldn't have if I just did a clear. Now
1:59:48look at this. If I open up my context
1:59:50window now, this is at 55,000 tokens.
1:59:53Did this all take 55,000 tokens? No,
1:59:56because system tools by default will
1:59:59take up some. MCP tools will take up
2:00:01some, memory files will take up some,
2:00:03skills will take up some, system prompts
2:00:04will take up some, and then the rest
2:00:06will basically be your messages and
2:00:07other things that you do. But on a blank
2:00:09fresh session, your context window might
2:00:11already be like 30k or 40k because of
2:00:14just what's in your project. And that's
2:00:16a big reason why, remember earlier when
2:00:18I was talking about this whole idea with
2:00:19with uh skills about progressive
2:00:21disclosure because every skill this is
2:00:24basically autoloaded in and that's what
2:00:26your agent can look at. But imagine if
2:00:29every single skill the full markdown
2:00:31file was loaded in. You know sometimes
2:00:32the full skill is you know hundreds and
2:00:34hundreds of lines. So because of the
2:00:36whole progressive disclosure that is why
2:00:38we're able to save a lot of context and
2:00:40only invoke a skill if and when it is
2:00:42needed. So it's all starting to come
2:00:44full circle here. Isn't that pretty
2:00:46magical? Anyways, that is the context
2:00:48window. My best practice is basically if
2:00:51we get past 250,000 300,000, I'm going
2:00:53to do a session handoff and then just
2:00:54paste it into a new chat and keep on
2:00:56working. Cloud code basically has
2:00:58something called autocompact or a
2:01:00/compact feature which is a summarize
2:01:02and keep going. But the autocompact
2:01:04kicks in way too late. It kicks in like
2:01:06way more around like this area where
2:01:08you've already probably gotten stuff
2:01:10worse. And um it also takes a long time.
2:01:13So, the session handoff skill will be in
2:01:14the free school community that you guys
2:01:15can go and grab completely for free. The
2:01:17link for that is down in the
2:01:18description. You're just going to go to
2:01:19the classroom. You'll go to all YouTube
2:01:21resources and then you can grab every
2:01:23single resource that I've ever given
2:01:24away for free. GitHub repos, skills,
2:01:26templates, anything. It's all found in
2:01:28here in my free school community. As you
2:01:30can see, this is also where we have a
2:01:317-day challenge as well as a build your
2:01:33own AI operating system course, which is
2:01:35what I want you guys to all take after
2:01:37you finish this course because it's
2:01:38really going to help you level up. But
2:01:40this is what I want you guys to do next.
2:01:41So go ahead and request to join the free
2:01:42school community. We've got almost half
2:01:44a million members in here. It is a great
2:01:46place to be. You can ask questions,
2:01:48collaborate with people, stuff like
2:01:49that. So the context window theory is
2:01:51really important to understand because
2:01:52there's a lot of things that we're going
2:01:53to learn later in this course that help
2:01:54you protect that context window and help
2:01:56you keep this as lean as possible
2:01:58because also all of this contributes to
2:02:00your 5-hour limits and your weekly
2:02:02limits. So managing context and managing
2:02:05tokens is a really important thing to
2:02:07get good at. But right now, I just
2:02:08needed you to understand, you know,
2:02:10clearing and context rot and where you
2:02:12can, you know, get visibility on this
2:02:14stuff. And now we're able to take that
2:02:16knowledge into the rest of this course.
2:02:18All right. So, let's talk about memory.
2:02:20What else can you do to make your agent
2:02:22actually remember things about you in a
2:02:25way where you know you're not repeating
2:02:27things? So, let me go back into cloud.
2:02:29I'm going to go to my Herk 2 project
2:02:30because this is just where I have tons
2:02:32and tons of stuff, right? I am currently
2:02:34making a course right now which actually
2:02:35you're aware of my knowledge work cloud
2:02:37code master class and I'm trying to
2:02:39explain the concept of memory to the
2:02:40audience. So what I want you to do right
2:02:42now is tell me where do you look in my
2:02:45project or globally what files where do
2:02:48you look when you need to find memories
2:02:50about me and things like my preferences
2:02:52and just having a bit of an automemory
2:02:55sort of feature so that we're
2:02:57continuously growing our relationship
2:02:58and you get smarter over time. Where do
2:03:00you actually look? Because I think of
2:03:01memory as a couple different things.
2:03:03Memory in my mind is first of all like
2:03:05the actual chat. So the fact that it
2:03:07remembers what we're talking about in
2:03:08one session. Then we also have the
2:03:10ability for it to remember like
2:03:11preferences, to remember decisions
2:03:13you've made, to remember meetings you've
2:03:15had, that sort of idea. So kind of like
2:03:17building cloud code as your second
2:03:19brain. So you guys remember earlier when
2:03:20I showed you this Obsidian sort of
2:03:22second brain of mine. All of these are
2:03:24not only knowledge, but I would also
2:03:25consider these memories because it has
2:03:27my YouTube transcripts, it has my
2:03:28meeting transcripts, it has decisions
2:03:30I've made, it has my chat threads, it
2:03:32has things that I found important enough
2:03:34to want to store so that Claude could
2:03:36also pull them. Okay, so let's take a
2:03:37look at what it said. It said the
2:03:40automemory folder, and this is something
2:03:42really cool. Cloud has auto memory which
2:03:44basically means after a certain amount
2:03:46of sessions or a certain amount of time
2:03:48cloud will look at what you've done and
2:03:49it will write memory to the auto global
2:03:53scope. So here is in the users in the
2:03:55nates in the docloud in the projects we
2:03:57have memory. It's a folder of 72
2:03:59individual markdown files one fact per
2:04:02file. Each file has a name like feedback
2:04:04noel hooks.mmd or reference corporate
2:04:06structure.mmd. A file holds one thing, a
2:04:08preference that you've corrected me on,
2:04:10a project state, a reference fact about
2:04:12your setup. The key file in there is
2:04:14called memory.mmd. So remember we have
2:04:16our claw.mmd. We also have a memory.mmd
2:04:18and that is the index. Those are tagged
2:04:20with front matter. So user, feedback,
2:04:22project and reference. Then it also
2:04:24considers the claw.mmd files as memory.
2:04:27So once again, we've got the global one
2:04:28and we've got the project one and we've
2:04:30got the private local one. I don't
2:04:31really touch these too much, but that is
2:04:33a third one you can you can pull. I'm
2:04:34not going to talk about this right now,
2:04:35but that's more so like when you're
2:04:37collaborating with a team, if you want
2:04:38to keep one cloudmd that's just for you
2:04:40and then one that's sort of like this is
2:04:42the general project system prompt. So,
2:04:44you know, those are the two differences.
2:04:45And then the herkbrain wiki, which is
2:04:47basically this thing that I just showed
2:04:49you guys, this is the wiki of
2:04:50everything. None of this is autoloaded.
2:04:53What happens is if it realizes, okay, I
2:04:55don't understand based on my memory,
2:04:56based on the cloudmd, or based on
2:04:58skills, I will look here. I will look
2:05:00here for things like team, finances,
2:05:02metrics, strategies. That is how this
2:05:04one works. But once again, the important
2:05:05mental model cloudmd is the rules.
2:05:08Memory is, you know, learned facts. And
2:05:10the cloudmd, if you guys remember,
2:05:12actually, you know what? I'll just open
2:05:13it up again. What was really important
2:05:15to me is that I treat this cloudmd as a
2:05:17router. Meaning cloud reads through this
2:05:20and it understands if I need this, I go
2:05:22here. If I need this, I go here. If I
2:05:23need this, I go here. If I need this, I
2:05:25go here. And that is how it's able to
2:05:27have this feeling of memory. Now, how do
2:05:30you actually turn on automemory? Is it
2:05:32enabled by default for everyone? How
2:05:34does that work? Because what you'll
2:05:35notice here is in cloud on the desktop
2:05:37app, if I go to do a slashmemory,
2:05:39we don't have slashmemory. We have
2:05:41consolidate memory, but we don't have a
2:05:44just a regular command called
2:05:45slashmemory. But if I go into my cloud
2:05:47code on the VS Code terminal, I'm just
2:05:51switching back to a different model
2:05:52here. We do have a slashmemory. If you
2:05:53see this, I can go slashmemory. And then
2:05:56right here, I can turn automemory on or
2:05:57off. So earlier in this video when I
2:06:00talked about where should you use cloud
2:06:01code for the majority of this video
2:06:03we're using the desktop app just because
2:06:04it's easy to understand and it just has
2:06:05a nice UI. But the one limitation about
2:06:08using it in something like the desktop
2:06:10app is there are a few a few very niche
2:06:13slash commands that you don't get for
2:06:15the majority of your driving. 99% of the
2:06:17time you don't need these commands but
2:06:19sometimes you know there's just little
2:06:21tiny things where the terminal version
2:06:25has more functionality than the others.
2:06:26So, it's good to get familiar with the
2:06:28terminal every once in a while, but like
2:06:30I said, for the most part, you are okay.
2:06:32And you can also see right here that it
2:06:33is on by default because once again,
2:06:36Claude itself has skills to read its own
2:06:39documentation. So, if you ever have a
2:06:40question about how Claude works under
2:06:42the hood or trying to figure out
2:06:43something about your setup, just ask
2:06:45Claude to to figure it out, to help you
2:06:47research it. Another big mindset shift
2:06:48here is treating this thing like a
2:06:50mentor, not just like a an engineer, not
2:06:53just like your your best friend. It is
2:06:54the smartest person you know and it is
2:06:56your mentor. So you can ask it
2:06:57questions. You can ask it why'd you do
2:06:59that? What would happen if you didn't do
2:07:00that? What was that tool call you did?
2:07:02Why did you need to do that tool call?
2:07:03All these sorts of questions. Being
2:07:04genuinely curious is what helps you get
2:07:06way more out of this thing. So anyways,
2:07:08automemory should be turned on by
2:07:10default for you guys. But if you want to
2:07:11drill in, you can find the actual memory
2:07:13files because they're just markdown
2:07:14files and you can go look at them. So
2:07:16remember this is the path that it told
2:07:17me to go to to see this. So, I'm just
2:07:19going to copy this, open up an explorer,
2:07:22paste that in, and in here we have the
2:07:2472 markdown files that it was referring
2:07:25to. So, like feedback LinkedIn balance
2:07:28line, um, feedback team names from wiki.
2:07:30So, I just opened one up real quick.
2:07:32This is a feedback internal doc visual
2:07:34style. It's got description, it's got
2:07:35metadata, and then it stores a memory,
2:07:37which is basically this is what Nate
2:07:39told us, why, how to apply, blah blah
2:07:41blah. So, this is super super cool. All
2:07:44right, so let's talk about AI slop. AI
2:07:46slop means different things to different
2:07:48people. Some people think AI slop means
2:07:50AI generated images that look bad or
2:07:52those Tik Toks you might see where
2:07:54there's like two AI generated fruits and
2:07:56there's like a love story between them.
2:07:57Whatever you consider AI slop, let's
2:07:59talk about it real quick because how I
2:08:01define it is basically when I can
2:08:04clearly tell something was generated by
2:08:05AI. Now, I don't necessarily think
2:08:08that's a bad thing because people should
2:08:11be using AI, but I do think you get to a
2:08:13point where you start to lose trust in
2:08:15people. And the most important thing is
2:08:16that you are judging all of your work
2:08:18and you want people to look at it and be
2:08:20like, you know, I don't care if he used
2:08:22AI or not, but I trust that he checked
2:08:24this and I trust that because he signed
2:08:26his name to this, he is taking
2:08:27accountability and responsibility for
2:08:29it. So, I actually wrote up a little
2:08:31post which I'm going to read for you
2:08:32guys word for word. The real problem
2:08:33with AI slop. So, I'm sure you guys have
2:08:35heard the term AI slop and everyone sort
2:08:37of defines it differently. Maybe you
2:08:38think of those Tik Toks, blah blah blah.
2:08:40I use that example again. But I want to
2:08:41talk about it in the context of
2:08:43communication, internal, external,
2:08:45content you put out in the world. I
2:08:46write my LinkedIn post with AI. My agent
2:08:48knows my business, how I write, how I
2:08:50speak. That's just how I work now. And
2:08:51there's nothing wrong with that because
2:08:52I think that everyone should be using AI
2:08:54to write if it makes them more
2:08:55efficient. But this isn't a binary yes
2:08:57or no. It's a spectrum. Sometimes AI can
2:08:59draft and send automatically. Most of
2:09:01the time I want it to just draft and
2:09:03then I review. If someone sends me an
2:09:05email with M dashes everywhere, I don't
2:09:06actually care at all that they used AI.
2:09:08But the fact that I can clearly tell
2:09:10it's AI generated isn't the problem.
2:09:11What I do start asking is, did they even
2:09:13proofread this? Is this even accurate?
2:09:16And subconsciously, I might start losing
2:09:18trust, not in the email, but in the
2:09:20person who sent it. Our job here has
2:09:21changed from writer to reviewer. And
2:09:23this quote really stuck with me. You can
2:09:24outsource your thinking, but you can
2:09:26never outsource your understanding. When
2:09:28your name is attached to the content,
2:09:30you take credit if it lands, as you
2:09:31should, but that also means you need to
2:09:33take accountability if it's incorrect.
2:09:34Taste and reviewing is becoming more
2:09:36important than ever. And maybe that
2:09:38should be R. So, you can tell I didn't
2:09:41write this with AI. AI is super
2:09:42intelligent and powerful, but I don't
2:09:44want to see a world where we trust AI so
2:09:45much that we stop reviewing things and
2:09:47then the human on the other end of the
2:09:49content starts losing trust in us.
2:09:50That's why even though I write with AI
2:09:52and people know that, I still try my
2:09:54best to disguise it and make it sound as
2:09:56innate as possible. And like that's
2:09:58exactly why I wanted to show you guys
2:09:59this AI phrase kill list, which I kind
2:10:01of showed earlier. But you should be
2:10:03building up something like this and you
2:10:05should be building up the way that you
2:10:07speak. So I've obviously got skills for
2:10:08helping me write LinkedIn posts or
2:10:10YouTube scripts or um emails, internal
2:10:13communication. And I want my
2:10:14communication to sound like me. And I
2:10:16want people to trust that what I'm doing
2:10:18is me. Because it's not just about
2:10:19communication. It's also about if you
2:10:21send over a report or if you send over
2:10:23some analytics or if you send over, you
2:10:25know, you're writing up a case study or
2:10:26the copy that's on your website. You
2:10:28don't want any of that to clearly be AI
2:10:30generated because once again, trust is
2:10:32like the biggest currency and you don't
2:10:34want to lose that currency. And not only
2:10:36for you, but it's important to
2:10:37communicate this down to your team when
2:10:39you guys are all as an organization
2:10:40learning how to use AI better. So
2:10:42anyways, just a really quick section. I
2:10:44hope that that hit and I hope that that
2:10:45little change of pace real quick was
2:10:46helpful to you guys and just kind of
2:10:48made you think about the way that you
2:10:50really truly want to be using your AI
2:10:51agents and how the teams should be using
2:10:54them together. And looking back at these
2:10:55mindset shifts again, you'll notice that
2:10:56number three, you can outsource your
2:10:58thinking but you cannot outsource your
2:10:59understanding is one that I referenced
2:11:01in that post as well because I think
2:11:03that one is just a really really good
2:11:05one of how the future is going. You
2:11:08know, like we're able to outsource the
2:11:09ability for agents to go do the research
2:11:12and to grab a bunch of sources and give
2:11:14you some sort of consensus, but you
2:11:16still have to read that. You still have
2:11:17to understand it and you still have to
2:11:18know how to apply it. Okay. Now, this
2:11:20next section of the actual kind of
2:11:22agenda up here is about picking what to
2:11:24build because that's a huge pain point
2:11:26that I hear from my audience when
2:11:27they're, you know, hey, I'm just getting
2:11:28started. I don't know what projects to
2:11:29do. Like, where do I start? So, let's
2:11:32look at just these mindset shifts real
2:11:33quick on the method. the method about
2:11:36how to decide five mindset shifts real
2:11:38quick that I'm just going to read off
2:11:39the constraint is the only place where
2:11:41work compounds everywhere else you're
2:11:43just busy automate a broken process and
2:11:46you don't fix it you scale it becoming
2:11:48AI native doesn't mean using AI for
2:11:50everything it means being a problem
2:11:52solver who sometimes uses AI you can't
2:11:55automate what you can't map every build
2:11:57needs a north star one number picked
2:11:59before you build that tells you whether
2:12:01the build was worth it and if these
2:12:02mindset shifts are interesting to you
2:12:03guys then you should definitely Check
2:12:05out the book. It's called Becoming AI
2:12:06Native. But anyways, what do these mean
2:12:08to me? The first one, the constraint is
2:12:10the only place where work compounds. We
2:12:12see too often people, whether you're
2:12:14working with a client or whether you're
2:12:15trying to automate your own business or
2:12:17whether you're trying to help yourself
2:12:18out, automating things that aren't
2:12:20actually being used very often. They're
2:12:22automating things that sound sexy,
2:12:23right? Like, oh, we need this sales
2:12:25agent to do all this. And they start to
2:12:26try to automate that or we need, you
2:12:28know, this fancy thing that I saw a demo
2:12:29of on LinkedIn. But really what you need
2:12:32to think about and the way that I always
2:12:33like to think whether I'm thinking about
2:12:34my processes and my team's processes or
2:12:36whether I'm consulting with a business
2:12:38owner. The way I think is let's take the
2:12:41flow of your business right like let's
2:12:42imagine it as a pipe which is you know I
2:12:44think a great analogy. So let me just
2:12:46draw a pipe real quick. This is your
2:12:48pipe and ideally what happens is you
2:12:50have water coming in the front and the
2:12:52water flows through and this is your
2:12:54business. When water comes out the other
2:12:55side, this is you basically this is your
2:12:57profit because things might happen in
2:12:59the middle and you know you have to pay
2:13:00people, you have to do whatever, you
2:13:01lose clients, they turn out this over
2:13:03here, all of this is profit and you want
2:13:05to maximize on both ends really how much
2:13:07water comes in and what percentage of
2:13:09that comes out. Now what we want to be
2:13:11thinking about are in this process where
2:13:13are the clogs, where are the constraints
2:13:16that sit here that basically make less
2:13:18water go through? And there's not just
2:13:20going to be one. You know, you're going
2:13:21to have multiple constraints and they're
2:13:22going to be different sizes and
2:13:23different priorities. But the point I'm
2:13:25trying to make here is I like to work
2:13:27from front back. I like to work with the
2:13:30first constraint. So I say tomorrow if
2:13:33you got 5x the amount of customers going
2:13:36through your pipe, so 5x the amount of
2:13:37water, what would break first? And that
2:13:39forces the business owner to walk
2:13:41through in their mind what do they do on
2:13:42the day-to-day? What does the team do?
2:13:44Where is water being stopped? And how
2:13:46can we, you know, basically that's the
2:13:48problem. That's the only way you grow a
2:13:50business is if you are attacking these
2:13:52constraints. So by attacking the
2:13:53constraint that is going to not only
2:13:55make the automation more powerful and
2:13:56and more successful for you, but it
2:13:58gives you a clear road map of where to
2:14:00start. And then guess what? Once this
2:14:02bottleneck has been unclogged, once the
2:14:04clog has been unclogged, now all the
2:14:05water is moving to the next constraint.
2:14:07And even though some more's, you know,
2:14:08sliding through right here, you're still
2:14:10having a big clog. And you want to get
2:14:12rid of that. And then after you get rid
2:14:14of these two, for example, guess what
2:14:16happens? Because there's more demand up
2:14:17front. another clog pops up. So this is
2:14:20a never- ending cycle, but that gives
2:14:21you at least a framework to work from.
2:14:23So that's why the constraint is the only
2:14:25place where work compounds because if
2:14:27you have, let's say, this main big clog
2:14:29up front and you've got a bunch of
2:14:31little ones in the back, oops,
2:14:34then does it make sense to start
2:14:36eliminating these? I don't know. I mean,
2:14:38you could argue that yes, so that when
2:14:39you open this one, water flows through.
2:14:42But really, this is just like you want
2:14:43to go for the constraints first. It's a
2:14:45theory of constraints. And then I just
2:14:47wanted to loop this back to number nine,
2:14:49which is every build needs a north star.
2:14:52So when you're deciding on a problem
2:14:53that you want to attack, before you
2:14:55build it, you want to think about what
2:14:57is the metric that you're looking at and
2:14:59which way do you want to move it. So
2:15:01let's say right here we realize that the
2:15:03clog is actually like up up front,
2:15:05right? Like there's just not much
2:15:07business coming in the front of the
2:15:08funnel at all. So we basically just have
2:15:10no water in the system and that's the
2:15:11problem. Okay. Well, let's say the
2:15:13metric here that we want to fix is um we
2:15:16are getting about five leads a week. So,
2:15:20we would think about this. Five leads a
2:15:23week is our metric and our north star is
2:15:26we want to design a system so that we
2:15:28can now be bringing in 15 leads a week.
2:15:30And that's how we're able to from the
2:15:32beginning everyone is aligned on this
2:15:34metric. If we're able to hit this metric
2:15:36in the next couple weeks or months, does
2:15:39everyone consider this product a
2:15:40success? If the answer is yes, then it's
2:15:42a great thing to work on. But the
2:15:44problem that we've seen from when I've
2:15:45done consulting, when I've built stuff,
2:15:47is that if we don't align on the metric,
2:15:48a lot of people are confused about,
2:15:50okay, well, how do we know if we got ROI
2:15:52on this? You know, like what where's the
2:15:53benefit? I can't actually see it.
2:15:55Because productivity isn't as onetoone
2:15:58as for something like running ads, you
2:15:59know, like you pay an agency to run ads,
2:16:01they are going to spend $200,000 a month
2:16:04and you can directly see that those
2:16:05$200,000 brought in an extra million to
2:16:08the business. But when you're removing
2:16:09constraints and when you're doing things
2:16:11on the back end, you don't always see it
2:16:13super clearly about how it affects the
2:16:14bottom line. So every project that you
2:16:16work on before you start building it,
2:16:18before you agree to it, pick the north
2:16:20star and make sure everyone agrees,
2:16:22okay, this would affect the bottom line
2:16:24in some way and it would be a success if
2:16:25we could move the metric from five to 15
2:16:27leads coming in a week. So at a large
2:16:29scale to really impact the business,
2:16:31that is typically what I think about. I
2:16:32think about constraints and I just stick
2:16:34to that. Now, when you're just getting
2:16:35started, so like you're sitting there
2:16:36today and you want to this week start to
2:16:38automate some stuff. Maybe you don't yet
2:16:40have first of all like the confidence in
2:16:42that or like the luxury to be able to
2:16:44make those decisions or the budget for
2:16:46it. Right? So, what I want you to do
2:16:48then is start to write things down.
2:16:51Literally get out a piece of paper and I
2:16:52want you to write down kind of the
2:16:54following information. I want you to
2:16:55think about from week to week and like
2:16:57doing an audit of yourself. So, like
2:16:59what processes do you have or you know
2:17:02what tools am I using? What processes
2:17:05did I repeat? What are my triggers? And
2:17:09by triggers, I mean what things happen
2:17:12throughout the week that basically
2:17:14trigger you to do something else. A lead
2:17:16comes in, what do you have to do about
2:17:17it? A customer support ticket gets
2:17:19submitted, what do you have to do about
2:17:20it? And I think this one is powerful
2:17:22because this will help you identify like
2:17:24five to 10 processes. And then once you
2:17:26have those five to 10 processes that are
2:17:27pretty event- based or trigger based,
2:17:29you then think, okay, in my typical week
2:17:32or month, which one of these happens the
2:17:34most often? And then you just want to
2:17:35probably go for that one. Unless that
2:17:37thing is like super super high risk to
2:17:38the business where you have to be
2:17:40involved right now and you can't really
2:17:42change the way you're doing it, then
2:17:43maybe you want to bump down. But then
2:17:45you just basically have that whole list
2:17:46and you drill down on those. And then
2:17:48this other mindset shift that I called
2:17:50out here was that you can't automate
2:17:51what you can't map. So remember how we
2:17:53talked about the way that you actually
2:17:54build these skills or you build these
2:17:56automations is you have to basically
2:17:58define how they work and whether that
2:18:00means you do it first and then you
2:18:02actually execute or whether that means
2:18:03you execute and then you build a skill
2:18:05around it. If you don't clearly
2:18:06understand the process already then how
2:18:09in the world do you expect an intern to
2:18:11to understand it or an AI agent to
2:18:13understand it? You have to know the
2:18:14process well enough or you have to be
2:18:16able to talk to the stakeholders or the
2:18:17subject matter experts that do in order
2:18:19to automate it because the subject
2:18:21matter expertise that goes into the
2:18:22system is the most important thing. We
2:18:24talked about this, right? No matter how
2:18:26good the AI model is, you could have the
2:18:28best AI model in the world and the best
2:18:29harness in the world, but it's not going
2:18:31to be able to do anything meaningful
2:18:32with your business and your business
2:18:34data unless you feed in that theme, that
2:18:36subject matter expertise. You guys have
2:18:38probably heard some of these phrases
2:18:39like garbage in garbage out. That's
2:18:42that's a really popular one. or like
2:18:43context is king. And these are two
2:18:45pillars that I believe will be true
2:18:47forever. Even though with new tools
2:18:49coming out, this these two things super
2:18:52super important. And then there's one
2:18:53more that I like to reference which is
2:18:54Abraham Lincoln. If I had 6 hours to
2:18:57chop down a tree, I would spend the
2:18:58first four sharpening the axe. Meaning
2:19:00there's so much importance in the
2:19:02planning stage, mapping things out
2:19:04clearly, writing down processes clearly
2:19:06before you try to actually automate them
2:19:08and you know build them or execute on
2:19:10them. So the picking what to build is
2:19:12just a very mindset oriented thing. And
2:19:14the actual building itself is once again
2:19:16very mindset. It's it's all about the
2:19:17planning. It's all about the clear
2:19:18communication and it's all about sitting
2:19:20there watching the agent go steering it
2:19:22to make sure it's going in the right
2:19:24direction and then giving feedback and
2:19:25iterating and iterating and iterating.
2:19:27And we've talked about all of this stuff
2:19:28where you guys could now pretty
2:19:29confidently go do that. We talked about
2:19:31prompting. We've talked about
2:19:32understanding how to connect your tools.
2:19:34We talked about skills. And we've talked
2:19:35about context windows. And there's going
2:19:36to be more later on about managing this
2:19:38right here, token management. But you
2:19:40guys have basically all the skills that
2:19:42you now need to start to pick processes
2:19:44and go automate them. And all of the
2:19:45stuff that we're about to dive into is
2:19:47just getting a little bit more advanced
2:19:48and taking it really to the next level.
2:19:50So that is how you pick what to build.
2:19:52All right. So we've mentioned sub agents
2:19:54a few times. When we talked about the
2:19:56skills, I talked a little bit about how
2:19:57sometimes you can have skills call on
2:19:59certain sub aents and you can also have
2:20:01sub agents call on certain skills. And
2:20:03sub aents are really important because
2:20:05you know how we talk about the context
2:20:06window. What happens is when we use a
2:20:08sub agent we can delegate tokens to be
2:20:10spent in a different context window in a
2:20:13fresh sub agent. That way our main
2:20:15orchestration agent right here is able
2:20:17to just call on a bunch of them. So,
2:20:19it's pretty cool because instead of
2:20:20having one main session where we just
2:20:22fill up the context window, we can have
2:20:24our main session keep it clean because
2:20:26it's delegating work to little smaller
2:20:28sub aents that are each having their own
2:20:31context windows and we're able to just
2:20:32disperse it out a little bit more. The
2:20:34other cool thing is let's say this main
2:20:36agent is on Opus, so it's the most
2:20:37expensive. This main agent can delegate
2:20:40work to all of these little workers that
2:20:41are all like maybe on sonnet or maybe
2:20:43even on haiku. So you can delegate work
2:20:45that's a little bit less like high
2:20:48priority or high risk to cheaper models
2:20:50and you can just get some really cool
2:20:52results by doing stuff like that. So I'm
2:20:54going to go ahead and shoot you guys
2:20:55into a video that I made pretty much
2:20:57breaking down everything there is to
2:20:59know about Claude Code sub aents. Just a
2:21:01quick warning before this next video
2:21:03starts playing. Some of the clips that
2:21:04I'm inserting into this course were
2:21:06recorded a few months back. meaning they
2:21:08might be shown in VS Code extension or
2:21:11the terminal instead of the cloud
2:21:12desktop app that we've been using so
2:21:14far. I just wanted to give you guys a
2:21:15warning. Functionally, exact same. So,
2:21:18don't worry about it too much. It just
2:21:19might look a little bit differently, but
2:21:20all you have to do is listen to what I'm
2:21:22saying and follow along with what I'm
2:21:23actually doing and you will be just
2:21:24fine. All of this stuff is still
2:21:26relevant. Otherwise, I wouldn't be
2:21:27putting it in this course. So, hopefully
2:21:29that makes sense. See you guys in the
2:21:31video. So, I don't know what's going on
2:21:32up here, but I just told Cloud Code to
2:21:34spin up five different sub aents, and
2:21:35they all have different personalities.
2:21:37One is going to be a complete beginner,
2:21:38one will be a software engineer, one
2:21:40will be a business owner, one will be a
2:21:41publisher. And it comes back, and it
2:21:43says, "Okay, I'm kicking off all five
2:21:44now, each with a distinct persona and
2:21:46lens. These will run in parallel." You
2:21:48can see that this is now running four
2:21:50agents. The fifth one's about to spin
2:21:51up. And on the bottom, if I click into a
2:21:53different session, so we've got the main
2:21:55or we've got like the beginner, and I
2:21:56enter this conversation, we can actually
2:21:58see what's going on here. Meaning if I
2:22:00scroll up I can see the actual prompt
2:22:02that the main session kicked off to this
2:22:05sub aent. So here we have Linda 58 years
2:22:07old a retired elementary school teacher.
2:22:09You are a complete beginner to AI and
2:22:11then we see the actual task which is to
2:22:13read all the chapters and give a bit of
2:22:15a review. And so all of the other sub
2:22:17aents probably have a very similar
2:22:19prompt if I go to like the enterprise
2:22:20exec. So same exact prompt except for
2:22:22here you're role playinging as David 52
2:22:25a COO at a 12,000 person Fortune 500
2:22:27financial services company. So anyways,
2:22:29the point being what we can do is have
2:22:31our main session up here and the main
2:22:32session can delegate to as many
2:22:34different sub agents as we want and all
2:22:35the sub agents can have different chat
2:22:37models, different personas, different
2:22:38skills, different subject matter
2:22:40expertise. And if you watch my video
2:22:41where I ranked all of my favorite Cloud
2:22:43Code features, sub agents ranked number
2:22:45six. So today, what I'm going to do is
2:22:47I'm going to tell you guys exactly how
2:22:49to use them, what they are, when you
2:22:50need to use them, and how you can use
2:22:51them better than 99% of people using
2:22:53Cloud Code. So let's not waste any time
2:22:55and get straight into today's video.
2:22:56Okay, so what is a sub agent? You guys
2:22:58just saw a demo. We have a main chat.
2:23:00So, right here is where I said, "Hey,
2:23:01can you spin up five different sub
2:23:02aents?" And what it did is it right here
2:23:04kicked off five different ones. And then
2:23:05it comes back with an overall review.
2:23:07Apparently, I need to do some work on
2:23:08this book because I only got about an
2:23:10eight. More info on my book will be
2:23:12coming soon. But anyways, the main
2:23:13session is basically the orchestrator.
2:23:15It says, "Okay, cool. So, I am the one
2:23:17who's actually talking to Nate, but what
2:23:18I can do is I can spin up a bunch of sub
2:23:20agents that can only talk to me, and I
2:23:21can assign them work. Go read these
2:23:23files. Go do this research. Go fix that
2:23:26bug." and then you come back to me with
2:23:27a report of what you did and I'll
2:23:29communicate that back to Nate. So there
2:23:30are a ton of different reasons why these
2:23:32sub aents are useful and why they exist.
2:23:33So let's just start with this first one
2:23:34which is that it keeps your context
2:23:36clean. So let's say I'm in cloud code,
2:23:38right? And I'm just talking. Hello, how
2:23:39are you doing? What's going on? Let's
2:23:41build something, right? Like maybe we're
2:23:42doing research, maybe we're building an
2:23:43app, whatever it is. You start to fill
2:23:45up your context window, which you guys
2:23:47can see right here with my status line.
2:23:48You can see right now we're about 48,000
2:23:50tokens in 5% of the way up. And so as
2:23:53this starts to fill up, it starts to get
2:23:54polluted with information. But if you
2:23:57kick something off to a sub agent, as
2:23:58you guys saw earlier, it's a completely
2:24:00fresh chat. So just to show you guys
2:24:02another real quick visual demo, I'm in
2:24:04the desktop app, which is a little bit,
2:24:06you know, easier to see and it's
2:24:07visually more pleasing than the terminal
2:24:08sometimes, but let's say I said, "Hey
2:24:11Cloud Code, go ahead and kick off a sub
2:24:12agent to do some research for me about a
2:24:15product called Fireflies.ai." And so
2:24:18this is my main session. You know, I can
2:24:20talk to this thing. It'll help me do
2:24:21research on different tools. And then
2:24:22what happens is it kicks off a
2:24:24researcher agent to do the research. And
2:24:26what's cool is right now you'll notice
2:24:27I'm using Opus, right, which is
2:24:29obviously the most expensive model, but
2:24:30we can have a sub agent kick off and do
2:24:32research with Haiku or Sonnet. So we're
2:24:34getting this research for cheaper and
2:24:36we're getting a fresh context. So if I
2:24:37click on this agent here, you can see
2:24:39this is basically the prompt that the
2:24:40main agent sent over to this sub agent
2:24:43which was, hey, research Fireflyy's dead
2:24:44AI. Give us what it is, core features,
2:24:47how it works, pricing, give us all this
2:24:48stuff. And now this agent is the one
2:24:50over here searching the web and creating
2:24:52its opinions rather than our main
2:24:54session. So this helps preserve your
2:24:55main context in case you're ever doing a
2:24:57ton of research or reading a ton of
2:24:59stuff that you don't want to fill up
2:25:00your main context window, right? So
2:25:02that's one thing. There's also built-in
2:25:04sub agents which is the one we just saw,
2:25:05right? That was like basically a
2:25:07built-in cloud code research agent that
2:25:09will you've probably seen it get invoked
2:25:11automatically without you even asking it
2:25:12to be invoked. And then you've got
2:25:14custom sub agents that you're actually
2:25:15able to build yourself. And if you guys
2:25:18remember earlier in the demo when we
2:25:19spun up those different agents, I said,
2:25:21"Hey, one should be a software engineer,
2:25:22one should be a beginner, blah blah
2:25:23blah." You remember those all said
2:25:25general purpose. So those were still
2:25:27builtin
2:25:29native generic agents that just had a
2:25:31prompt. So that doesn't mean that we
2:25:33built those custom agents. That was just
2:25:34a general purpose agent that cloud code
2:25:36prompted differently. If we wanted to
2:25:38actually build a custom agent, that
2:25:39would be a markdown file. So if I open
2:25:41up my VS Code, you guys know in the
2:25:43cloud folder, we have different things.
2:25:45And the one you probably know the best
2:25:47is called skills. So in the skills
2:25:48folder, let's just take a look at real
2:25:50quick my agent builder skill. What this
2:25:52is is it's a markdown file. This lives
2:25:54as markdown so that I could send it to
2:25:56you guys. I could put it in my
2:25:57community. I could send it to my team.
2:25:58And all someone has to do is put this
2:26:00markdown file in there. Claude in a
2:26:02skills folder and then they're able to
2:26:04use it. And so a sub agent is the exact
2:26:06same actual tangible thing as a skill.md
2:26:10file. It's just called something else.
2:26:12You know, we've got the YAML front
2:26:13matter up here. And then we have the
2:26:14instructions of what the skill does and
2:26:16the actual steps to take. So if I open
2:26:18up my agents folder also in my
2:26:20do.claude. You can see I've got a
2:26:21different a couple different agents
2:26:22here. Right. So this one let's just look
2:26:24at is called the clickup-archer.md.
2:26:27And that's an agent that's called
2:26:28clickup searcher. We've got the yaml
2:26:30front matter up here name clickup
2:26:31searcher. We've got the description.
2:26:33We've got the model which I've defined
2:26:34here. We've got the color which means if
2:26:36I actually use the clickup searcher
2:26:38agent it shows the color. So actually
2:26:40let me just show you. can you go ahead
2:26:41and use the ClickUp searcher agent to
2:26:43show me what we've talked about today in
2:26:45the weekly commitments channel? And so
2:26:46what you'll notice is I invoked that
2:26:47with completely natural language. I'll
2:26:49have the ClickUp searcher agent pull
2:26:50today's messages and then right here I
2:26:52can see the green color. So that's all
2:26:53it means when you actually assign an
2:26:55agent a color. It's just so you can
2:26:57actually see it right there. And down
2:26:58here, you know, earlier right here is
2:27:00where it said general purpose. What it
2:27:02says now is ClickUp Searcher. And that's
2:27:04how we know that that's a custom agent
2:27:05that we built ourselves. So anyways,
2:27:07those are the two differences. And like
2:27:09I said, it's just one markdown file. And
2:27:12what this is called up here, the YAML
2:27:13front matter, that's called progressive
2:27:15disclosure. Which basically means if you
2:27:17say, "Hey, go do X, Y, and Z for me,"
2:27:20Cloud Code will naturally go search
2:27:22through your sub agents and your skills
2:27:24to see if you have any sub aents or
2:27:26skills to use. And so for the rest of
2:27:27this video, I'm just going to say sub
2:27:28agents, not skills, but they both work
2:27:29with this kind of progressive disclosure
2:27:32um process. But the idea is that cloud
2:27:34code is able to read just the front
2:27:36matter, just the name and the
2:27:37description and then decide does this
2:27:39apply to this prompt. If so, I'll pull
2:27:42in the sub agent and I'll run all of the
2:27:44extra stuff and read it. But otherwise,
2:27:45I'm not going to waste my tokens by
2:27:47reading everything if I'm not going to
2:27:49end up invoking that sub agent. So
2:27:51that's why we have this YAML front
2:27:53matter and that's why that's very
2:27:54important besides the fact that it also
2:27:55defines things like tools, model, and
2:27:57then there's tons of other levers you
2:27:58can pull there, but not going to dive
2:28:00into that right now. So anyways,
2:28:01settings up top and then your
2:28:02instructions go below that. And these
2:28:04are the four that I think matter the
2:28:05most. The name obviously so you can
2:28:07reference the sub aent. The description
2:28:09is really important. This is basically
2:28:10the trigger and this is how you can make
2:28:12sure that your sub aents are getting
2:28:13invoked without you actually saying,
2:28:15"Hey, go invoke this X Y and Z sub
2:28:18aent." So the more precise that your
2:28:20descriptions are, the more often cloud
2:28:22code will actually trigger them and you
2:28:24won't get misfires. Misfire is basically
2:28:26meaning you want it to invoke a sub
2:28:29agent but it doesn't or you don't want
2:28:31it to invoke a sub agent but it does.
2:28:33And so sometimes the only way that you
2:28:35can really make sure that you're you're
2:28:36tuning the actual front matter so that
2:28:37you're not getting this misfires is you
2:28:39just have to test it out and you just
2:28:40have to use it more and more and then
2:28:41like when it doesn't fire and you think
2:28:43it should you just think about okay why
2:28:45didn't that happen and then you update
2:28:46the description and then same thing if
2:28:49it's the opposite way around. And then
2:28:50if you go to the actual cloud code
2:28:51documentation on these sub aents, you
2:28:54can see all of the different things that
2:28:55you can actually put in the front
2:28:57matter. You can put tools like we just
2:28:58mentioned, but you can also put
2:28:59disallowed tools. So if you don't want
2:29:01it to ever write or edit files, you can
2:29:03put that so that these sub aents are
2:29:04explicitly read only. You can also
2:29:06define things like which MCP servers
2:29:08it's allowed to use. And you can even
2:29:10give it skills. So basically any setting
2:29:12that you want to configure for your
2:29:15custom sub aents, you can pretty much
2:29:17do. just come to the documentation, have
2:29:18cloud code read the documentation and
2:29:20say, "Hey, I want to set up a sub agent
2:29:21that does X, Y, and Z. It should not be
2:29:23able to do X, Y, and Z. It should be
2:29:24able to look at this data, not look at
2:29:26this data." And it will help you build
2:29:29the right YAML front matter. So, how do
2:29:30you actually write a great sub agent?
2:29:32So, obviously not having a weak
2:29:34description. So, having, you know, a
2:29:36very precise type of description. You
2:29:37can even say something like use
2:29:38proactively if you want it to fire off,
2:29:40you know, pretty generously. And then
2:29:43after you have the actual front matter
2:29:44dialed in, it's all about the body. The
2:29:46body is the way that the sub aent
2:29:48actually works, what skills it invokes,
2:29:50because yes, sub agents can invoke
2:29:51skills and skills can invoke sub agents.
2:29:54So, keep that in mind. They work
2:29:55together. They're not um you know,
2:29:57competitors. And you have to have that
2:29:58same idea once again that that you have
2:30:00to iterate. It's not going to be perfect
2:30:02on the first try. Every time you use
2:30:03your sub agent, you have the opportunity
2:30:05to give it feedback on what it didn't do
2:30:07well and how to make that better and
2:30:09then what it did really well and how to
2:30:10make sure that it does it every time.
2:30:12And real quick, what's the difference
2:30:13between a skill and a sub aent? Well,
2:30:14honestly, at their core, they're very
2:30:16similar because you're able to define,
2:30:18do X, Y, and Z in this order. You know,
2:30:20here's a prompt, here's a persona,
2:30:22whatever. But the main difference really
2:30:23is that one has a clean context window,
2:30:25and one doesn't. And you can run a ton
2:30:28of different sub agents in parallel in,
2:30:29you know, independent sessions, as we
2:30:31saw earlier, whereas the skill is
2:30:33typically more of something that I'm
2:30:34kind of triggering in my main session
2:30:36all the time. But once again, that
2:30:37doesn't mean that I don't have a great
2:30:38LinkedIn research skill that I hand off
2:30:40to sub agents to use. You know what I
2:30:42mean? So really I think of it as kind of
2:30:44like the parallel use and the clean
2:30:46context window and of course the ability
2:30:48to use a different model. Now there is
2:30:50something that I'm going to show you
2:30:50guys real quick in cloud code which is
2:30:52like it allows you to build agents very
2:30:54easily with a slash command. You can
2:30:56also do it with natural language but um
2:30:58I'll show you that in a sec. Before we
2:30:59show you that I did want to kind of go
2:31:01over this real quick which is project
2:31:03level versus global level sub aents. And
2:31:06this is the same, you know, if you
2:31:08understand how the cloud code like
2:31:09settings files work or the cloud code
2:31:11like hooks and skills, MCP servers even,
2:31:14it's all the same. You always have
2:31:15project level stuff or you have global
2:31:17level stuff. So project level stuff is
2:31:19basically what lives in your project in
2:31:21that repo. So right here we're in my
2:31:22Herk 2 project and anything that you see
2:31:24inside of mycloud right here is project
2:31:27level. So all of these agents are
2:31:28product project level. All of these
2:31:30skills are project level. And then I've
2:31:31got other sub aents and other skills
2:31:33that are global. So, for example, if I
2:31:35say, hey, where does my session handoff
2:31:38skill live? That's going to find that
2:31:41globally because if I go to my skills,
2:31:43there's no skill in here called session
2:31:44handoff, as you guys can see. But right
2:31:46here, the session handoff skill lives in
2:31:47your global skills directory at the, you
2:31:50know, the user level. And so, global
2:31:51ones are usable by every product on your
2:31:53machine. So, no matter which project or
2:31:55repo I'm working in, I can always use
2:31:57that session handoff skill or I can
2:31:58always use that, you know, sub agent
2:32:00that I've built and it belongs to me.
2:32:02So, if I share this GitHub repo to
2:32:03someone, they won't get that skill or
2:32:05they won't get that sub agent. It's not
2:32:06a big deal because you can easily say,
2:32:08"Oh, you know, you accidentally made
2:32:09that sub agent globally, but I actually
2:32:10wanted in this project, can you just
2:32:12move it?" And because it's just a
2:32:13markdown file, they move super easily.
2:32:15Or you can even have them both. You
2:32:16know, you can have them in both spots.
2:32:17But the reason why I wanted to explain
2:32:19that before I showed you this is because
2:32:21if you know, you have to choose. So, if
2:32:23you do a slash agents, you can look at
2:32:24what agents are currently running. If
2:32:26you've got a bunch of sub aents, you can
2:32:27go to your library and you can see a
2:32:29bunch of different built-in agents down
2:32:31here like claude, claude code guide,
2:32:33explore plan, and then you can also look
2:32:35at your project level agents. So, for
2:32:37example, we could look at, you know, the
2:32:39AI trend hunter. We can look at carousel
2:32:41planner. You'll notice that some of
2:32:42these have different models like all of
2:32:44these are sonnet, but then some of these
2:32:45have different project memory. You know,
2:32:46this one has project memory. This one
2:32:48has none. But anyways, what I wanted to
2:32:49show you guys if you go to create a new
2:32:51agent, you choose here if it's a
2:32:53personal or global or if it's a project.
2:32:55So let's just make a new project one
2:32:56right now. In order to create it, we can
2:32:58generate with claude or we can do manual
2:33:00configuration. So I would probably come
2:33:02in here and choose generate with claude.
2:33:03And then you basically just describe
2:33:04what this agent should do and when it
2:33:06should be used. And it says to be
2:33:07comprehensive for the best results.
2:33:09Create me a sub agent that criticizes
2:33:11all of my work. I basically want to be
2:33:13able to hand it off ideas and I want it
2:33:15to not agree with me, but I want it to
2:33:17um criticize it. I want it to roast it.
2:33:19I want it to play devil's advocate and
2:33:21look for every possible hole in the plan
2:33:23and what could go wrong and give me back
2:33:26basically that report. I want this thing
2:33:28to be invoked whenever I say roast my
2:33:30plan or review my plan. Anything like
2:33:32that. So that's my prompt. Obviously
2:33:34that's pretty concise. So like if you
2:33:36really had a good sub aent use case,
2:33:38you'd probably want to give it some more
2:33:39detail and some more nuance there. But I
2:33:41just want to show you how it's able to
2:33:42generate this file from the description.
2:33:44And because I chose project level, it's
2:33:46going to create that in the agents
2:33:48folder within my do.cloud. So in a sec
2:33:50here, we'll see an agent pop up. It'll
2:33:52probably be called like um devil's
2:33:54advocate.md or roast agent.mmd,
2:33:57something like that. Oh, but before
2:33:58that, it also says what tools do we want
2:34:00it to be able to use? So like for
2:34:02example, in this one, maybe I only want
2:34:04it to go with readonly tools. So I could
2:34:06say just readon. And then you know, we
2:34:09could look at some advanced options too,
2:34:11which would be all these MCP servers and
2:34:13a bunch of other things. and even like
2:34:15individual tools. Whoops. Even
2:34:16individual tools like bash, cron create,
2:34:18cronde delete. Like you can get really
2:34:20granular here. But in this case, I'm
2:34:22just going to go ahead and hit continue
2:34:23with readonly tools. And then we're able
2:34:25to choose the model. And in this case,
2:34:26we're going to go with haiku. But you
2:34:28can also inherit from the parent. So if
2:34:29the parents running on haiku, all sub
2:34:31aents that all of that sub aent will be
2:34:33inherited. Or same thing with opus or
2:34:35sonnet. And then finally, we can choose
2:34:36our background color. I'm just going to
2:34:38go ahead and choose pink. And then we
2:34:40get to choose the memory. So whether
2:34:41that's project, none, user, or local.
2:34:44And so really for the sub agents that
2:34:46I'd be creating and the ones that I
2:34:47would recommend you guys do, I'd
2:34:48probably just say project scope. Unless
2:34:50you want all these sub agents to be
2:34:51completely completely innocent, wake up
2:34:54completely blind, no memory at all, then
2:34:55you would choose none. But as far as
2:34:57between project user and local, I'm
2:34:59probably just going to always choose
2:35:00project. All right, there we go. So I'm
2:35:01going to go ahead and save this new
2:35:02agent. And you can see it just popped up
2:35:04right here. It's called the plan roster.
2:35:06And what happens when you create them
2:35:08with Claude is it makes this huge
2:35:11because it doesn't yet understand what
2:35:13you might say and how you want it to
2:35:14trigger. So my first recommendation
2:35:16would be trim this down a little bit
2:35:18because once again this is part of the
2:35:20progressive disclosure. So there's no
2:35:22need for the description to be so long.
2:35:24So I'm literally going to delete all of
2:35:26this. I mean it's good to look at but
2:35:27I'm gonna delete all of this up to here.
2:35:29And really in my case this is good
2:35:30enough, right? Use this agent when Nate
2:35:32wants an adversarial critique of an
2:35:34idea, plan, strategy, blah blah blah.
2:35:35Trigger on phrases like roast my plan or
2:35:37review my plan. We've got the tools, the
2:35:39model, the color, the memory, and the
2:35:41name. So now I'm just going to open up a
2:35:42new session of Claude. And um let's real
2:35:45quick just say so I've got this plan and
2:35:48I want to create an ice cream stand in,
2:35:51you know, uh Chicago. I want to create
2:35:53this ice cream stand on Oak Street Beach
2:35:56and I don't yet have a refrigerator and
2:35:58I want to sell the ice cream all day
2:36:01long for about, you know, 20 bucks a pop
2:36:04and it's just a little a little piece of
2:36:06ice cream. So, um, go ahead and roast my
2:36:08plan. This is actually interesting. So,
2:36:10I created a skill called roast and it's
2:36:12going to use that instead. So, it
2:36:13defaults to that because it thinks that
2:36:15it's good enough. And in the skill, the
2:36:18roast skill, I actually have it spinning
2:36:20up five different sub agents. So, that's
2:36:22a good demo. I didn't mean for this to
2:36:23happen. These are all general purpose
2:36:25sub aents that live within my row skill,
2:36:27but I'm going to go ahead and cancel
2:36:28that. I'm going to run this prompt
2:36:29again, but this time I'm going to
2:36:30explicitly tell it to not use a skill,
2:36:32but like I said, that's a good example
2:36:34of showing you that in a skill, you can
2:36:36have it fire off a bunch of sub aents.
2:36:37Anyways, the whole reason why the
2:36:38roasting thing is so top of mind is
2:36:40because cloud code and AI in general can
2:36:42be a little bit of a sickopant. It can
2:36:44just be a yes man. So, having things
2:36:46worked out like a roast skill or like a
2:36:48plan roaster sub agent is pretty
2:36:50helpful. Okay, so look what happened
2:36:51here. It did not invoke our roast, our
2:36:54plan roster sub aent. So what I'm going
2:36:56to say is go ahead and take a look
2:36:58within ourclaw agents folder. We've got
2:37:01a sub aent called plan roster.mmd and
2:37:04you didn't invoke it here and I'm not
2:37:06sure exactly why. Go ahead and read the
2:37:08description of that and and look back at
2:37:10my prompt and help me understand why did
2:37:12you not fire off the sub agent so we can
2:37:14make this better because that exact
2:37:15prompt is something where I'd want you
2:37:17to use that that sub agent. And so
2:37:18that's really the way that I think about
2:37:20um iterating on my descriptions both for
2:37:23skills and for sub agents. Just just
2:37:24understanding like why didn't it fire or
2:37:26why did it fire and how do we then
2:37:28rework the description. I do think
2:37:30there's a little bit of you know foul
2:37:32play here because my roast skill got
2:37:35invoked earlier and it's probably like
2:37:37defaulting to those skills before a sub
2:37:39agent. So, you know, maybe that wasn't
2:37:41the best example, but I guess it's good
2:37:42that it happened so I can show you guys
2:37:43the way that you might think about
2:37:46improving your YAML front matter. Okay,
2:37:49so completely my fault. I didn't close
2:37:52out the front matter. So, good tip. You
2:37:55have to close off the quotes if you open
2:37:58them up, right? That can break your
2:37:59JSON. It can break other things as well.
2:38:00So, it will break your YAML front
2:38:02matter. It said the problem wasn't
2:38:04judgment, it was mechanical. It also
2:38:07said, "Hey, you know, you do have a rose
2:38:08skill, so maybe there was a little bit
2:38:09of, you know, cloudiness there." So, I
2:38:11completely get that, but it went ahead
2:38:12and it updated the description. You can
2:38:14see it made it a little bit longer, but
2:38:16there's still collision between the
2:38:17roast skill and the plan roster, right?
2:38:19They both get invoked kind of similarly.
2:38:21So, really, the best thing to do here is
2:38:23you would combine the skill to say,
2:38:24"Hey, whenever you run the skill, you
2:38:26also invoke the plan roster agent
2:38:28instead." But for the sake of the demo,
2:38:30I am just going to actually be way more
2:38:32specific about what to use. So, there
2:38:34goes our prompt once again. The copy and
2:38:36pasting out of the terminal is horrible.
2:38:38So, usually if I want to copy and paste
2:38:40something from the terminal, I will tell
2:38:41it to write it to a text file or I will
2:38:43just use it in the desktop app. But
2:38:45either way, I was way more explicit
2:38:46here. You can see I said use the plan
2:38:48roster sub agent, not the roast skill.
2:38:50And now it's initializing our pink plan
2:38:52roster sub aent. And what I can do is I
2:38:54can go down to this section down here. I
2:38:57can open up this other terminal and we
2:38:59can see the exact prompt that got sent
2:39:00over to our plan roster, which was roast
2:39:03this business plan hard. Here it is in
2:39:04full. I want to create an ice cream
2:39:06stand in Chicago, blah blah blah. So,
2:39:07it's basically exactly what I said. Tear
2:39:08it apart, hit every flaw, the missing
2:39:10refrigerator, the absurd $20 price, and
2:39:13then the sub agent already finished up.
2:39:14So, it sent us back to the main session.
2:39:16And now the main terminal is going to
2:39:18interpret what the plan roster sub agent
2:39:20said and then give us the rundown. And
2:39:22what you'll notice here is the the plan
2:39:24roster took 22.8K tokens, but those
2:39:2622.8K tokens did not pollute our main
2:39:29session. All we got was basically this
2:39:31much, which is pretty awesome. So
2:39:33anyways, that's a real quick a little
2:39:34bit of a sloppy example, but hopefully
2:39:36it showed you guys the different
2:39:37elements to play with and you know the
2:39:39way that you think about using these sub
2:39:40aents, but that's what it looks like in
2:39:42cloud code. The way that I like to think
2:39:43about these is the same way that I've
2:39:45thought about AI since the beginning of
2:39:46my YouTube channel, which is your AI.
2:39:49You know, it's very fun and cool to have
2:39:50one mega personal assistant agent that
2:39:52can do everything, but really the best
2:39:53way to do it is to have each AI be a
2:39:55specialist. And that's where your main
2:39:57general ones can be pretty good at, you
2:39:59know, a jack of all trades because of
2:40:00skills, right? You invoke a skill and
2:40:02now it's good at LinkedIn post. Now it's
2:40:03good at doing research. Now it's good at
2:40:04scripting videos, whatever. But really,
2:40:06the sub agents are actual specialists.
2:40:08They have subject matter expertise. So
2:40:10you can have one that's a security
2:40:11auditor, you can have one that does
2:40:13tests, you can have one that writes
2:40:14docs, you can have one that's an expert
2:40:15with databasing, whether that's the
2:40:17architecture or the queries or anything
2:40:18like that. And you can just silo
2:40:21basically this assembly line or parallel
2:40:23work of a bunch of agents that are good
2:40:24at one thing and really really good at
2:40:26that one thing. And the other thing
2:40:27that's cool about that is you can borrow
2:40:30subject matter expertise from other
2:40:32people. This is just one of the hundreds
2:40:34of thousands of examples out there, but
2:40:35there's a GitHub repo which I'll link in
2:40:36the description. And this one's called
2:40:38awesome cloud code sub aents. So if you
2:40:40scroll down here, you can see there's a
2:40:41bunch of sub aents that you can use and
2:40:43in different categories, right? You've
2:40:44got an API designer, a back-end
2:40:46developer, a GraphQL architect. We've
2:40:49got other language specialists like
2:40:51TypeScript or SQL. You know, you can
2:40:53scroll through and find a lot of these
2:40:55custom same way that you look for skills
2:40:56from other people, custom sub aents that
2:40:58other people have already built, and
2:41:00they maybe know a lot more about CLI
2:41:02developing than you do. So, they've put
2:41:04all their subject matter expertise into
2:41:05a sub agent, and now you can just use
2:41:06that because all it is is a markdown
2:41:08file. Now, yes, because everything's
2:41:10open source and because all these
2:41:11markdown files are out there, you want
2:41:12to be careful, right? Like, if you're
2:41:14downloading a file or you're putting
2:41:15into your system, just make sure there's
2:41:16no prompt injections in there. Make sure
2:41:18there's nothing, you know, malicious.
2:41:20And you can even do it by having maybe a
2:41:22sub agent that verifies open source
2:41:24repos. And it's read only. It can never
2:41:25send data. It can never do anything. And
2:41:27all it does is verifies that there's
2:41:29nothing malicious inside of that
2:41:31markdown file. So anyways, we looked a
2:41:33little bit about how cloud picks out the
2:41:34agents. It can be automatic and it can
2:41:36automatically invoke things when you are
2:41:38like looking through your codebase or
2:41:40whether you are doing research. It'll
2:41:41automatically chuck some out there. You
2:41:43can also have them very proactively use
2:41:45sub aents if you have things like that
2:41:47in the description so it fires
2:41:48frequently. You [snorts] can also list
2:41:50them explicitly by name. You know, you
2:41:52can tag the agent name or you can say,
2:41:53"Hey, use the plan roster sub agent like
2:41:55I just did in that example. And you can
2:41:57also launch a session as a sub agent. If
2:42:00you do claude with a flag of the sub
2:42:02agents name, it'll actually put you
2:42:03right in a terminal right away with one
2:42:05of those sub aents." And honestly, I
2:42:06never do this, but it's nice to know
2:42:08that that feature exists. So, once
2:42:10again, just wanted to hit on the point
2:42:11that you can do readonly sub agents,
2:42:13which is pretty cool, just by using tool
2:42:15restrictions and giving them only
2:42:17certain things. It's always nice to have
2:42:18basically the mindset of if my AI could
2:42:21touch data or could read data, I have to
2:42:24assume that it will. Even if I never
2:42:25prompt it, I have to assume that it
2:42:27will. And that's the difference between
2:42:28a permission layer being explicit tools
2:42:30that it's allowed to use and explicit
2:42:32MCP servers it's allowed to use and just
2:42:35prompting and saying, "Hey, don't do
2:42:37that or you don't need to read that.
2:42:38Don't worry about it." There's a big
2:42:40difference between those types of
2:42:41permission layers. And then, of course,
2:42:43you have the ability to save a ton of
2:42:44money here. Let's say you have to read a
2:42:46300page research report and just get,
2:42:48you know, maybe three fun facts from it
2:42:50or just get a summary. There's probably
2:42:52no reason unless it's a really really,
2:42:54you know, technical report to use opus
2:42:56for that. Probably not even sonnet. So
2:42:58delegate that to a haiku sub agent to
2:43:00read everything and then send back just
2:43:02a small summary to your main session.
2:43:04And that's how you have the system where
2:43:05you have your smart boss, which is the
2:43:07opus model that you talk to on the
2:43:08day-to-day that just works with a bunch
2:43:10of little haiku agents. It's going to
2:43:12save you a lot of money. it's going to
2:43:13keep things moving faster and that's the
2:43:15way that you want to start utilizing
2:43:17these things. Another way that you can
2:43:18also keep them from getting out of
2:43:19control is you can have a max turns set
2:43:22on these sub agents. So maybe they're
2:43:23starting to do loops of research or
2:43:24they're doing loops of reviewing through
2:43:26a codebase. You can say, "Hey, max turns
2:43:28equals 10." Honestly, I don't use this
2:43:30very often because most of my sub aent
2:43:32delegation is research or very specific
2:43:34workflows where it doesn't really I'm
2:43:36not worried about a loop and I'm keeping
2:43:38my hands on either way. But that is once
2:43:40again just another nice lever to pull.
2:43:42So then after we've seen all these
2:43:44benefits, hopefully it's starting to
2:43:45become a little bit more clear, but a
2:43:47lot of people might also still wonder,
2:43:48okay, so when do you actually use a sub
2:43:50agent? When is it really better to? So
2:43:52one core question you can think about
2:43:53is, is this about to dump a pile of
2:43:55stuff into my chat that I'll never read
2:43:57again? If that's ever yes, delegate it
2:44:00to a sub agent. If it's no, then maybe
2:44:01you keep it in line. But there's also
2:44:03some other things to think about, too,
2:44:04right? So let's look at some signals. If
2:44:06you're about to read a lot of files, do
2:44:09some sub aents. If you're going to spit
2:44:11out a wall of output, maybe do sub
2:44:13agents. If it's a job that you keep
2:44:14repeating, build a custom sub aent for
2:44:16it. If it is independent stuff and you
2:44:18can run a ton of things in parallel,
2:44:20like you know, maybe you have 15
2:44:22chapters of a book and you want each
2:44:23chapter to be reviewed and like it
2:44:25doesn't have to be in chronological
2:44:26order. All of them can be reviewed at
2:44:28the same time. Then that that's parallel
2:44:29and then you can go ahead and do those
2:44:30independent jobs. And also if you want
2:44:32like an unbiased reviewer because once
2:44:34again sub agents can wake up no context
2:44:37completely fresh no memory and you can
2:44:39get an honest review. Now you don't need
2:44:41a sub agent if you're just doing a quick
2:44:43edit if the steps depend on each other
2:44:45right so if it's like 1 2 3 then four if
2:44:47the agents need to talk to each other
2:44:49then that's when you would need more of
2:44:50like an agent team or a different type
2:44:52of orchestration. I've made a video on
2:44:53agent teams before. Um they are more
2:44:55expensive than sub agents because
2:44:56they're they're talking and stuff like
2:44:57that but they share task list and
2:44:59everything. Sub aents do not work that
2:45:00way. It's just a onetoone relationship
2:45:02between sub agent and main session, not
2:45:04like a one to many. If you've got five
2:45:06sub agents running, they cannot talk to
2:45:07each other. You would also skip them if
2:45:09you need the sub aent to have like the
2:45:11context of the entire conversation or if
2:45:13it needs to ask you a question because
2:45:14you don't really get to talk to the sub
2:45:16agents. You know, the main agent is the
2:45:18orchestrator. Now, there's also
2:45:19something to think about which is a
2:45:20fairly newer feature. It was with with
2:45:21the release of Opus 4.8, which is the
2:45:24dynamic workflows. And what that does is
2:45:26it spins up a workflow that typically
2:45:29delegates to a ton of different sub
2:45:30agents in parallel. So remember the idea
2:45:32is that the main chat is the
2:45:33orchestrator and you've got a bunch of
2:45:34different sub aents running whether
2:45:35that's three or whether that's 40. A lot
2:45:37of times if you're asking for a big
2:45:39project and it decides to use a dynamic
2:45:41workflow, then all that's doing is it's
2:45:43creating a bunch of sub aents and it's
2:45:45delegating to them all at one time. So I
2:45:47made a video about those. I will tag
2:45:49that right up here if you want to check
2:45:50out the dynamic workflows video. You'll
2:45:52see an example I did where it spun up 41
2:45:54sub agents at the same time and just ran
2:45:56them. I've also done some examples, not
2:45:58on video, but like when I was testing it
2:45:59out, where I did some workflows and one
2:46:02of them spun up like 210 sub agents at
2:46:04the same time, which was great, but it
2:46:06ate through my context or sorry, it ate
2:46:08through my session limit like crazy. So,
2:46:10you definitely want to be careful when
2:46:11you're spinning up dynamic workflows.
2:46:13They actually then a few days after this
2:46:15came out, they said, "Hey, we changed
2:46:16the trigger word for dynamic workflows
2:46:18from workflow to to ultra code." You can
2:46:20still say to use a workflow for this,
2:46:21but when you're clearly referring to
2:46:23something else, Claude won't kick off a
2:46:24dynamic workflow. So, you want to make
2:46:26sure that you are being very careful
2:46:28about when you kick off those dynamic
2:46:29workflows because like I said, they are
2:46:31expensive. So, anyways, that is pretty
2:46:33much all of the stuff that I wanted to
2:46:36talk about here with sub agents. So, the
2:46:38whole thing on one slide, just to do a
2:46:39quick recap, if it's just one quick
2:46:41thing, you don't need a sub agent,
2:46:43right? Just because this feature is
2:46:44awesome, which it really is a great
2:46:45feature, that doesn't mean to force it.
2:46:47because sometimes if you're forcing too
2:46:49many sub agents, you're going to get
2:46:50worse results. So, play around with
2:46:51them, understand the benefits, and start
2:46:53to kick them off when you really do need
2:46:54them. If you want to share them with
2:46:55your team, keep them in your project,
2:46:57keep them in your repo. If you want to
2:46:58keep sub agents just for you that you
2:46:59can use across any project, then put
2:47:01them in your home folder. Kind of, you
2:47:03know, make them globally or personally.
2:47:05You can save a lot of money by having
2:47:06cheap workers with one smart lead. You
2:47:08can get better results by letting a
2:47:09fresh agent review your work or do work
2:47:11in parallel. If you want to do a giant
2:47:13parallel job, go ahead and check out a
2:47:15dynamic workflow. Just be careful of
2:47:16your session limit. And if you're not
2:47:18sure, if it's a pile of stuff that
2:47:20you're never going to reread, then go
2:47:21ahead and spin off a sub agent. Whether
2:47:23you are using cloud code in the terminal
2:47:25or whether you're using in the desktop
2:47:26app or even the VS Code extension in VS
2:47:29Code or, you know, on the web, wherever
2:47:30you're using cloud code, everywhere that
2:47:32you use cloud code can run sub aents.
2:47:34And the principles that I just talked
2:47:35about are always the same. This is where
2:47:38they live. That's how you invoke them.
2:47:39They're always YAML front matter. And
2:47:41those are pretty much the best
2:47:42practices. So, I know we covered a ton
2:47:45of information. If you guys want to
2:47:46download this exact slide deck, all you
2:47:48have to do is join my free school
2:47:49community. The link for that is down in
2:47:50the description. Once you join here, all
2:47:52you have to do is click on the
2:47:53classroom, click on all YouTube
2:47:54resources, and then you'll be able to
2:47:56find everything that I've dropped in
2:47:58here for free. GitHub repos, skills,
2:48:00templates, slide decks, whatever you
2:48:01want. It's all in there for free. All
2:48:03right, so we got sub agents crossed off
2:48:05the list. We're just going to keep
2:48:06making our way down through the rest of
2:48:08this course. So, next we have websites.
2:48:11And websites kind of go handinhand with
2:48:12GitHub because Cloud Code's really
2:48:15really good at building websites for us.
2:48:16It can build websites out of any coding
2:48:18language that you want. So HTML or you
2:48:20know CSS or all the other ones out
2:48:22there. It's really good at that because
2:48:24websites are code. But what happens is
2:48:26when we build code, it might give us
2:48:28something like a local host, which if
2:48:30you've never heard that before, don't
2:48:31worry. It's it's we'll break it down.
2:48:33But it's basically a URL that only you
2:48:35locally could open. If you tried to copy
2:48:37a local host URL and give it to your
2:48:39friend, nothing would pull up on their
2:48:41laptop if they tried to open that up.
2:48:42I've seen some funny tweets where it's
2:48:43like, "Hey, I'm a beginner. I just
2:48:45started using Claude Code. Check out
2:48:46what I built." And then they, you know,
2:48:48they attach a local host URL. And
2:48:50obviously that's like a meme. It's a
2:48:51joke, but it is pretty funny. So, just
2:48:53keep that in mind. We're building the
2:48:54website in code and then the code we
2:48:56push that to GitHub so that we can
2:48:58actually deploy that somewhere on the
2:48:59cloud. So, that is what I'm about to
2:49:01walk you guys through in this next
2:49:02video. Just a quick warning before this
2:49:04next video starts playing. Some of the
2:49:05clips that I'm inserting into this
2:49:07course were recorded a few months back,
2:49:09meaning they might be shown in VS Code
2:49:11extension or the terminal instead of the
2:49:14Cloud Desktop app that we've been using
2:49:15so far. I just wanted to give you guys a
2:49:17warning. Functionally, exact same. So,
2:49:19don't worry about it too much. It just
2:49:21might look a little bit differently, but
2:49:22all you have to do is listen to what I'm
2:49:23saying and follow along with what I'm
2:49:25actually doing and you will be just
2:49:26fine. All of this stuff is still
2:49:27relevant. Otherwise, I wouldn't be
2:49:29putting it in this course. So, hopefully
2:49:31that makes sense. See you guys in the
2:49:33video. Today I'm going to be showing you
2:49:34guys five simple hacks that you can use
2:49:36to make sure that Claude Code is
2:49:37building you websites that don't look
2:49:39like they were AI vibe coded, but they
2:49:41actually feel professional and branded.
2:49:43And we're going to be going through this
2:49:44in a way where even if you've never used
2:49:45Claude Code before, that's completely
2:49:47fine. You're going to be able to by the
2:49:48end of this video spin up some really
2:49:50awesome looking landing pages and
2:49:51websites. All right, so I don't want to
2:49:53waste any time at all. The first thing
2:49:54that you need to do is you need to go
2:49:55download Visual Studio Code. So go to a
2:49:57browser and type in VS Code and download
2:49:59this for your operating system. This is
2:50:01essentially just the IDE that we're
2:50:03going to be using Claude Code within. So
2:50:05once you've done that and you've opened
2:50:07it up, this is what it will look like.
2:50:08You're going to go to the lefth hand
2:50:09side right here and click on extensions
2:50:11and you're going to type in cloud code
2:50:13and install it like what you see right
2:50:15here. Now once you do that, it's going
2:50:16to prompt you to sign in with your
2:50:18anthropic subscription or your cloud
2:50:19subscription, which you do need a paid
2:50:21account. As you can see here, if you're
2:50:22on free, you don't have access to cloud
2:50:24code, but here on pro, you actually can
2:50:27use cloud code. Whether you're on pro or
2:50:29max, you can use it. I'd probably just
2:50:31start with pro. If you hit limits, which
2:50:32you probably will if you want to, you
2:50:34know, build websites all day, then you
2:50:36should probably upgrade to max. So once
2:50:38you've got that installed, you will see
2:50:40this little button up here, which is
2:50:41cloud code. And when you click on that,
2:50:43this is where it opens up the ability to
2:50:45actually use cloud code, talk to this
2:50:47little crab agent. And this is very
2:50:49similar to sort of like a chatbt or
2:50:50using cloud in the web. Now, the way
2:50:52that this works when you're using Cloud
2:50:54Code in Visual Studio Code or really
2:50:55wherever you use it is you have files on
2:50:57the lefth hand side and then you have
2:50:58your agent on the right hand side. So,
2:51:00first thing we need to do is open up a
2:51:02project so that we can start working
2:51:03with some files. So, I'm going to go up
2:51:05here to the top left and I'm going to
2:51:06click on explorer. What you can see is
2:51:08that it says you have not yet opened a
2:51:09folder. So, I'm going to go ahead and
2:51:11open up a fresh folder that has nothing
2:51:12in it. So, here we are in my website
2:51:15building YouTube folder, which like I
2:51:17said, it's a blank project. If you don't
2:51:19have a folder, just go ahead and create
2:51:20one. Whether that's in your desktop or
2:51:22your documents, just create one to start
2:51:24and then open that up. And that is where
2:51:26we will be working on this project. So,
2:51:28let's get started going through these
2:51:29five hacks. The first one is actually
2:51:31number zero. And the reason that I did
2:51:32this is because the first one is a
2:51:34claw.md file. And I put this as number
2:51:36zero because it's kind of a
2:51:37prerequisite, but also a lot of times
2:51:39near the end, even after 1 2 3 and four,
2:51:41you might have to rego back and update
2:51:44your claw. MD file or just have Claude
2:51:46do it itself. So what is a claw.md file?
2:51:48Just think of it as a system prompt.
2:51:50Think of it as every time before you ask
2:51:53cloud code to do something, it will read
2:51:56the claw.md file first. It will always
2:51:59process that. So what you want to do is
2:52:00make sure that that is pretty concise.
2:52:02You don't want to bloat it too much with
2:52:03context, but you want to give it the
2:52:05rules that it needs. So every time you
2:52:07are doing something in this project,
2:52:08this website building project, do this,
2:52:10this, and this. And always remember
2:52:12that's kind of the end goal. And so if
2:52:13you don't exactly know your full process
2:52:15yet or the end goal, then you might
2:52:17start without a claw.mmd file. But
2:52:18luckily for you guys, if you go over to
2:52:20my free school community, the link for
2:52:21that's down in the description. You go
2:52:23to the classroom, you go to claude code,
2:52:25and right here you will see the web
2:52:26designcloud.md file, which is the one
2:52:28we're going to be using today. You can
2:52:30go ahead and just download that for free
2:52:31right here. Now, once you've done that,
2:52:33you can actually just drag it right over
2:52:34here to the lefth hand side. Like I told
2:52:36you guys, the lefth hand side is where
2:52:37we can see our files and our folders.
2:52:39And what that does is it opens up the
2:52:40claw.md file which if I drag over here
2:52:42we can see it kind of full screen. Now
2:52:44the MD stands for markdown which is
2:52:47basically just this right here. We've
2:52:48got the pound signs. We've got um
2:52:50asterisk and it just helps keep the text
2:52:52organized so that the agent can read you
2:52:54know what's a header, what's a
2:52:55subheader, what's bold, what are bullet
2:52:57points, things like that. So you could
2:52:58obviously read through this entire
2:53:00claw.md file if you want to to kind of
2:53:02understand what we're telling it to do
2:53:03in this project. I'm not going to read
2:53:05everything because you guys can just,
2:53:06you know, look at it here or download
2:53:07it. And as we go through these other
2:53:09hacks, you will see why I put some of
2:53:10this stuff in here. But that actually
2:53:12brings me over to our first technically
2:53:14our first hack, which is the front-end
2:53:16design skill, which is why you can see
2:53:18right here in our cloud.MD, the first
2:53:20thing I wrote is always invoke the
2:53:22front-end design skill before writing
2:53:23any front-end code every session. No
2:53:26exceptions. So, first of all, real
2:53:27quick, what are skills? Well, if you go
2:53:29to the cloud code docs, you can read
2:53:30about skills right here. Essentially,
2:53:32they are custom instructions. So every
2:53:34time you build like a custom GBT or
2:53:36cloud project, you're usually putting in
2:53:37knowledge and you're putting in
2:53:38instructions. And basically skills are
2:53:41just that but in a markdown file. And
2:53:43why it's so important and cool is
2:53:45because every time you ask Claude a
2:53:47question, first it reads its cloud.MD
2:53:49file, but then it will think, okay, the
2:53:51user asks me this, do I have any skills
2:53:53in my library that help me do this
2:53:55better? If yes, I'll grab the skill,
2:53:57I'll read it, and then I'll take action.
2:53:58If no, I'll just use my general
2:54:00knowledge. So that's why we need to have
2:54:02the front-end skill because it helps us
2:54:04create designs that are way more modern
2:54:06and professional and they don't look as
2:54:08much vibecoded AI vibe coded. And the
2:54:11good news is it's super super simple.
2:54:13You just have to install it. So here's a
2:54:15tweet that showed the power of this. All
2:54:17they prompted Claude Code to do was use
2:54:19the front-end design skill, create a
2:54:20music player app, and it created this
2:54:22that has some, you know, animations. It
2:54:24has some dynamic elements. And if you
2:54:26would have just told Cloud Code to do
2:54:27this without that skill, it would have
2:54:28looked much worse. So, I'll leave a link
2:54:30to this tweet in the description of this
2:54:31video. You basically just have to run
2:54:32this command and then you run this one
2:54:34and then you should be good with the
2:54:36skill installed globally across any
2:54:38cloud code project that you might use in
2:54:39the future. And when I say run these
2:54:41commands, you can literally just copy
2:54:42this if you wanted to and just paste
2:54:44that right into here in cloud code and
2:54:46it would install that for you. All
2:54:47right, so let me go ahead and show you
2:54:49guys how good this front-end design
2:54:51skill really is with such a minimal
2:54:53prompt. So, before we prompt this agent,
2:54:55I just wanted to show you guys something
2:54:56else you can do, which is kind of a
2:54:57bonus hack. What I'm going to do is I'm
2:54:59going to create a new folder. I'm going
2:55:00to call this brand_assets.
2:55:03And our claw.mmd file actually explains
2:55:05that this might be a file or a folder
2:55:07that cloud code needs to look at. And
2:55:09what I'm going to put in here are two
2:55:12things. My logo and brand guidelines so
2:55:15that it creates this website and it
2:55:17feels very branded towards me and my
2:55:18business. So right here I'm dragging in
2:55:20the Amazon Society logo as you can see
2:55:22like that. And then I'm also going to
2:55:24drag in our brand guidelines which has
2:55:26stuff like our colors, our typography,
2:55:29icons, stuff like that. And so now that
2:55:31Claude can look at that, I'm going to
2:55:33just give it a very, very simple prompt.
2:55:35So all I'm saying is build me a modern
2:55:36and professional landing page for AI
2:55:38Automation Society. And I'm also going
2:55:40to tell it that here's my logo and
2:55:41here's my brand guidelines. It would be
2:55:43able to figure it out either way because
2:55:44we put it in the claw.md. But I just
2:55:46wanted to show you guys that you can
2:55:47actually tag assets directly. So, if I
2:55:50do an at, it will basically pop up and
2:55:52let me choose or point at the right
2:55:54things. So, now I can explicitly say,
2:55:56hey, here are the, you know, here's the
2:55:58brand guidelines and here's the logo
2:55:59because maybe they're not named in a way
2:56:01that's super intuitive. And now I'm just
2:56:04showing cloud code exactly what I want.
2:56:05So, I'm going to shoot this off. I'm not
2:56:07even in plan mode. I just want to show
2:56:08you guys how good this front-end design
2:56:10skill is. And what you're going to
2:56:11notice is first of all, what it did is
2:56:13it read the cloudmd file and now it's
2:56:16reading the brand assets. And now what
2:56:18it's going to do is it should hopefully
2:56:19invoke the front-end design skill and
2:56:21start building out that website for us.
2:56:23There we go. Right on Q. It has invoked
2:56:26the front-end design skill right there.
2:56:28All right. So, that has finished up. You
2:56:29can see that we've got a nav, a hero,
2:56:31tools, marquee. We've got stats, about
2:56:33benefits. So, a full onepage landing
2:56:35page. And it should be completely
2:56:36matching our brand as far as the logo,
2:56:38the colors, and the typography. It also
2:56:41added some animations. So, I'm excited
2:56:42to see how that works. And it threw it
2:56:44on local host for us to check out. So,
2:56:45let's head over there. All right. Look
2:56:46at that. We've got like a little
2:56:48animation up here. We've got a a line
2:56:49going down. We can see that we do have
2:56:51our logo up here as well as our exact
2:56:54colors and font. We've got a community
2:56:56rating. Ooh, that's super nice. We've
2:56:58also got some scrolling tech companies.
2:57:01So, we've got Editen, Make Claude,
2:57:03GBT40, Zapier, Air Table. We've got some
2:57:05random stats here. Obviously, we'd have
2:57:07to fill this in with our own copy, but
2:57:09keep in mind all of this happened with
2:57:11only us saying, "Create me a landing
2:57:13page for our community called A
2:57:14Automation Society." That was literally
2:57:16it and it created all of this. We've got
2:57:18testimonials. We've got a final call to
2:57:20action here. The logo is doing a little
2:57:21floating for basically a one sentence
2:57:23prompt. This is super super solid with
2:57:25the front-end design skill. Now, there
2:57:27was another secret thing going on here
2:57:28that I didn't yet tell you guys about,
2:57:29but if you've already read the Claw
2:57:31Denm, you might have noticed. And that
2:57:33brings us on to hack number two, which
2:57:35is the screenshot loop. So, the idea
2:57:38here is that AI is really good at
2:57:40getting you where you want to go, but it
2:57:42takes a lot of manual correction and
2:57:44steering. So, let's say I just told
2:57:46Claude Code to build us that website.
2:57:47Without the front-end design skill, it
2:57:49might have gotten us like 40% of the way
2:57:51there. But now that we added the
2:57:52front-end design skill, it's going to
2:57:53get us maybe let's let's just call it
2:57:5560. What we can do now is use
2:57:57screenshots to help AI iterate upon
2:57:59itself. So, instead of it getting 60% of
2:58:01the way there and then we make an
2:58:02improvement and then we make another
2:58:03improvement and we keep doing this, it
2:58:05basically should just bridge this gap
2:58:06itself because it's able to take a
2:58:08screenshot, look at the browser, see
2:58:10what it looks like, and then make make
2:58:11changes. So, what you guys didn't
2:58:13notice, or maybe you did, is over here,
2:58:15it created a new folder for us called
2:58:16temporary screenshots. And we can see
2:58:18that in that process of building out
2:58:20that first version of our workflow, it
2:58:22took 10 screenshots. So, I can click
2:58:24here, and I can see what it looked at.
2:58:25It looked at the hero section, which
2:58:27kind of was a a random full page. It got
2:58:30the viewport, which was that's more of
2:58:31the hero section. It looked at the
2:58:33stats. It looked at the about page. And
2:58:35what it did is it used these screenshots
2:58:36as it kept clicking through and looking
2:58:39and improved things. So, you guys didn't
2:58:41see this, but in the actual to-dos, it
2:58:44wrote the index html, it started the
2:58:46server and screenshotted the workflow,
2:58:48and then it did a two pass screenshot
2:58:49review and polish. So, it basically uses
2:58:51its eyes to check that what it's
2:58:53building actually looks good. And in
2:58:54order to set that up, it's actually
2:58:55really, really easy. If you go to the
2:58:57cloudMD file, you can see that I've got
2:58:58a section for screenshot workflow. And
2:59:00we're just doing this using Puppeteer.
2:59:02So, literally, if you take this claw.md
2:59:04and say, "Hey, Claude Code, can you set
2:59:05up Puppeteer to take screenshots?" it
2:59:08should be able to install all of that
2:59:09stuff for you right there really simply.
2:59:11And so, yes, that's cool on its own, but
2:59:13where it actually comes into handy a lot
2:59:15more is when we look at hack number
2:59:17three, which is using other websites as
2:59:20inspiration. Because what we're able to
2:59:22do is say, "Hey, Claude Code, take this
2:59:24website right here and build me a
2:59:26clone." So, you should build one that
2:59:27looks exactly like this one. And then
2:59:29what it's able to do is use its eyes,
2:59:30use its screenshot tool to screenshot
2:59:32what it's building and look at the
2:59:33reference and keep going back and forth
2:59:35until it's close enough. So, let me show
2:59:38you guys that in action right now. So,
2:59:40there's tons of sites that you could go
2:59:41to for [snorts] website inspiration.
2:59:43Here's one example called Dribble.
2:59:44Here's another example called godly
2:59:46website. And here's another really cool
2:59:48example called Awards with three W's.
2:59:50So, there's tons of places that you can
2:59:52find inspiration. So, for the sake of
2:59:54this video, I found this one that I want
2:59:56to use. It's got a nice little animation
2:59:58in the background. It's obviously not
2:59:59our color scheme, but it has some cool
3:00:01things as you scroll down like a
3:00:03dashboard. It's got some other little
3:00:04cards down here. None of this is really
3:00:06too animated. Well, I guess that is. But
3:00:08let's just say we wanted our website to
3:00:10look like this one for example. First
3:00:11thing that I would do is I would hit
3:00:12F12. I'm on Windows, [snorts] by the
3:00:14way. I would go to console and I would
3:00:17do control shiftp and search for
3:00:19screenshot. What this lets me do is
3:00:21capture a full-size screenshot of the
3:00:23entire page rather than just my current
3:00:25view.
3:00:26So here you can see it downloaded this
3:00:27screenshot and you can see that that is
3:00:29indeed the entire website. Now if you're
3:00:31on Mac that's still doable but you just
3:00:34have to Google the different buttons to
3:00:36do it. And then the next thing what I
3:00:37want to do is on the top right here I'm
3:00:39going to go to elements and in the style
3:00:41section down here I'm just going to copy
3:00:43everything. So I'm actually copying
3:00:45basically like the raw code or HTML or
3:00:48you know whatever you want to consider
3:00:49this as that tells the website how this
3:00:51is styled and we're going to give Claude
3:00:53code that. So, I'm going to go ahead and
3:00:55do a clear so we can start a fresh
3:00:56session. I'm going to first of all just
3:00:59paste in the code that we just copied,
3:01:00which is the style information. So, I
3:01:02said, I want you to spin up a new
3:01:03website for us. Get rid of the old one
3:01:05and you can put this one on local host.
3:01:07I basically want you to clone this
3:01:08website. So, I'm going to give you the
3:01:10screenshot, which what I'm going to do
3:01:11is just drag it in from my files and put
3:01:13it right over here. As you can see, that
3:01:14is the screenshot we just took. And I'm
3:01:16going to point to it so it knows what to
3:01:18use, which is the www right there. And
3:01:22then I said, here's the screenshot.
3:01:23here's the style and just go ahead and
3:01:24clone this website for us. So that is
3:01:27all we're going to do to start and then
3:01:29we can come back in later and tell it to
3:01:31use our branding and our you know colors
3:01:33and logo and everything like that. Now a
3:01:35couple things to keep in mind when
3:01:36you're doing some of the big processes
3:01:38like spinning up a website from scratch
3:01:40or comparing two websites and cloning
3:01:42them that coding process and thinking
3:01:44will take longer. But once you have a
3:01:46working version making small changes or
3:01:49tweaks that happens pretty quickly. And
3:01:51one other thing is you might have
3:01:52noticed that this really isn't stopping
3:01:54to ask me questions. And that's because
3:01:55I'm using bypass permissions mode. So if
3:01:58you don't see this in your instance,
3:01:59you're going to go to settings. You're
3:02:01going to type in clawed code. And then
3:02:03right here, you should see allow
3:02:05dangerously skip permissions. And that
3:02:06is where you turn that on. Now I
3:02:08definitely have a responsibility to tell
3:02:09you that this is dangerous. It has the
3:02:12potential to just kind of like run any
3:02:14command that it wants. But in my
3:02:16practice, I've never really had this be
3:02:18an issue, especially because I'm never
3:02:20like setting this to code all night long
3:02:22and then going to sleep. I'm always
3:02:23still kind of like watching it or I'm
3:02:25nearby and nothing bad really is going
3:02:28to happen. All right, awesome. So, we
3:02:30just got to the point where now it is
3:02:31creating a to-do list. And what you can
3:02:33see here is once it actually writes the
3:02:35code for the website, it's going to
3:02:37start up the server and it's going to
3:02:38take screenshots and it's going to do
3:02:40two rounds at least of comparing. It's
3:02:43going to look at what it built versus
3:02:44the reference. It's going to fix any
3:02:45mismatches and then it's going to do
3:02:46that again. And that is why the
3:02:48screenshot loop is so powerful. So
3:02:50logically, this is really cool. I mean,
3:02:51it's going through and it's looking
3:02:53section by section and analyzing how
3:02:55well it's stacking up. But we will have
3:02:56to see how it actually turns out. Okay,
3:02:58so that just finished up and before we
3:03:00actually see how good it really built
3:03:02this, I wanted to point out one thing
3:03:03about the screenshots. So you can see
3:03:05that we have screenshot 1 2 3 4, all
3:03:07this kind of stuff, but we don't really
3:03:09know which one is which without clicking
3:03:11on them. So, it looks like these are the
3:03:13clones as you can see because they're
3:03:14coming out looking like the website that
3:03:16we gave it. Well, we either should have
3:03:18before we started this new build. We
3:03:19should have told cloud code, hey, you
3:03:21can delete all of those temporary
3:03:22screenshots or in the claw.md, we should
3:03:25be more specific about the naming
3:03:27convention of the screenshots so that we
3:03:28can actually tell. Now, realistically,
3:03:30these temporary screenshots are more for
3:03:31Claude codes benefit than for ours, but
3:03:34that is something else that you can be
3:03:35thinking about if you do want to be able
3:03:36to click through and see the changes
3:03:38that were made with each version. But
3:03:39anyways, let's go ahead and open up this
3:03:40link and see what we got. All right, so
3:03:42I'm going to switch this to English for
3:03:44my head. But we can see we've got the
3:03:46purple colors. We've got your strategic
3:03:49ally for a truly profitable business.
3:03:50We've got the same top menu bar. Um a
3:03:52similar type of dashboard here. We've
3:03:54got some cards. And as we scroll down,
3:03:56it feels very similar to the real
3:03:58version that we gave it, which was this
3:04:00one. Obviously, some of the dynamic
3:04:02elements in the background and some of
3:04:04the actual images could not have been
3:04:05the exact same, but for a clone, this is
3:04:08very, very similar. And it is a really
3:04:11good spot for us to actually start. And
3:04:13now we can just start to integrate our
3:04:14own colors and logos and copy right into
3:04:17this template. And it's as simple as
3:04:20just asking it to do so. So, I'm going
3:04:21to go ahead and clear this out. I'm
3:04:24going to say go ahead and delete all of
3:04:26the temporary screenshots in the
3:04:27temporary screenshots folder. And so now
3:04:29all of those have been deleted as you
3:04:30can see. And we're basically going to
3:04:32say the most recent landing page looks
3:04:34really good. What I want you to do now
3:04:36is work in our brand assets. So our
3:04:40brand guidelines and our AIS logo. And
3:04:44this is for our community called AI
3:04:46Automation Society. So just work in
3:04:48those changes to that website clone that
3:04:50you just built. And once again, we are
3:04:52just going to stay on bypass
3:04:53permissions. I'm going to shoot that
3:04:54off. One shot prompt this thing. And
3:04:57hopefully we should get something that
3:04:58looks pretty solid. Now, what I'm
3:05:00interested to see is what it ends up
3:05:01doing with this dashboard and what it
3:05:03ends up doing with this iPhone screen
3:05:04because we haven't given it any other
3:05:06pictures. As you saw in our website, we
3:05:08obviously gave it some different
3:05:09pictures like the school games dashboard
3:05:11or me with Hermosi and Sam Ovens. But
3:05:13that's what you could do is you would
3:05:14come back into Claude Code and you would
3:05:16say, "Hey, I gave you some more pictures
3:05:17in the brand assets. Put this one here.
3:05:19Put this one there." And it would figure
3:05:21that out for you. And of course, you
3:05:22would also have to say, "Cool. When they
3:05:24click on start for free, take them to
3:05:25this link." or when they click on see
3:05:26the demo, take them to this link. So,
3:05:28there's other little pieces that you
3:05:29would obviously have to configure as
3:05:31well, but those changes take basically
3:05:33no time. Okay, so that finished up
3:05:35pretty quickly. We've got three
3:05:36screenshots here, but I'm not going to
3:05:38click into them because I don't want to
3:05:39ruin the final reveal here. But it used
3:05:41our colors. We have our primary accent,
3:05:43our secondary, our dark background, and
3:05:45our mid background. We've got the right
3:05:47typography. We've got the right logo,
3:05:48and everything was fully translated from
3:05:50French to English, thank goodness. And
3:05:52now it's rewritten for our community,
3:05:54which once again, we didn't actually
3:05:55give it facts about the community yet.
3:05:57This is just very simple prompting. It
3:05:59also mocked up a dashboard. So, let's
3:06:00head over to our local host. Let's give
3:06:02this a hard refresh. And boom. We now
3:06:04have our new site, master a automation,
3:06:07build faster, earn more. For the
3:06:09dashboard, it worked in like a little
3:06:10bit of a it's got members. It's got
3:06:11automations, courses. It's got it's kind
3:06:14of like a community tracker dashboard,
3:06:16and it uses our colors in there, too,
3:06:18which is cool. We've got different
3:06:19things on here, workshops, templates,
3:06:21expert community. It also changed this
3:06:23iPhone thing to member growth this
3:06:25month. So, it's keeping all of this on
3:06:27brand with the actual original reference
3:06:29site, which once again looked like this.
3:06:32However, now it has our colors and it
3:06:34has our information in here. We've got
3:06:36two paths and then we have some other
3:06:37stats down here and a nice little call
3:06:39to action at the bottom. So, cool. What
3:06:41we could do now is obviously go back and
3:06:42forth a little bit, maybe change some
3:06:44text, make things bigger, you know,
3:06:45change the images and stuff like that.
3:06:47But let's say we're at a spot where we
3:06:48like the overall feel and vibe of the
3:06:50website. But now, how do we really up it
3:06:52to the next level to make it feel
3:06:53unique? Well, what we're going to do is
3:06:55unlock the final hack, which is
3:06:58individual components. And what I mean
3:07:00by that is taking inspiration from
3:07:01different places, but for very
3:07:03individual components for small pieces,
3:07:06not entire websites. So, what we can do
3:07:08is we can go to a website called
3:07:0921st.dev,
3:07:11which has some of the best website
3:07:13components you might be able to find.
3:07:14It's got shaders. It's got backgrounds.
3:07:16It's got home screens. It's got buttons.
3:07:19It's got, you know, mouse highlights.
3:07:20It's got so many different things that
3:07:21you can do. So, here you can see I've
3:07:23got buttons and I could make them have a
3:07:24rainbow outline. I could make them
3:07:26shiny. We could toggle, you know, dark
3:07:27mode or light mode. There's lots of
3:07:29different things we could do here. Or I
3:07:30could just click on backgrounds in here
3:07:31and I could look at other ways that we
3:07:33could have our background. So, maybe we
3:07:34want these little kind of drop down
3:07:36pills instead. Or maybe we want these
3:07:38hero waves in the background. I think we
3:07:39should actually do this instead. So,
3:07:40what I'm going to do is just copy this
3:07:42prompt right here. This will basically
3:07:43copy a chunk of code for us to give to
3:07:46claude code. And I'm just going to say,
3:07:47I want you to work in this background
3:07:50element right behind the hero text. And
3:07:52after I give it that prompt, I just
3:07:54paste in what we grabbed from 21st.dev.
3:07:56And it should be able to use all of this
3:07:59and understand how to put that into our
3:08:02site. So, I'm just going to go ahead and
3:08:03shoot this off and we will see.
3:08:05Actually, one thing that I forgot to
3:08:06mention is in this case, because we're
3:08:09working with an animation, the
3:08:11screenshot might not always work the
3:08:12best. So, sometimes you might want to
3:08:14tell it not to do the screenshot flow.
3:08:16So, I'm basically actually just going to
3:08:17copy all of this text. Once again, I'm
3:08:20going to clear this out. I'm going to
3:08:21paste it back in. But then I'm also
3:08:24going to say
3:08:26because this is an animated background,
3:08:28do not use the screenshot tool to
3:08:30compare. just work in the code and then
3:08:32I will let you know if we need to make
3:08:34any changes. So hopefully with that
3:08:35mention, even though it's going to read
3:08:37the claw.mmd, it won't do a bunch of
3:08:38screenshots here because I've actually
3:08:40tested this out and I've had, you know,
3:08:42different background elements come
3:08:43through and because they're dynamic,
3:08:45sometimes the screenshot doesn't fully
3:08:46capture it. So it gets stuck in this
3:08:47loop of thinking, I haven't built this
3:08:49good enough. I'm going to keep trying
3:08:50and it like overengineers and it just
3:08:52doesn't really work. So sometimes you
3:08:54may want to turn off the screenshot
3:08:56tool. All right, so that just finished
3:08:57up. It didn't take a bunch of
3:08:58screenshots, so it didn't take forever.
3:09:00Let's go to the website. Let's give it a
3:09:03refresh and see. Okay. Okay. So, we've
3:09:06got a background. It looks a little bit
3:09:09um distracting. It also looks a little
3:09:10bit cheap. It looks like too pixelated.
3:09:12So, what I'm going to do now is just
3:09:13iterate. I'm going to tell it that
3:09:16I think that it's a little bit
3:09:18distracting as far as it makes the hero
3:09:21text right behind it a little bit tough
3:09:23to read. Also, in the hero text, I'd
3:09:26like it if the earn more was maybe a
3:09:30blue or a different color. I think that
3:09:31doesn't really feel good to have that be
3:09:33orange. It would be good if there was
3:09:34maybe some sort of background behind the
3:09:36hero text so that we could see it and it
3:09:38would still stand out and contrast
3:09:40against the background animation, but
3:09:43the background animation looks super
3:09:44fuzzy and super pixy. If you could make
3:09:46that look a little bit more professional
3:09:49and clean, that would be great. And if
3:09:51you guys were curious why I was just
3:09:52like staring at that and talking is
3:09:53because I was dictating and I wanted to
3:09:55be able to look at what I was talking
3:09:57about. So, we've given some feedback.
3:09:59Now, let's see if it can go ahead and
3:10:01make those changes. And once again, like
3:10:03we're being pretty vague here and it
3:10:05would be up to the creativity of the
3:10:07model to understand what we're asking
3:10:08for and be able to make these changes.
3:10:10Now, if you were on plan mode, it might
3:10:12be able to do a little bit better job of
3:10:13asking you some questions and maybe
3:10:15helping you get to a better solution
3:10:17first before it starts coding. But for
3:10:19the sake of the video, let's see how
3:10:21well it does with this prompt. All
3:10:23right, that just finished up and you can
3:10:24see that that looks much much better.
3:10:26This is definitely more what I was
3:10:28looking for when we copied over that
3:10:30animation into this website. So from
3:10:33here, we would just keep going through
3:10:34and we keep being really nitpicky about
3:10:35what we want to change. We'd add our own
3:10:37pictures in. We'd maybe want to change
3:10:38some of these buttons to be more
3:10:40dynamic. We'd want to maybe animate some
3:10:41of this other stuff, which we could
3:10:43easily do just by asking Claude Code to
3:10:45do so. So from here, the question is,
3:10:46how do you actually get this onto a real
3:10:49landing page? Because right now, we're
3:10:50still developing all of this code and
3:10:52we're previewing this in our local host.
3:10:54So what we're going to do is we're going
3:10:55to use a combination of GitHub and
3:10:56Versell to do this. Cloud code is where
3:10:58we're working right now. All of these
3:11:00folders, all of these files are local,
3:11:02meaning if I pulled up my laptop, I
3:11:03wouldn't be able to access them. And
3:11:05when we're building our website, which
3:11:06is obviously this website right here,
3:11:08this is all made up of a bunch of code
3:11:10in our cloud code project. So what we
3:11:12need to do with that is we sync that
3:11:14code to GitHub and GitHub has version
3:11:17control. We can see all of our commits,
3:11:19other people can work on it, stuff like
3:11:20that. We basically host our code or our
3:11:22project in the cloud and we set up a
3:11:24really cool auto deploy between Verscell
3:11:27and GitHub. And Verscell is basically
3:11:28just where we deploy our code to a live
3:11:31site. So basically what this means is
3:11:33whenever we tell cloud code, hey this
3:11:35looks good, push these changes to
3:11:37GitHub, GitHub grabs the new changes and
3:11:39then Verscell automatically grabs those
3:11:41from GitHub and then updates the real
3:11:44working version of our site. And I will
3:11:45show you guys that. But let's first of
3:11:46all do this pipeline. So the first thing
3:11:49that you're going to need to do is go to
3:11:51GitHub, create an account if you don't
3:11:52already have one, and you're going to
3:11:53need to create a new repository. So I'm
3:11:56going to create a repository right here
3:11:57called AIS test website. I'm not going
3:12:01to worry right now about a description
3:12:02or all of this and I'm just going to go
3:12:04ahead and create that repository. Now,
3:12:06what you also could do is you could tell
3:12:08Claude Code, hey, create me a GitHub
3:12:11repository and it could actually do
3:12:13that. But right now, I just wanted to
3:12:14show you guys so you can get a feel for
3:12:15GitHub if you've never used it before.
3:12:17So, anyways, now we have this repository
3:12:19called AIS test website. I'm just going
3:12:20to copy the name of that real quick and
3:12:22I'm going to come back into Cloud Code.
3:12:24We're going to clear this out and say
3:12:25awesome. So now that this site looks
3:12:27good, we need to actually deploy this on
3:12:29our domain. I need you to help push this
3:12:32to GitHub and we're going to push it to
3:12:34a GitHub repository called
3:12:37and then I'm going to paste in the name.
3:12:39Now, so far it has not yet gotten our
3:12:41GitHub credentials. So we're going to
3:12:42have to obviously authenticate into
3:12:44GitHub first so it can push that into
3:12:46GitHub. So I just got logged in as Nate
3:12:48Herk AI and now it's going to create
3:12:50the.get ignore and get everything set up
3:12:52so it can actually do so. Now, it's not
3:12:54too big of a deal right now because
3:12:55nothing that we'd be pushing into the
3:12:57public GitHub or, you know, onto the
3:12:59cloud has API keys or has any usernames
3:13:02or passwords or any sensitive
3:13:03information or, you know, web hook
3:13:05abilities. But that is something to be
3:13:06aware of once you actually are pushing
3:13:08automations and things like that to the
3:13:09cloud. Make sure that you're not putting
3:13:11any of your sensitive information out
3:13:13there. Awesome. So, it now says that our
3:13:15site is live on GitHub. So, if I click
3:13:17into this link, we should see that we
3:13:18now have a new commit. We have all of
3:13:20this stuff like our claw.mmd. We have
3:13:22our screenshot stuff. We have brand
3:13:23assets. And now we can sync this to
3:13:26Verscell. So that would be step two is
3:13:27you're going to go to versell.com,
3:13:29create an account. When you create that
3:13:30account, it's much easier if you just
3:13:31sign in or create that account with your
3:13:33GitHub credentials. And then all we have
3:13:36to do is go ahead and add a new project.
3:13:38And then we're able to just choose a
3:13:40GitHub repository. As simple as that. So
3:13:42I can literally just hit import on our
3:13:44AIS test website, which you guys just
3:13:45saw me set up. And then all I have to do
3:13:47is go ahead and deploy this project.
3:13:50Awesome. So I've deployed a new project
3:13:51to my project. I can go ahead and
3:13:53continue to the dashboard here. And what
3:13:55this now does is we can actually visit
3:13:57this by going to
3:13:58ais-test-website.vercell.app.
3:14:01I open that up. And now this is no
3:14:03longer local. I could open up my phone
3:14:05and type in this. You could open up your
3:14:06browser and type this in. And you guys
3:14:08could all visit this site because it's
3:14:09now deployed on the cloud. But of course
3:14:12it's got an ugly domain. So, what you
3:14:14would have to do now is you would have
3:14:15to go to your project settings. You
3:14:18would go to domains. And then this is
3:14:19where you would actually just have to
3:14:20either buy a domain right here or add an
3:14:23existing one. And it's really simple. It
3:14:24would walk you through the DNS
3:14:25configuration that you need to set up.
3:14:27And it's not too difficult, but I'm not
3:14:29going to actually do that live in this
3:14:30video. So, what I wanted to show you
3:14:31guys real quick before we end off this
3:14:33video is what actually happens if we
3:14:35realize that we want to make a change to
3:14:36our website that is on the cloud. Well,
3:14:38that's why it's good that we still have,
3:14:40you guys can't see because you can't see
3:14:42the URL, but we still do have our local
3:14:44version because if I make a change here
3:14:46and I don't like it, I don't want that
3:14:47to automatically get pushed to um
3:14:49Verscell. So, what you'd probably want
3:14:51to do is in your claw.md file, you would
3:14:53say ultimately what's going to happen is
3:14:55we're syncing all of the changes to
3:14:56GitHub. GitHub's going to automatically
3:14:58push them to Verscell and we'll be good
3:14:59to go. But when I'm making changes with
3:15:01you here, we're always going to test on
3:15:03a local host until I tell you explicitly
3:15:05to push that to GitHub or commit those
3:15:07changes to GitHub. Okay. So, this is our
3:15:10local version. And let's just say, for
3:15:11example, we wanted to make this button a
3:15:13little bit cooler. So, I'm going to ask
3:15:15in Cloud Code, could you go ahead and
3:15:18make the join the community button in
3:15:20the main hero text section, make it more
3:15:23professional. So, give it like a cool
3:15:24glow. And once you've made this change,
3:15:26let me see it in local host. Don't push
3:15:29it to GitHub until I tell you to. This
3:15:31thing is getting pretty screenshot
3:15:32happy. I may have to adjust the wording
3:15:34in the cloud. Mmd file a little bit. It
3:15:36literally took one of the main screen
3:15:37and then it took one of where it just
3:15:39cropped the actual button, but hey, it
3:15:41looks good. Okay, so what happens is
3:15:43here's the local host. I'll refresh
3:15:44that. Now we can see the little glow
3:15:46behind the join the community button and
3:15:48here is the web app version. I refresh
3:15:50this and we don't have that change yet,
3:15:52which is great because we don't want to
3:15:53push changes if they're not good, right?
3:15:55But now what I'll do is say awesome. I
3:15:57love that change. Go ahead and push that
3:15:59to GitHub. All right, so it just pushed
3:16:00that. We have a new commit. If I go to
3:16:02GitHub and I give this a refresh, we can
3:16:04see that we should see right here two
3:16:05commits. This one was add glowing pulse
3:16:08effect to hero join the community
3:16:10button. And then if I go to Verscell and
3:16:13we go to our deployments, we should see
3:16:15that we just got a second one come
3:16:16through as well just now. And now if I
3:16:18go to the site on the web and I refresh,
3:16:21we see the actual glowing join the
3:16:23community button. All right, so those
3:16:24are the five hacks that I wanted to
3:16:25cover today. We have our claw.md file,
3:16:27which as you could tell by this video,
3:16:29yes, it's nice to have something to
3:16:30start, but you are going to continue to
3:16:32iterate upon it throughout your project
3:16:33until you get to a good spot. We've got
3:16:35the front-end design skill, which is
3:16:36just like way too easy to not use. We've
3:16:38got the screenshot loop, which you got
3:16:40to be careful about, but it is very
3:16:41helpful. We've got inspiration websites,
3:16:43and then we have inspiration individual
3:16:45components, and now it's just a matter
3:16:47of making small tweaks and iterating
3:16:49upon your website. Awesome. So now that
3:16:52you guys understand how building the
3:16:54websites work and how cloud code is able
3:16:56to push to GitHub and push to versel to
3:16:57actually deploy them, you can continue
3:16:59going down some other like design stuff
3:17:01if you want. And if you do, you should
3:17:02try out claude's tool called claw
3:17:04design. So it's literally just a
3:17:06different web-based app that is
3:17:07basically specifically only designed for
3:17:10design. So if you want to check that
3:17:11out, I did drop a full course on cloud
3:17:13design which I will tag right up here.
3:17:14But anyways, now that we have finished
3:17:17talking about websites and GitHub, we
3:17:19are going to move on to talking about
3:17:21trusting the output. This is a really
3:17:23important topic and it's pretty nuanced
3:17:26because the the trust factor differs
3:17:29based on the type of automation that
3:17:30you're actually building. So, let me
3:17:32explain what I mean by that. So,
3:17:33deterministic versus non-deterministic.
3:17:35Deterministic means essentially
3:17:37predictable input and then we know
3:17:39what's going to happen and we know what
3:17:40we're going to get out. So, think a
3:17:42vending machine. we know that we have,
3:17:44you know, A1, E4. We can click these
3:17:47buttons and when we click the button, we
3:17:48know exactly what's going to happen and
3:17:49exactly what we're going to get. Now,
3:17:51nondeterministic is basically just the
3:17:53exact output. We don't exactly know what
3:17:55the trigger is going to look like. We
3:17:56don't exactly know what's going to
3:17:57happen inside the process, and we don't
3:17:59exactly know what we're going to get on
3:18:01the other side. We obviously have things
3:18:02that we're steering the system towards
3:18:04and we're trying to get one type of
3:18:06output. But because AI is essentially a
3:18:09black box and it's kind of like a slot
3:18:11machine, AI automations are more
3:18:13non-deterministic and they're designed
3:18:15to be. That's good about them. You know,
3:18:17they're flexible and that's what unlocks
3:18:19so much extra opportunities in this
3:18:21world is that now we have intelligence
3:18:22inside these automations. But with that
3:18:24non-determinism comes a lot of other
3:18:26things to think about. Think about a
3:18:28quick example of like responding to
3:18:29emails or writing cold emails. With the
3:18:32deterministic side, what we would do is
3:18:34we would have variables. So variables
3:18:35would fit into some sort of templated
3:18:38email copy and the variables would say,
3:18:40okay, the person's name goes here, their
3:18:42company name goes there and their, you
3:18:44know, MR goes there or something like
3:18:46that and it was very static. But now on
3:18:48the nondeterministic side, the AI can
3:18:50just look at all this input and then
3:18:51write a more personalized email. So it's
3:18:53not a template anymore. It's more of an
3:18:56actual, you know, generative AI. It's
3:18:57generated content. Now, from there
3:18:59though, I think about this as an AI
3:19:02systems pyramid. On the bottom, we have
3:19:04chat bots and at the top we have agents.
3:19:06So, let me explain kind of what I mean
3:19:07by this and why I like to think about it
3:19:09like this. So, really the core message
3:19:12that I think about and that Enthropic
3:19:14talks about and that other, you know,
3:19:15tech labs and leaders in the space talk
3:19:16about is you want to design the simplest
3:19:20solution for the job. Which means if the
3:19:22process doesn't need AI, there's no
3:19:24reason you should put AI in there
3:19:26because as you move up this pyramid,
3:19:27which is basically, you know, you're
3:19:29moving up in, let me actually change the
3:19:31arrow type. You're basically scaling up
3:19:33in autonomy as you move up, which also
3:19:36means you're scaling up in basically
3:19:37like unpredictability. And as you're
3:19:40scaling up these two things, you're also
3:19:41scaling up a few other things like cost
3:19:44and risk, just to name a few. So as you
3:19:47move up, you have things to think about
3:19:49as well. So, a chatbot, what do I mean
3:19:51by this? I think about a chatbot as
3:19:53basically something where a human has to
3:19:55trigger it. So, we're in full control of
3:19:57when we want the chatbot to run. I even
3:19:59would think of this as things like your
3:20:01claw skills that we're building, right?
3:20:02Because typically what's happening is
3:20:04you are invoking the skills by yourself,
3:20:06which means you're able to sit there,
3:20:07watch it, you're still kind of driving,
3:20:09you're still in control, you take the
3:20:10output, and now you're doing something
3:20:12with it. That's not really a high-risk
3:20:14environment because it's not like it's
3:20:15going to shoot off thousands of requests
3:20:17and do a thousand things without you
3:20:19asking unless you know that's what the
3:20:21skill entails on the inside. But think
3:20:23about this more like something where you
3:20:25actually trigger it as the human. And
3:20:27then what we have are workflows. These
3:20:28are deterministic workflows that are
3:20:30nothing new. It's basically just
3:20:31automation. Maybe like moving data from
3:20:34one spreadsheet to a database or
3:20:36something like that that takes no AI at
3:20:38all, but it's triggered by an event or
3:20:40on a schedule. So this thing runs while
3:20:43you sleep and you're not really there.
3:20:44It's not as much human in the loop. So
3:20:46it really starts to scale up here, but
3:20:48still not super dangerous because this
3:20:49is still like basically 100%
3:20:51deterministic. You know the input and
3:20:53you know the output. Then you get up to
3:20:55AI workflows where basically the order
3:20:57of events is still the same. It still
3:20:59goes 1 2 3 4 5 and then somewhere in
3:21:03that process there's an AI step. So let
3:21:05me try to make this a little more
3:21:06practical with an example. Let's take
3:21:08first of all let's just do the customer
3:21:10support example. responding to emails.
3:21:12So with a chatbot or something like
3:21:14that, what happens is an email comes in
3:21:17and then the human would trigger the
3:21:20skill. So the email comes in, the human
3:21:22triggers the skill, the skill looks up
3:21:24the database, it you know generates the
3:21:26email and then the human would go off
3:21:27and send the email back to the person.
3:21:30So that is very like you know we are in
3:21:31control the whole way. Now what happens
3:21:33with a workflow? Okay, so the email
3:21:35would come in that would maybe trigger a
3:21:38like data lookup tool and that could be
3:21:39a lookup based on the email account and
3:21:42then maybe that could just trigger a
3:21:43notification for a human to be able to
3:21:45actually go look at the ticket and
3:21:46process the email. So that is like a
3:21:49workflow because it's basically just
3:21:50moving data from one side to the other
3:21:52from left to right and there's no AI.
3:21:54Now let's take a look at what an AI
3:21:56workflow here could look like. Let's say
3:21:57an email comes in. We then do a data
3:21:59lookup so we can collect more
3:22:00information about the user. But then
3:22:02what happens is instead of notifying a
3:22:04human, what we want to do here with the
3:22:05AI is maybe we generate an email. And
3:22:07then this is where the non-determinism
3:22:09kicks in because we don't know what the
3:22:10email is going to look like. But over
3:22:11here, this is still like one 1 2 3. It's
3:22:15still like a workflow in nature because
3:22:16we know the steps that are going to
3:22:18happen and in what order. But then when
3:22:20we get to an agent, it's much different
3:22:22because what happens with an agent is we
3:22:24don't know what's going to happen. We
3:22:25would basically have an AI agent that
3:22:27has a bunch of different tools. And
3:22:28obviously I'm just making up what these
3:22:30tools would be, but we don't know if the
3:22:31email is going to come in and then the
3:22:33agent will go, "Okay, I'm going to use
3:22:34none of these tools." Or maybe I'm going
3:22:35to use the data tool twice and then the
3:22:37ticket tool. Or maybe I'm going to use
3:22:38the ticket tool, then the data tool,
3:22:40then the ticket tool again. Because that
3:22:41loop of reasoning and decision is all
3:22:44inside of this AI step right here. And
3:22:46that is where we lose the element of a
3:22:48workflow. We don't know the steps, but
3:22:50we do know that this is like the fully,
3:22:52[snorts] you know, most autonomous
3:22:53situation that we could possibly have.
3:22:55And we know that on the end result, what
3:22:57we want is for the agent to respond to
3:23:00the user with an accurate email. But
3:23:02that's why we have to be able to build
3:23:03the right systems to make sure that we
3:23:04actually trust that output. So hopefully
3:23:06that paints a little bit of a picture
3:23:07about the different ways you can
3:23:08orchestrate these systems and what
3:23:10you're kind of sacrificing or what you
3:23:12have to be aware of as you move up this
3:23:14pyramid of the AI systems. And all of
3:23:17this stuff will come into play later
3:23:18when we start to talk more about like
3:23:20routines and deploying automations and
3:23:22things like that. Just keep in mind you
3:23:23want to build the simplest solution
3:23:25possible for what you're trying to do.
3:23:27And there's one more thing I wanted to
3:23:29talk about here when it comes to
3:23:31actually trusting the outputs. And
3:23:33that's what a permission layer looks
3:23:34like because we've got prompts and then
3:23:36we've got like an actual permission
3:23:38layer like a tool level permission
3:23:39layer. So I should maybe move this up
3:23:41here and then just say here like tool
3:23:43layer. So anyways, let's say this is a
3:23:46visualization of like an agent we're
3:23:48building and we've prompted it to do
3:23:49things like never send an email only
3:23:51draft or never delete our database or
3:23:54never delete emails. Well, the prompt
3:23:56layer is one kind of guardrail, but it's
3:23:59not solid enough. The agent could follow
3:24:01instructions 100 times, but on the 101st
3:24:03time, maybe it forgets to or maybe it
3:24:05just gets a little bit clouded or
3:24:06confused and it ignores that prompt. So
3:24:08the point I'm trying to make here is
3:24:10let's say you said, "Hey, don't ever
3:24:12send an email." A lot of times the email
3:24:14boundary will stay within the prompt
3:24:15layer, but every once in a while it'll
3:24:17sneak out of the prompt layer because it
3:24:20has a tool to do so. So if you take away
3:24:22the tool to send an email, then there's
3:24:24no way that the agent could actually get
3:24:26through the tool layer to hit that email
3:24:28tool. So that's the idea. If you want
3:24:29the agent to draft emails but not send,
3:24:31then don't put a send email tool inside
3:24:34of its tool layer. Keep that on the
3:24:35outside. But then it's perfectly fine to
3:24:37have the draft email tool right here and
3:24:39say you can draft emails, but just don't
3:24:41ever send them. Similarly with delete
3:24:43functions, right? Like why would you
3:24:45ever allow this to be inside of the tool
3:24:47permission layer, even if your prompt
3:24:49says never ever ever ever delete
3:24:52anything, then just move the delete tool
3:24:54out of there so that it physically
3:24:56couldn't, even if it tried to, if it
3:24:57wanted to go rogue, I don't know,
3:24:59whatever the situation is. Don't give it
3:25:01tools if it shouldn't be able to use
3:25:03those tools. And I know that sounds
3:25:04obvious, but it happens way more often
3:25:06than you'd think. You've seen maybe
3:25:08stories on X or LinkedIn where massive
3:25:11companies had databases deleted because
3:25:12an agent went rogue or we actually
3:25:14internally had a situation where an
3:25:16agent sent out like 100,000 emails or
3:25:18150,000 emails to our list with a
3:25:20discount code and we didn't want it to
3:25:21do that. But what we found is that
3:25:23because it had access to the tool even
3:25:25though it was never told to send that
3:25:26email, it just did. So you have to
3:25:29assume that if an agent can read or
3:25:31touch or use something, assume that it
3:25:33will assume worst case scenario. Let me
3:25:35show you guys another quick example that
3:25:36I think will really resonate because
3:25:38when I first started building AI agents
3:25:40for people or AI automations, this was
3:25:42like the most common request was like an
3:25:44inbox triaging responding agent. So like
3:25:46an an email agent really. So a lot of
3:25:49people wanted one that looked like this
3:25:51and I built a lot that looked like this.
3:25:53We would basically have the trigger
3:25:55being a Slack message. The agent would
3:25:57have, you know, a little bit of a memory
3:25:59as well as a system prompt. And the
3:26:01system prompt would be massive. It would
3:26:02be a bunch of rules about what do you do
3:26:05with the emails, how do you label them,
3:26:06how do you respond, how do you do this,
3:26:08what is important, what's not important,
3:26:10and then after you respond to the to the
3:26:12actual email, you would send a message
3:26:14to the user and say, "Hey, here's what I
3:26:15did. There's the draft or I sent this or
3:26:17whatever." So, you'd have a agent would
3:26:19have a bunch of different tools, a bunch
3:26:20of email tools like draft an email,
3:26:22label, mark them as read, mark them as
3:26:24unread, um, pull them back, like
3:26:26actually get them and be able to read
3:26:27them and search through old ones. And
3:26:29there was probably more. But the point
3:26:30I'm trying to make here is I built a lot
3:26:32of agents like this and they weren't
3:26:34super reliable because of the fact that
3:26:36there was they were agents and there was
3:26:38so much reasoning. There were so many
3:26:40decisions. There were so many logic
3:26:42rules baked into the prompt. And as we
3:26:45know, a prompt isn't exactly a hard
3:26:47layer. Prompting is almost just more
3:26:48like a suggestion. So I then started to
3:26:52design all of my inbox agents like this.
3:26:55Now, this looks a little bit scarier,
3:26:57but it's so much simpler because all of
3:26:59this is a workflow. So, I moved this
3:27:01down from an AI agent at the top of the
3:27:04pyramid to AI workflows where now it is
3:27:07so so simple because what happens is the
3:27:09trigger would come in and then we're
3:27:10just using routing rules. Like for
3:27:12example, right here, if the contact is
3:27:15in um Google contacts, if yes, then we
3:27:18will do nothing. And if no, then we will
3:27:20extract the information using AI. So
3:27:22we'll pull in their name, their email,
3:27:23phone number, information about them,
3:27:25and we have to use AI for that. And then
3:27:27we would create a new record in the
3:27:28Google contacts. Now, every single one
3:27:31of these nodes is just one step. And the
3:27:33workflow is always going to follow the
3:27:35chain. So after we would basically
3:27:37figure out are they a contact or not.
3:27:39Then we would look to see how we respond
3:27:41to the email. We had three routing rules
3:27:44here. If the email [email protected],
3:27:47then we would label it over here as
3:27:49internal. If the email domain was
3:27:52[email protected],
3:27:54then we would label as bill or sorry, we
3:27:57would label it as billing. We would then
3:27:59summarize the email and then we'd send a
3:28:00summary to the right person in Slack
3:28:02that hey, we just got this new billing
3:28:04inquiry. And if it was a VIP email, then
3:28:08we would basically just label it as VIP.
3:28:10And so this is a very very simple
3:28:12example of different routing rules. What
3:28:14what you'll notice is that because the
3:28:17decisions are basically being made by
3:28:19legitimately objective facts rather than
3:28:23an AI prompt, this type of system had
3:28:27basically no failures. The only failures
3:28:29would be if it extracted information
3:28:31wrong or if it summarized an email
3:28:33weird, right? Because all of these blue
3:28:36steps are objective. They were simple
3:28:39logic. They were easy. The green where
3:28:42you have decisions and where you have
3:28:43generative AI is where you get the
3:28:45variability. They could perform the
3:28:47exact same these two systems. Like
3:28:48essentially they're doing the same thing
3:28:49in theory. This one on the right was so
3:28:52much more consistent. It was also
3:28:55cheaper and it was just faster because
3:28:57over here you have so many decisions and
3:28:59you have different things going on and
3:29:00you get the visibility is worse. And
3:29:02it's harder to improve this because it's
3:29:03harder to drill down where did something
3:29:05go wrong. Whereas over here if something
3:29:06goes wrong we just follow the trail and
3:29:08we see exactly what happened. So
3:29:11hopefully that makes sense. So here's a
3:29:12message that I actually sent to my team
3:29:15um about a month ago or so. Let me just
3:29:16read this out and then I'll break it
3:29:18down real quick. So the above weekly
3:29:20stats update was not meant to go here.
3:29:22This was like a public channel in our
3:29:23ClickUp. That was an automation of mine
3:29:25that I completely forgot existed. Back
3:29:27at the end of March, I set up an
3:29:28automation to keep the stats on my
3:29:30website current. So every week it pulls
3:29:31my YouTube subscriber count and school
3:29:33member numbers and it updates the
3:29:34website. And I built this on a separate
3:29:36account as a cloud routine. And it was
3:29:38mostly because I wanted to test it and
3:29:40then I made a YouTube video about it,
3:29:41but I completely forgot that that
3:29:42automation existed. It's been running
3:29:44quietly every week for 2 months and I
3:29:46completely forgot it existed, right? It
3:29:48works, which was the whole point of the
3:29:49test, but I never really vetted it. If
3:29:51we think back to my teaching a kid to
3:29:52ride a bike analogy, which I'll explain
3:29:54in just a sec. I basically put the kid
3:29:56on a bike with a helmet and training
3:29:57wheels and then I just walked back
3:29:58inside and took a nap. I told the
3:30:00automation to send me a personal DM each
3:30:02week, but prompting is not a permission
3:30:04layer. And the first five weeks it was
3:30:06sending me a DM in ClickUp, but then it
3:30:08started to go rogue and it started to
3:30:10send other people DMs and it started to
3:30:12drop these weekly stats updates in like
3:30:14general public channels rather than a
3:30:16private DM. Once again, just because I
3:30:17said, "Hey, only send it here," doesn't
3:30:19mean it's going to because it has access
3:30:21to every single channel in ClickUp.
3:30:22Sometimes it might make a mistake. So
3:30:24anyways, no harm done. Nothing sensitive
3:30:26happened. But it made me realize how bad
3:30:28that could have been. the fact that I
3:30:30completely forgot that an automation
3:30:31existed and the fact that I didn't have
3:30:33hard tool layer walls inside of this
3:30:36automation and was just relying on a
3:30:37prompt that could get bad really quick,
3:30:40especially because it was silently doing
3:30:41things that I forgot about. So anyways,
3:30:43hopefully there's some lessons there
3:30:44that you can take out of this. But what
3:30:46is the bike analogy? So basically the
3:30:47bike analogy is the idea that when you
3:30:50are building an automation, it's very
3:30:52much like you're teaching a kid to ride
3:30:53a bike. You can't expect to just put a
3:30:55kid on a bike and that they're going to
3:30:56be able to go 25 miles an hour down the
3:30:59road without falling ever. What you have
3:31:01to assume is that they're going to fall.
3:31:03So, when you put them on the bike,
3:31:04they've got training wheels, they've got
3:31:05helmets, they've got elbow pads and knee
3:31:07pads, and you're still going to hold the
3:31:09bike. You're going to push them. You're
3:31:10going to make sure that they are feeling
3:31:11comfortable. You're going to see, okay,
3:31:13you're leaning too much to the left,
3:31:14maybe shift more of your body weight to
3:31:15the middle. You're going to help them
3:31:16adjust. And slowly, you get to a place
3:31:18where you feel more confident. And if we
3:31:20relate this back to skills, you run the
3:31:22skill, you watch it, you very closely
3:31:24watch it, you give feedback, you run the
3:31:25skill again, you watch it, and you do
3:31:27that and slowly you're able to remove
3:31:28yourself more and more from that
3:31:30process. But what happens is even when
3:31:32the skill is like pretty battle tested,
3:31:34do you want to still just like say,
3:31:35"Hey, go off, kid. Take off your helmet.
3:31:38Go bike down the busy road and I'm going
3:31:40to go inside and take a nap." You're
3:31:41probably still going to like watch them
3:31:43a little bit. And you're going to wait
3:31:44till you feel comfortable to the point
3:31:46where there's basically nothing that
3:31:47could go wrong in this automation
3:31:49because you did all of the things that
3:31:50we talked about because you have the
3:31:52right system in place. That's the right
3:31:54amount of autonomy because you
3:31:55understand the difference between
3:31:56deterministic and nondeterministic
3:31:58because you set up the right prompting
3:31:59layers and tool layering so that you
3:32:01actually feel comfortable with the
3:32:02permissions that this thing has access
3:32:04to. And so like the majority of the work
3:32:06that we've been talking about in this
3:32:07course so far pretty much all lives down
3:32:09here. Like everything that we've been
3:32:10building are systems for a second brain
3:32:13and skills and capabilities so that we
3:32:16can have our agents help us do things
3:32:18much quicker and those are pretty much
3:32:19all things that are being triggered by
3:32:20us. And later when we start to talk a
3:32:22little bit more about deploying things
3:32:23and using cloud routines and stuff like
3:32:25that, that's where you really need to
3:32:27start thinking about these other
3:32:28elements. You know, like if you have a
3:32:29cloud skill that does something like
3:32:30let's just say um when a new lead comes
3:32:33in, you research the lead and then you
3:32:35shoot a message to the team in ClickUp.
3:32:37that probably doesn't need to be a human
3:32:39triggered thing. That's probably where
3:32:41an AI workflow actually comes into play
3:32:43a little bit better and you deploy that
3:32:44after you've evaluated it and you feel
3:32:46comfortable in it. So that is what I
3:32:48wanted to talk about here when it comes
3:32:49to trusting the outputs and obviously
3:32:52there's a lot to drill into once you
3:32:54really start to get into the weeds of
3:32:55building out like pretty big automations
3:32:57then eval comes into play. Evals are
3:33:00basically the idea of having a golden
3:33:02data set. So maybe 100 or a few hundred
3:33:04examples of input and expected output
3:33:07and then running those inputs on your
3:33:09system and seeing how many times did it
3:33:11pass, how many times did it fail, when
3:33:13it failed, why did it fail, and
3:33:14iterating on that automation, changing
3:33:16the prompt, changing the tools, watching
3:33:18for edge cases, and evaluating your
3:33:20systems or QA, quality assuring your
3:33:22systems before you ever push them into
3:33:24production or before you ever roll out a
3:33:26different model or a different prompt or
3:33:28whatever it is. Evals are very
3:33:29important. So hopefully now you're
3:33:31starting to understand some of these
3:33:32things that go into how you actually
3:33:34trust the output of these AI systems.
3:33:37And what you guys are probably starting
3:33:38to pick up on, which I think is
3:33:39important for me to call out here, is
3:33:41that so much of this stuff is
3:33:43non-technical. So much of this stuff is
3:33:45mindset and theoretical like oriented.
3:33:48And that is really really important to
3:33:50realize because as you start to doubt
3:33:52yourself or doubt your automations or
3:33:54have any sorts of doubt or discomfort,
3:33:57just use your words to figure it out.
3:34:00When I first started building
3:34:01automations in Cloud Code and I wanted
3:34:02to test them to see if they would
3:34:03actually work and see if they'd survive
3:34:05edge cases, I would literally say, "Hey,
3:34:07so you know how we just pushed this
3:34:08automation to Modal, which is something
3:34:10I'll talk about later, but we basically
3:34:12built an automation and we put it on a
3:34:13deployment site called Modal." I said,
3:34:16"Okay, so now that that's live, I'm a
3:34:18little bit confused and concerned about
3:34:20what could go wrong. So, help me figure
3:34:21out what are all the edge cases that we
3:34:23might want to think about here and how
3:34:25do we make sure that if something goes
3:34:26wrong, it's not like a bad failure. It's
3:34:29a safe failure." And so, this thing
3:34:31basically helped me look at all of the
3:34:33scenarios and then it designed tests for
3:34:35me. So, me and Claude with natural
3:34:37language went back and forth and we
3:34:38tested all these different scenarios. We
3:34:40tested invoices that came in weird. We
3:34:42tested what would happen if the API was
3:34:43down. We tested a bunch of these things
3:34:45and then because of those tests I said,
3:34:47"Okay, cool. Let's bake in all those
3:34:49guardrails so that if anything errors, I
3:34:51get a notification or the automation
3:34:53gets shut down or we always are doing
3:34:55things in a safe fashion so that the
3:34:56database won't get deleted so that
3:34:58records won't get merged so that nothing
3:35:00ever goes out externally so that we can
3:35:02fail safely and that we have logs of
3:35:04what's actually going on in our systems.
3:35:07Okay, so let's continue talking about
3:35:09how do we keep building out our AIOS
3:35:11here and giving it more memories, more
3:35:12subject matter expertise, more context
3:35:15about what we do. And that is kind of
3:35:17the element of building out our second
3:35:18brain. Just a quick warning before this
3:35:20next video starts playing. Some of the
3:35:22clips that I'm inserting into this
3:35:23course were recorded a few months back,
3:35:25meaning they might be shown in VS Code
3:35:28extension or the terminal instead of the
3:35:30cloud desktop app that we've been using
3:35:32so far. I just wanted to give you guys a
3:35:33warning. Functionally, exact same. So,
3:35:36don't worry about it too much. It just
3:35:37might look a little bit differently, but
3:35:39all you have to do is listen to what I'm
3:35:40saying and follow along with what I'm
3:35:41actually doing and you will be just
3:35:42fine. All of this stuff is still
3:35:44relevant. Otherwise, I wouldn't be
3:35:45putting it in this course. So, hopefully
3:35:47that makes sense. See you guys in the
3:35:49video. What you're looking at right here
3:35:50is 36 of my most recent YouTube videos
3:35:53organized into an actual knowledge
3:35:55system that makes sense. And in today's
3:35:57video, I'm going to show you how you can
3:35:58set this up in 5 minutes. It's super
3:35:59super easy. You can see here how we have
3:36:01these different nodes and different
3:36:02patterns emerging. And as we zoom in, we
3:36:05can see what each of these little dots
3:36:07represents. So, for example, this is one
3:36:09of my videos, $10,000 aentic workflows.
3:36:11We can see it's got some tags. It's got
3:36:13the video link. It's got the raw file.
3:36:15And it gives an explanation of what this
3:36:17video is about and what the takeaways
3:36:19are. And the coolest part is I can
3:36:20follow the back links to get where I
3:36:22want. There's a backlink for the WAT
3:36:24framework. There's a backlink for Claude
3:36:26Code. There's a backlink for all these
3:36:28different tools I mentioned like
3:36:29Perplexity, Visual Studio Code, Nano
3:36:31Banana, Nen. It also has techniques like
3:36:33the WT framework or bypass permissions
3:36:35mode or human review checkpoint. So as
3:36:38this continues to fill up, we can start
3:36:39to see patterns and relationships
3:36:41between every tool or every skill or
3:36:43every MCP server that I might have
3:36:45talked about in a YouTube video and I
3:36:46can just query it in a really efficient
3:36:48way now that we have this actual system
3:36:50set up. And the crazy part is I said,
3:36:52"Hey Cloud Code, go grab the transcripts
3:36:54from my recent videos and organize
3:36:56everything." I literally didn't have to
3:36:58do any manual relationship building
3:37:00here. It just figured it all out on its
3:37:02own. And then right here I have a much
3:37:04smaller one, but this is more of my
3:37:05personal brain. So this is stuff going
3:37:06on in my personal life. This is stuff
3:37:08going on with, you know, Upai or my
3:37:10YouTube channel or my different
3:37:11businesses and my employees and our
3:37:13quarter 2 initiatives and things like
3:37:15that. This is more of my own second
3:37:16brain. So I've got one second brain here
3:37:18and then I've got one basically YouTube
3:37:20knowledge system. And I could combine
3:37:22these or I could keep them separate and
3:37:24I can just keep building more knowledge
3:37:25systems and plug them all into other AI
3:37:27agents that I need to have this context.
3:37:29It's just super cool. So, Andre Carpathy
3:37:32just released this little post about LLM
3:37:33knowledge bases and explaining what he's
3:37:35been doing with them and in just a
3:37:36matter of few days, it got a ton of
3:37:38traction on X. So, let's do a quick
3:37:40breakdown and then I'm going to show you
3:37:41guys how you can get this set up in
3:37:42basically 5 minutes. It's way more
3:37:44simple than you may think. Something
3:37:45I've been finding very useful recently
3:37:47is using LLM to build personal knowledge
3:37:49bases for various topics of research
3:37:51interest. So, there's different stages.
3:37:53The first part is data ingest. He puts
3:37:55in basically source documents. So he
3:37:57basically takes a PDF and puts it into
3:37:59cloud code and then cloud code does the
3:38:00rest. He uses Obsidian as the IDE. So
3:38:03this is nothing really too gamechanging.
3:38:04Obsidian just lets you visually see your
3:38:06markdown files. But for example, this
3:38:08Obsidian project right here with all
3:38:10this YouTube transcript stuff that
3:38:11actually lives right here. This is the
3:38:13exact same thing. Here are the raw
3:38:14YouTube transcripts. And here's that
3:38:16wiki that I showed you guys with the
3:38:17different um folders for what Cloud Code
3:38:20did with my YouTube transcripts. And
3:38:22then there's a Q&A phase where you
3:38:24basically can ask questions about
3:38:26YouTube or about the research and it can
3:38:28look through the entire wiki in a much
3:38:29more efficient way and it can give you
3:38:31answers that are super intelligent. He
3:38:33said here, I thought that I had to reach
3:38:35for fancy rag, but the LLM has been
3:38:37pretty good about automaintaining index
3:38:38files and brief summaries of all
3:38:40documents and it reads all the important
3:38:42related data fairly easily at this small
3:38:44scale. So right now he's doing about 100
3:38:46articles and about half a million words.
3:38:48So there's a few other things that we'll
3:38:49cover later, but the TLDDR is you give
3:38:52raw data to cloud code. It compares it,
3:38:54it organizes it, and then it puts it
3:38:55into the right spots with relationships,
3:38:57and then you can query it about
3:38:59anything. And it can help you identify
3:39:00where there's gaps in that node or in
3:39:02that, you know, relationship, and it can
3:39:04go do research and fill in the gaps. All
3:39:06right. So why is this a big deal?
3:39:07Because normal AI chats are ephemeral,
3:39:09meaning the knowledge disappears after
3:39:11the conversation. But this method, using
3:39:13Carpathy's LLM wiki, makes knowledge
3:39:16compound like interest in a bank. People
3:39:17on X are calling it a game changer
3:39:19because it finally makes AI feel like a
3:39:20tireless colleague who actually
3:39:22remembers everything and it stays
3:39:23organized. It's also super simple. It
3:39:25will take you five minutes to set up.
3:39:27I'll show you guys. You don't need a
3:39:28fancy vector database embeddings or
3:39:30complex infrastructure. It's literally
3:39:32just a folder with markdown files.
3:39:34That's it. You literally just have a
3:39:35vault up top. So, in this example, it's
3:39:37called my wiki. You've got a raw folder
3:39:39where you put all of the stuff. And then
3:39:41you've got a wiki folder, which is what
3:39:42the LLM takes from your raw and puts it
3:39:45into the wiki. So in here you have all
3:39:46the wiki pages which it will create but
3:39:48then you also have an index and you have
3:39:50a log. So for example in my YouTube
3:39:52transcripts vault here is the index. You
3:39:54can see that I have all these different
3:39:55tools which I could obviously click on
3:39:56and it would take me right to that page
3:39:58or after that I have all the different
3:40:00techniques agent teams sub agents
3:40:02permission modes the WAT framework and
3:40:05then we've got different concepts MCP
3:40:07servers rag vibe coding we've got all
3:40:09these different sources which are you
3:40:10know the YouTube videos and then when I
3:40:12have people or when I have comparisons
3:40:14they will be put in here in the index
3:40:15and then we also have a log which is the
3:40:17operation history so in this case in the
3:40:19YouTube project the log isn't huge cuz I
3:40:21only ran one huge batch of the initial
3:40:2336 YouTube videos, but now every time I
3:40:25have one, I say, "Hey, can you go ahead
3:40:27and ingest the new YouTube video into
3:40:30the wiki and then we'll see every single
3:40:32time we update this." And then, of
3:40:33course, you need your claw.md to explain
3:40:35how the project works and how to search
3:40:37through things and how to, you know,
3:40:39update things. It's also a big deal from
3:40:41a cost perspective, token efficiency,
3:40:42and long-term value. One ex user turned
3:40:45383 scattered files and over 100 meeting
3:40:48transcripts into a compact wiki and
3:40:50dropped token usage by 95% when querying
3:40:52with Claude. And obviously token
3:40:54management and efficiency is a huge
3:40:56conversation right now and will always
3:40:57be. The other thing that's really cool
3:40:59about this is there's not really like a
3:41:01GitHub repo you go copy or there's not a
3:41:03complicated setup. You literally just
3:41:04say hey cloud code read this idea from
3:41:06Andre Karpathy and implement it. And
3:41:08people on X are now talking about like
3:41:10this is how 2026 AI agentic software and
3:41:13products will be made. You just give it
3:41:15a highle idea and it goes and builds it
3:41:17out. And Karpathy even said hey you know
3:41:19I left this prompt vague so that you
3:41:21guys can customize it and I'll show you
3:41:22the ways in my two different vaults
3:41:24right now that it changed things a
3:41:25little bit based on the context and
3:41:27understanding of what the project is
3:41:28actually for. Okay. So this was the
3:41:30original tweet I just showed you guys
3:41:31and then he followed up and said, "Hey,
3:41:32this one went viral. So here is the idea
3:41:34in a gist format." So if you open this
3:41:36up, this is basically just another
3:41:37explanation of the core idea of how this
3:41:39works and why the architecture indexing
3:41:42all this kind of stuff. And by the way,
3:41:44this is the part where he says, "Hey,
3:41:45this is left vague so that you can hack
3:41:46it and customize it to your own
3:41:48project." So we're going to come right
3:41:50back to this in a sec, but the first
3:41:51prerec that we're going to do, it's not
3:41:53necessary, but I like to have a nice
3:41:55little front end to see the
3:41:56relationships, is we're going to go to
3:41:58Obsidian and download it. So, if you
3:42:00just go to obsidian.mmd, you can see
3:42:02this is the completely free tool and
3:42:03you're going to go ahead and download
3:42:05it. So, just for your operating system,
3:42:07download this and then open up the
3:42:09wizard and open up the app. So, when you
3:42:11open up the app, it'll look like this.
3:42:13And what we're going to do here is we're
3:42:14going to create a new vault. So, down
3:42:15here, you can see I have Herk Brain and
3:42:17I have YouTube transcripts. I'll just
3:42:18make it a little bigger. I'm going to go
3:42:20to manage vaults. I'm going to create a
3:42:22new one. And now, we just have to give
3:42:24this a name. So, I'm just going to call
3:42:25this one demo vault. and you're going to
3:42:27choose a location where you want to put
3:42:29this. So, I'm just throwing this on my
3:42:30desktop and I'm going to go ahead and
3:42:32create this vault. Then, what you're
3:42:34going to do is go to wherever you like
3:42:35to run cloud code. So, in this case, I'm
3:42:37doing it in VS Code. And I open up that
3:42:39folder. So, demo vault. We get an
3:42:41Obsidian and then we get a welcome.md.
3:42:44So, I'm going to open up Claude. So, I'm
3:42:46going to do it in my terminal. I'm going
3:42:47to run Claude. And lately, I've been
3:42:48liking using my terminal better for
3:42:50Claude. I like to do it inside of VS
3:42:52Code, but the reason is just because I
3:42:53like to see the status line and I have,
3:42:55you know, a little bit more
3:42:56functionality. So, anyways, now that we
3:42:58have Cloud Code open, here's what we're
3:43:00going to do. We're going to go back over
3:43:01to the LLM wiki thing that we got from
3:43:03Andre Carpathy. We're going to copy all
3:43:05of this and we're going to go back into
3:43:08Cloud Code and then just paste it in
3:43:10there. So, that is the prompt from
3:43:13Carpathy that's going to build out
3:43:14everything we need. And then before we
3:43:16send that off, we're dropping this in
3:43:17which you guys can screenshot and then
3:43:19just throw into yours. But I'm saying
3:43:21you are now my LLM wiki agent. Implement
3:43:23this exact idea file as my complete
3:43:25second brain. Guide me step by step.
3:43:27Create the cloudmd schema. Blah blah
3:43:29blah. So this is just telling it what it
3:43:31needs to do with this idea that we just
3:43:33got from Kpapathy. So anyways, on the
3:43:35right we have this cloud code running
3:43:36and on the left we have our Obsidian
3:43:38vault. And you can see it just created
3:43:39those two folders. So it created the raw
3:43:41and it created the wiki as you can see.
3:43:43Now, by default, it threw in four
3:43:45folders. It threw in analysis, concepts,
3:43:47entities, and sources. Once we start to
3:43:48populate stuff, we can talk to it to see
3:43:50if that's actually the way we want to do
3:43:51it or not. Because it's interesting in
3:43:53my personal kind of second brain, the
3:43:56wiki is literally just markdown files.
3:43:57There's no structure to it. And in some
3:43:59cases, that's good. Carpathy actually
3:44:01said, "Sometimes I like to keep it
3:44:02really simple and really flat, which
3:44:04means like no subfolders and not a bunch
3:44:06of over organizing." But then you guys
3:44:08did see in my YouTube transcript one,
3:44:10there were different subfolders. And I
3:44:12think that in this case it actually
3:44:13makes more sense. But you can see that
3:44:15it went ahead and it created a claw.mmd.
3:44:17It created an index and a log and then a
3:44:19few different folders in our wiki. But
3:44:20now it's saying, "Hey, let's go ahead
3:44:21and try it out. Drop in your first
3:44:22source into the raw folder and tell me
3:44:24to ingest it." Okay, so I'm at this
3:44:26website called AI2027. If you guys
3:44:28haven't read this before, it's kind of
3:44:29an interesting read. So go check it out.
3:44:31And now let's say I want to get this
3:44:33into my vault. What I could do is just
3:44:35copy the whole page, right? And it might
3:44:37just come through a little weird. or we
3:44:39can just get an Obsidian extension which
3:44:41lets us basically take articles right
3:44:42from the web and just put it right into
3:44:44our vault. Super easy. So, search for
3:44:46this extension called Obsidian Web
3:44:47Clipper. You would go ahead and add this
3:44:49to Chrome. So, then when you're in the
3:44:50article that you want, you basically
3:44:52just click on your extensions, you open
3:44:53up Obsidian Web Clipper, and then you
3:44:55can just chuck it into your vault. And
3:44:56then right here, you're going to want to
3:44:57set this to RAW because this is the
3:44:59actual folder that it's going to put it
3:45:00in. So you can go ahead and click add to
3:45:01Obsidian. Open Obsidian. And then now
3:45:03you can see in my RAW section, we have
3:45:06this AI 2027 source with the title, the
3:45:09source, and it's not super super
3:45:11populated yet because the LLM in cloud
3:45:14code is going to do that. So here is our
3:45:16file. I'm going to open up Cloud Code
3:45:18and say, awesome. I just threw in an
3:45:19article called AI2027 into the RAW. Can
3:45:22you please go ahead and ingest that? It
3:45:24might ask you some questions. It might
3:45:26also be helpful to before you start
3:45:27ingesting stuff say, "Hey, by the way,
3:45:29this project is specifically for my
3:45:30second brain." So, personal things,
3:45:32business things, whatever. Or this is
3:45:34just a research project. This is where
3:45:36I'm going to chuck you all the articles
3:45:37and all the things that I want to learn
3:45:38about and all the things that I know.
3:45:40So, there's different ways that you can
3:45:41set up the project as you saw with mine.
3:45:43One for YouTube, one for just personal
3:45:45second brain. So, now what it's doing is
3:45:47it's going to read through this article
3:45:48and then it's going to figure out where
3:45:50should I chuck everything into the wiki.
3:45:52It's not just going to create one MD
3:45:54file for this. it might create five or
3:45:56it might create 10. And there's going to
3:45:57be relationships between each of the
3:45:58different sections that it creates. So,
3:46:00it's kind of doing its own method of
3:46:01chunking. Now, one thing I want to call
3:46:03out real quick is with this extension,
3:46:05if you go here and you open up the
3:46:07options for it, you can see that you can
3:46:09actually change where by default the
3:46:12folders are dropped, which is in the
3:46:14location section. By default, it'll be
3:46:15going to a place called clippings, but
3:46:17just go ahead and change that to raw.
3:46:19Okay. So, here it came back with all
3:46:20these questions, right? It said, "Here
3:46:22are my key takeaways from this article,
3:46:23blah blah blah." And now it'll ask you,
3:46:25what do you want to emphasize from this
3:46:26article? What's your focus? How granular
3:46:28do you want to be? What's your plan? So,
3:46:30I'm just going to say, I want this to be
3:46:32extremely thorough. This is my passion
3:46:35looking at where AI is going to go. Um,
3:46:37and this whole project, by the way, that
3:46:38you're setting up in this vault is
3:46:40basically just going to be my place to
3:46:42dump in research about AI. So, help me
3:46:44keep all that organized so that I can
3:46:45query it and that I can, you know, keep
3:46:47my thoughts related. So that's just a
3:46:49quick example of what it might look like
3:46:51for you to give it some more context to
3:46:53continuously build your project. So I'm
3:46:55going to switch over over here to the
3:46:56graph view because I think it'll be
3:46:58interesting to see as it is starting to
3:47:00go through and create those different
3:47:01wiki files. It's going to go ahead and
3:47:03it's going to create all those
3:47:04relationships and we'll be able to watch
3:47:05it in real time. All right, so it's
3:47:07creating all of the wiki pages now and
3:47:09you can see that it said it's going to
3:47:10make about 25 because there's so much
3:47:12stuff going on in the original AI 2027
3:47:15article. Okay, so our first one just
3:47:17popped in here and there a second one
3:47:18just came through and now you can
3:47:20understand you're starting to see where
3:47:21do you have hubs or where do you just
3:47:22have little individual nodes so this is
3:47:24a major hub someone named Eli someone
3:47:27named Thomas Daniel and you can see all
3:47:29the different relationships here with
3:47:31things like AI governance with things
3:47:33like open brain superhuman coder okay so
3:47:36that ingest took about 10 minutes so
3:47:38sometimes you have to be a little
3:47:39patient with you know it reading through
3:47:40everything and organizing everything but
3:47:42it does a lot of heavy lifting of course
3:47:44when I uploaded the 36 6 YouTube
3:47:46transcripts in batch. It took about 14
3:47:48minutes. So, it kind of just depends,
3:47:50but it created 23 wiki pages. We have
3:47:52the source. We have six people, five
3:47:55organizations, and one AI systems page.
3:47:57Different concepts, so technical,
3:47:59alignment, and geopolitical. And then an
3:48:01analysis, and then it asks some
3:48:03questions about it so that it can help
3:48:04make the relationships and make the
3:48:06structure even better. Now, let's just
3:48:08open this one up a little bit deeper and
3:48:10see what it actually did in here with
3:48:12this stuff. So we have this is the
3:48:14source with all the main relationships.
3:48:16So as we start to add other articles, we
3:48:18will see other big kind of like nodes
3:48:20and maybe in some cases we'll have
3:48:21relationships between like compute
3:48:23scaling with different articles that we
3:48:25upload as well. So let's just see if I
3:48:26click into the main source we can see
3:48:28the tags that it got. We can see the
3:48:30authors and we can click around. So
3:48:32here's a link to OpenAI. Okay, what's
3:48:34OpenAI? Here's references in AI 2027.
3:48:36Here's some other connections with
3:48:37OpenAI like model spec. Okay, we're in
3:48:39modelsp spec. Let's take a look. We can
3:48:41see other things about model spec and we
3:48:43could also go to how the LLM psychology
3:48:45model works. So this is just super super
3:48:47cool all the relationships that we get
3:48:49and once again all of this stuff that
3:48:51we're looking at was derived from one
3:48:53article and automatically organized and
3:48:55related. So the question now is like
3:48:57what do we do from here? Do we query it
3:48:59inside of this environment? Do we query
3:49:00it from somewhere else? And that's
3:49:02completely up to the way that you want
3:49:03to use this. So, for example, with my
3:49:05YouTube project, I'm probably just going
3:49:07to keep this here and whenever I want to
3:49:08ask questions about YouTube or if I want
3:49:10to turn this into like a website, I can
3:49:12just do that from here. Or if I need to,
3:49:14I can point a different project at this
3:49:16folder since everything's here and it
3:49:18can crawl through the wiki. It can read
3:49:19the index and it knows how this stuff
3:49:21works because you can give it the claw
3:49:23MD so it understands the project as
3:49:24well. So, for example, in this one,
3:49:26which is just my second brain where we
3:49:28have all of the different things about
3:49:29like I drop in my meeting recordings, I
3:49:31drop in, you know, ClickUp channels,
3:49:33summaries, and things like that. This is
3:49:34something that I want to use in my
3:49:35executive assistant. So, what I did in
3:49:37my executive assistant here called Herk
3:49:392. If I go to this claw.md, you can see
3:49:41that we have a wiki path. So, whenever
3:49:43you need to read things about me and my
3:49:46business that you don't have already,
3:49:47you would basically go to my Herkbrain
3:49:49vault. You would go to that directory
3:49:51and then you would read through the
3:49:52wiki. You can read the hot cache, which
3:49:53I'll explain in just a sec. You can read
3:49:55the index. You can read the domain
3:49:57subindex. And then you can also just
3:49:59search through everything here. And I
3:50:00said don't read from the wiki unless you
3:50:02actually need it. Here are some things
3:50:03that you might do that you don't need to
3:50:05go read the wiki for. And all of this is
3:50:06my business knowledge. Now, if you guys
3:50:08remember, if you watch my video on
3:50:09setting up an executive assistant, I
3:50:11used to do this with context files
3:50:13inside of this project. And when I
3:50:15changed over to this method, I actually
3:50:16saw a reduction in tokens that I was
3:50:19actually calling in this project. So the
3:50:21thing about the hot cache, right, I
3:50:23didn't actually have this in my YouTube
3:50:25one. So if I go to YouTube, you can see
3:50:26there's no hot cache, but if I go to the
3:50:29herk brain in the wiki, you can see
3:50:31there's a hot.md right here. And this is
3:50:33basically just a cache of like 500 words
3:50:35or 500 characters that it saves, which
3:50:37is like what is the most recent thing
3:50:39that Nate just gave me or that we talked
3:50:41about. In the context of my executive
3:50:42assistant, this is really helpful. You
3:50:44know, it might save me from having to
3:50:45crawl different wiki pages. But in
3:50:48something like the YouTube transcript
3:50:49project, I don't really need a hot
3:50:51cache. So, another thing that I alluded
3:50:53to but didn't really cover was the idea
3:50:54of linting. So, Karpathi says that he
3:50:56runs some LLM health checks over the
3:50:58wiki to find inconsistent data, impute
3:51:01missing data with web searches, find
3:51:03interesting connections for new article
3:51:05candidates, things like that. So, it
3:51:07basically helps you run a lint, you
3:51:09know, every day, every week, whenever
3:51:10you want, which helps make sure that
3:51:12everything is scalable and structured in
3:51:14the right way. And it might even come
3:51:16back and say, "Hey, I don't fully
3:51:17understand this. Can you give me some
3:51:18more info or can you grab some more
3:51:20articles that might help me out here?"
3:51:21So now the final question about this
3:51:23that I wanted to cover is like, "Does
3:51:24this kill semantic search rag?" And the
3:51:27answer is no, but kind of yes. And it
3:51:30all depends on the goal of the project
3:51:32and the goal of the context, how much
3:51:34context you have. So here's a really
3:51:35quick chart that I had my claude code
3:51:37make. I was in my Herk brain where I
3:51:39dumped in a bunch of information about
3:51:41Karpathy's LLM knowledge and I just
3:51:43said, "Hey, can you please explain
3:51:45Karpathy knowledge as simple as
3:51:46possible, keep it super concise, and um
3:51:49compare it to typical semantic search
3:51:52rag." So, it found Carpathy's idea.
3:51:54Instead of a database, you just give the
3:51:56LM well organized markdown files and it
3:51:58compares it here to the actual semantic
3:52:01search rag. So, actually, I might as
3:52:02well just read it off from here. So it
3:52:04finds it by reading indexes and follows
3:52:06links rather than using similarity
3:52:08search. So we're getting a deeper
3:52:09understanding of relationships because
3:52:11they're links rather than just saying,
3:52:13"Hey, these chunks seem similar." As far
3:52:14as infrastructure, it is literally just
3:52:16markdown. So like I said, you don't even
3:52:17need the obsidian. You just need these
3:52:19markdown files. Whereas with semantic
3:52:22search, you need an embedding model. You
3:52:23need a vector database and a chunking
3:52:24pipeline. The cost over here is
3:52:26basically free. Your only cost is going
3:52:28to be tokens. Whereas over here, you
3:52:29might have ongoing compute and storage.
3:52:31And for maintenance, you just run a
3:52:33lint. You clean up things. You add more
3:52:35articles. You give it more context
3:52:37rather than having to re-mbed when
3:52:39things change. But right now, the
3:52:40weakness of course with the uh LLM
3:52:43knowledge wiki is that it doesn't scale
3:52:46huge across enterprises, right? Because
3:52:47it's just a bunch of files. Um and that
3:52:50is where the cost will probably get more
3:52:51and more expensive than going to
3:52:53something like standard semantic search
3:52:55or knowledger graph or light rag or
3:52:57whatever other tool is out there for
3:52:59that. So here you can see if you have
3:53:00hundreds of pages with good indexes,
3:53:02you're fine with wiki graph. But if you
3:53:03were getting up to the millions of
3:53:05documents, then you're going to want to
3:53:06actually do more of a traditional rag
3:53:08pipeline. Today I'm going to explain the
3:53:09different levels of building your own AI
3:53:11second brain. You can see here we have a
3:53:13visual of three very different types of
3:53:15data. This one is where we have our
3:53:17context really starting to form and
3:53:18we're starting to see some relationships
3:53:20and we're starting to see some different
3:53:22nodes and entities form. And then as we
3:53:24continue to scale this up, add more
3:53:25knowledge, more knowledge, more
3:53:27relationships, we start to get something
3:53:28that looks a little bit more like this,
3:53:30where we have clearly different
3:53:31clusters, and inside of all of these
3:53:33nodes, we can see how they relate to
3:53:34each other. And then over here, we're
3:53:35taking all of those relationships a step
3:53:37further, and we're able to then start to
3:53:39see how everything really pieces
3:53:40together rather than just having files
3:53:42that sort of link back to each other.
3:53:44This is relationship mapping. And so
3:53:46really the idea of an AI second brain
3:53:48has blown up because we're all trying to
3:53:50get as much information out of our heads
3:53:52into our systems as possible. That's the
3:53:54true value. Your moat is your data. It's
3:53:56your IP. But the process of organizing
3:53:58that into a system so that you can use
3:54:00it with a bunch of different AI models
3:54:02and so that it can actually recall
3:54:04things in a way that makes sense rather
3:54:05than just hallucinating or spending a
3:54:07bunch of your time and tokens trying to
3:54:08look through everything. That's the
3:54:10issue. So clearly all of this is my real
3:54:12data and this is what the actual project
3:54:13looks like. It is my Herku project. I
3:54:15have a bunch of folders and files here.
3:54:17And at the end of the day, that's
3:54:18basically all it is. It is markdown
3:54:20files that are organized in a way that I
3:54:21understand and that my agents
3:54:23understand. And so, yes, I'm going to
3:54:24walk you guys through what I have here
3:54:25and how it works. But I also have this
3:54:27other project where I'm going to show
3:54:28you if you're starting from scratch or
3:54:30if you feel like maybe you're in between
3:54:32level two and three, how we can actually
3:54:34look at the differences and what it
3:54:35might look like to scale up your own
3:54:37systems and start to add context in
3:54:39different ways. So, super excited to dig
3:54:41into this today and I don't want to
3:54:42waste any of your guys' time. So let's
3:54:43just start looking at these five levels
3:54:45and how they differ. All right. So every
3:54:47level of a claude code second brain and
3:54:49I'm going to be obviously kind of
3:54:50referring to claude code a lot but keep
3:54:52in mind this can be used with any AI
3:54:54model. I use my second brain all the
3:54:56time with codecs as well. I use it with
3:54:57Hermes agent. This can be used by
3:54:58different agent harnesses because it's
3:55:01just files and folders. So what is the
3:55:04actual job of a second brain? A lot of
3:55:05people probably define this differently,
3:55:06but the way that I think about it is
3:55:08that it's a place for me to save notes,
3:55:11meeting recordings, ClickUp threads,
3:55:13stuff like that. I can save it there and
3:55:15then it helps me basically ingest it and
3:55:17get it into the right spots so that it
3:55:19can actually find it later. And so
3:55:20that's really the thing to think about
3:55:22is can your agent find it again and
3:55:24could you find it again? Because if the
3:55:26answer is no, then you probably don't
3:55:27have the right routing or folder
3:55:29architecture set up, which is what I'm
3:55:31here to talk about today. And one other
3:55:32sort of mindset thing that I want to get
3:55:35out there before we dive into these five
3:55:36levels is that you kind of have to work
3:55:39backwards. You want to reverse engineer
3:55:41based on the question. So this will
3:55:42start to make more sense as we get into
3:55:44it. But really what you should be
3:55:45thinking about is how do I want to use
3:55:47this data in the future because how it's
3:55:50going to be accessed and recalled
3:55:52determines the way that you put it in in
3:55:54the first place. For example, a
3:55:56basketball hoop and a basketball. We
3:55:58know what shape the hoop is and we know
3:56:00that the ball needs to go through. So,
3:56:01why would we ever design the ball to be
3:56:04a giant square? Because it just wouldn't
3:56:07fit through the hoop. So, that would
3:56:08make no sense. So, you need to start
3:56:09with the end in mind a little bit. Once
3:56:11again, I will show you exactly what I
3:56:13mean by that as we continue on. Because
3:56:15remember, we're trying to get to the
3:56:16point where your second brain knows
3:56:18everything about your business, about
3:56:20you, your relationships. It knows
3:56:21everything to the point where it
3:56:23probably can recall stuff better than
3:56:25you can because it has a better memory
3:56:26and it can search through things way
3:56:28faster than you can. So, we've got five
3:56:29different levels to talk about, and they
3:56:32each kind of have different questions.
3:56:34So, level one is, can you find the file
3:56:35or the info by looking for an exact word
3:56:38or name? Level two is, can you pull
3:56:40everything on a certain topic together?
3:56:42Level three is, I searched for different
3:56:44words than I wrote. So, semantic search,
3:56:46you're searching for meaning rather than
3:56:47an exact word match. And then trace
3:56:50relationship chains. Can you ask about
3:56:52topic X? And then trace that all the way
3:56:55back to topic A. And then level five is
3:56:57just kind of making this whole second
3:56:59brain thing super autonomous to the
3:57:01point that you don't even have to think
3:57:02about it. And by the way, this isn't me
3:57:04saying that number five is best. I have
3:57:06some arguments about why I do not
3:57:07currently sit on level five. The point
3:57:09I'm trying to make here is each level is
3:57:11different and you want to find the
3:57:13simplest level or the lowest level that
3:57:15actually fits your needs. If you don't
3:57:17have a painoint in your system, then I
3:57:19don't really think there's a need to go
3:57:21experiment or develop a new sort of, you
3:57:23know, architecture. If there's not pain,
3:57:26then why create more? Okay, so level one
3:57:29is pretty simple and this is where you
3:57:30always start. So you start with a
3:57:32claw.md or if you're using codeex or
3:57:34something, you would start with an
3:57:35agents.mmd. But you start with a
3:57:37cloud.mmd which is kind of, you know,
3:57:39that gets loaded up. That's almost like
3:57:40the system prompt for that session for
3:57:42that project. And then you've just got a
3:57:43bunch of folders and files. But the key
3:57:44part there is the cloudmd is kind of
3:57:47treated as a router. So yes, you've got
3:57:48some, hey, this is your role. Here is
3:57:50what's important. But you also have
3:57:52routing rules. If you ever need to find
3:57:54information about me personally, look in
3:57:55this folder. If you need information
3:57:57about our quarter one priorities, look
3:57:58in this folder. Because if you've ever
3:58:00had a point where you ask Claude to do
3:58:02something, and then it asks you, "Hey,
3:58:04can you give me more info? I don't know
3:58:05what you're talking about, but you know
3:58:06there's files and folders in your
3:58:08project, then you probably just didn't
3:58:09give Claude the knowledge to go look
3:58:12there." It's not just going to go search
3:58:13your entire codebase automatically. I
3:58:15mean, you wouldn't want it to do that
3:58:16because it's going to waste your time
3:58:17and your tokens. So, if it doesn't know
3:58:19if something lives somewhere, then it's
3:58:21probably not going to be able to find
3:58:22it. So, when this is properly set up,
3:58:24you will stop having to reexplain
3:58:25things. You will talk to it and it will
3:58:27just know where to go look and why. But
3:58:29the problems with this is that if it
3:58:30grows too big, it can start to get messy
3:58:32and feel ignored. And this is typically
3:58:34more of like an exact word type of
3:58:36search depending on the way that you
3:58:37route. So, if I open up my um example
3:58:40project here, let's open up level one.
3:58:43So, in level one, what you can see,
3:58:44pretend this is its own cloud project.
3:58:46We've got a clawmd. So let me click into
3:58:48that. We can see here it says this file
3:58:50loads automatically every time you open
3:58:52cloud code in this folder. It is the one
3:58:53file that tells the AI who you are, how
3:58:55you work, and where things live. At
3:58:56level one, this file plus a few folders
3:58:58is your entire second brain. So here's
3:59:00kind of like that basic knowledge. And
3:59:01then right here, it's this simple where
3:59:03things live in the context folder.
3:59:05Always true background about you and how
3:59:07you work. Read this first projects
3:59:09decision log. And that's basically it.
3:59:11So right here, you can see there's a
3:59:12context folder. We have an about me file
3:59:14which you could grow. We have stack and
3:59:16conversations file. We have decisions.
3:59:18So this is a decision log where you can
3:59:20have your cloudmd always append new
3:59:22decisions and the dates whenever you
3:59:24make a big change to your project or to
3:59:25your life or to your business. And then
3:59:27we have projects. So this is where you
3:59:28could have a markdown file or even
3:59:30folders within the projects for all of
3:59:32your ongoing projects, all of your
3:59:33ongoing clients, whatever it is, however
3:59:35you want to organize it. That's where
3:59:36you can have some projects. And you can
3:59:38even start to organize these things by
3:59:39dates if you want. So, if you want to
3:59:40just have one that's for like May and
3:59:42then you have all of those stuff and you
3:59:43have one for June. The thing that I
3:59:45really want to stress here with level
3:59:46one and the thing that I answer a lot in
3:59:48my community in the comments is that
3:59:50there is not yet a standard way that has
3:59:53been proven the best way to set up your
3:59:55projects or your second brain besides
3:59:56some of the most common things like your
3:59:58context and your cloudmd and your, you
3:59:59know, whatnot. But the point I'm trying
4:00:01to make there is don't see what I do and
4:00:05think that that's the right way or see
4:00:06what someone else you watch does and
4:00:08think that that's the only right way.
4:00:10All that matters is do you have proper
4:00:13routing in place and does it make sense
4:00:15to you and does it make sense to your
4:00:17AI? Okay, so let's say I have my Herk 2
4:00:20project right here and I need to find
4:00:21something in here but I can't ask AI for
4:00:23some reason. what I need to find is easy
4:00:25because I understand the drill downs.
4:00:27You know, I understand my base folders.
4:00:28And let's say I'm looking for the HTML
4:00:30slide deck I built for my ranking cloud
4:00:33code features video. I would come into
4:00:35here and I say, "Okay, I know that's a
4:00:36project." So, I'll go there. Within my
4:00:38projects, I've got another project for
4:00:39YouTube videos. I'll open that up. And
4:00:41now I know I made this video right here,
4:00:44May 30th Claude Code top 50 features. In
4:00:47here, I have the actual tier list deck.
4:00:49And when I open that up, now I have the
4:00:50slide deck. And not only can I find it
4:00:52easily, but my agent can find it because
4:00:54it all makes sense and I have routing
4:00:55rules. Real quick, guys, if you're
4:00:57watching this video, you're probably
4:00:58interested in building your own AI
4:01:00operating system. Lucky for you, I have
4:01:01a full free course on that in my free
4:01:03school community. The link for that is
4:01:04down in the description. Join the free
4:01:06school community. Hop in here, take the
4:01:077-day challenge, build your own AI
4:01:09operating system, and apply these
4:01:10principles into building your second
4:01:12brain, which will make your AI operating
4:01:14system even more powerful. So, links in
4:01:15the description. Let's get back to the
4:01:16video. Awesome. Okay, so that is how you
4:01:18start. Now, as you move up to level two,
4:01:21you might be able to start to work in
4:01:22some things like the LLM wiki, which is
4:01:24what I've got set up for a few different
4:01:26things. This is the whole Carpathy LLM
4:01:28wiki, which I did make a full video
4:01:29about. If you want to check that out,
4:01:30I'll tag that right up here. But this is
4:01:32when you start to have more files and
4:01:34and they start to take a bit of a
4:01:36different shape and you want to organize
4:01:37them together in a bit of a different
4:01:39way. So, it could be really good for
4:01:41researching all on a certain project. It
4:01:42could be really good for, you know, a
4:01:44few of the ones that I've got set up is
4:01:45my YouTube transcripts all live in their
4:01:47own wiki. I've got all of like my
4:01:48meeting transcripts that live in their
4:01:50own wiki. So, for example, this is the
4:01:51obsidian view of my wiki for all of my
4:01:54YouTube video transcripts. You can see
4:01:56here if I go to wiki, you can see
4:01:57there's main concepts like aentic
4:01:59workflows, AI coding market, context
4:02:02window, and all of these in here start
4:02:03to relate back to other tools and
4:02:06concepts and videos and stuff like that.
4:02:07So, we've got the sources, we've got
4:02:09platforms, we've got um context
4:02:11management techniques, and all of this
4:02:13was autocreated by our cloud code when I
4:02:16told it to ingest this YouTube
4:02:18transcript into our wiki. So, I'm not
4:02:19going to dive super super deep into all
4:02:21this right now, but definitely check out
4:02:22that YouTube video I linked. Now, what
4:02:24else is cool about this is this
4:02:26transcript wiki actually lives within my
4:02:28main Herk 2 project. So, here's Herk 2.
4:02:30If I go right here to other worlds and
4:02:33then I go down to YouTube OS and I click
4:02:35into the transcript wiki right here.
4:02:38This is what we were just looking at in
4:02:39Obsidian. We could see the concepts. We
4:02:41could see the comparisons. We could see
4:02:42sources, techniques. This is what we
4:02:44were looking at in Obsidian. So all
4:02:46Obsidian is is it basically just
4:02:47visualizes your markdown files. You see
4:02:50here wiki concepts, comparisons,
4:02:52techniques. This is what we were just
4:02:53looking at. All we get now is we just
4:02:55get a visual view of all that. And so
4:02:57the reason I wanted to bring that up as
4:02:58well is because I think a lot of people
4:03:00obviously get pretty infatuated by that
4:03:03visual view. And obviously I started the
4:03:05video with that because I think that's
4:03:06what hooks a lot of people in. But all
4:03:08that really matters is can your system
4:03:10grab that and give it to you. If you are
4:03:12a visual person and you really want that
4:03:13view, then by all means install Obsidian
4:03:16and set it up. It's super easy. But I'm
4:03:18saying that you don't always need that
4:03:19visual layer if it's not beneficial to
4:03:21you. I hardly ever open Obsidian just to
4:03:23be honest because I know that it all
4:03:25lives here and I know that my second
4:03:27brain and my OS can find all of that. So
4:03:29anyways, in level two here, let's look
4:03:30at this. It's very similar in shape to
4:03:33level one. It's just building on top of
4:03:34it because now we have our cloudMD which
4:03:37starts to route to some other things
4:03:38because it routes to the wiki and it
4:03:40still routes to context projects
4:03:42decisions, but it's also routing to
4:03:44references and memory. So we're just
4:03:46starting to add a bit more of these
4:03:47routing rules inside of the claw.md. We
4:03:50can grow the context. We can grow the
4:03:52decisions. We can grow projects and
4:03:53references. And we can also start to get
4:03:55this idea of memory. And what's really
4:03:57cool about this is you can turn on
4:03:58automemory in cloud code. And the AI
4:04:00will basically start to write this file
4:04:02and update it on its own. So you don't
4:04:04even have to think about it. If you come
4:04:05in here and you do /memory, it'll say
4:04:07automemory on or off. And if it's off,
4:04:09if you want to turn that on, just turn
4:04:11it on. And now one thing to think about
4:04:12is I mentioned earlier that we want to
4:04:14make our second brains tool agnostic.
4:04:17And this is one thing that's pretty
4:04:18specific about cloud code is it uses
4:04:20claw.mmd and it uses this memory.mmd and
4:04:23it keeps that updated on its own. So if
4:04:25you wanted to move this over to codeex,
4:04:26what you would do is you would first of
4:04:28all transition your claw.md. You'd make
4:04:30a copy of it called agents.mmd. As you
4:04:32can see here in my herk 2, I've got my
4:04:34if I scroll down claw.md right here and
4:04:36then I've got agents.mmd right here. And
4:04:38they're essentially the exact same file.
4:04:39Just so codex can read this one and
4:04:41cloud code can read this one. But
4:04:42because claude code keeps that
4:04:43automemory, all you need to do is make
4:04:45sure you have that memory MD file and
4:04:47just tell codecs, hey, by the way, for
4:04:49memories, look in our memory.mmd file.
4:04:52It's all about the routing there.
4:04:53Anyways, just felt like that was
4:04:54important to throw out. But at a certain
4:04:56point when you have these, you know,
4:04:57wiks, they do start to degrade a little
4:04:59bit because what's what's great about
4:05:00them is that they have indexes, right?
4:05:02So when your AI starts to look in the
4:05:04wiki, it knows, okay, if the user is
4:05:07asking about agentic workflows, I'm
4:05:08probably going to start here. And then
4:05:10from here, I'm going to drill down and
4:05:11read this to see what else is important
4:05:14to them. Maybe they're asking about the
4:05:15WAT framework, and then I can drill into
4:05:17that. And maybe from there, I need to
4:05:19learn a little bit more about the
4:05:19cloudmd system prompt, and then I will
4:05:21drill into that. So there are
4:05:23relationships here a little bit, but
4:05:25this isn't the same as like semantic
4:05:27relationships or knowledge graph
4:05:29relationships that have more meaning.
4:05:31This is more about just actually
4:05:33following a trail and reading the page
4:05:35in its entirety. And I'll be fully
4:05:37honest with you guys. I pretty much sit
4:05:40my entire Herk 2 project in this level
4:05:42in level two because this has been
4:05:44working really well for me. Like I
4:05:45mentioned earlier, I haven't felt a pain
4:05:47yet big enough to switch over to level
4:05:49two. And here's what I meant by that. My
4:05:50wiki has links. Isn't that a knowledge
4:05:52graph? Not exactly. Because this doesn't
4:05:56have connections of how they are
4:05:58related. like this is endorsed by this
4:05:59or this has cron to here. These just
4:06:02have connections because it's like a a C
4:06:05also. It's like backlinks. So, they're
4:06:07very similar and yes, they can achieve a
4:06:08similar effect, but it's still a little
4:06:11bit different. Anyways, let's take a
4:06:12look at level three, which is where you
4:06:14start to do things like semantic search.
4:06:16Whether you do that in Obsidian, whether
4:06:18you do that with Pine Cone or Superbase,
4:06:20however you start to grab the actual
4:06:22semantic search, that is what level
4:06:24three is. And so just as a quick visual
4:06:26for you guys, let's take a look at this
4:06:29quadrant cluster of images. So every one
4:06:32of these vector points is an image. And
4:06:35what we see in here as the payload is
4:06:37stuff like the file name, the URL, the
4:06:38name of the author or the artist and the
4:06:41URL. But we don't actually see like
4:06:43what's in the image. We don't get a
4:06:44description. So what we have to do is we
4:06:46have to organize these images by meaning
4:06:48or by similarity. So, when I open up
4:06:50this graph and we start to visualize the
4:06:52stuff here, what you see is that we have
4:06:54this main image, these owls, these kind
4:06:56of like I I don't even know. Um, it's a
4:06:58very trippy style, like hallucenic
4:07:00style. Anyways, then this one is kind of
4:07:02similar, right? It's got those colors.
4:07:04It's got the paint. This one is also
4:07:06similar, but they're not the same. They
4:07:08just share similarities. And as we start
4:07:10to expand these more and more, we can
4:07:12start to get into different styles. So,
4:07:14this one has like some creepy eyes and
4:07:15mushrooms or whatever. This one is kind
4:07:17of more down that fantasy lane. And as
4:07:19we start to build out more of these
4:07:20relationships and meanings, we can
4:07:22expand and grow away from them. And so
4:07:24quadrant really just gives you a
4:07:25visualization here. I mean, it's a it
4:07:27has clusters and vector store, but the
4:07:30reason I pulled this up as a demo is
4:07:31just because we start to see the actual
4:07:33relationships form here based on
4:07:35meaning. And that's what's important
4:07:37about semantic search is that we're no
4:07:38longer doing keyword matching. We're
4:07:40searching based on meaning. So here in
4:07:42my YouTube transcript second brain if I
4:07:45go to the smart lookup over here this is
4:07:47very different from just the regular
4:07:49search. So for example if I search here
4:07:51for um feedback let's say we're actually
4:07:56doing a match on the word feedback and
4:07:58it's only showing me where that word
4:08:00actually appears inside of our second
4:08:02brain. But if I come over here in the
4:08:04smart lookup and I search for feedback
4:08:06we are getting matches that have things
4:08:09in here that mean feedback. So live test
4:08:11results, cloud code skills, which was
4:08:13talking about evaluations and stuff. So
4:08:15there's a big difference between keyword
4:08:16matching and semantic search, you know,
4:08:18similarity matching. This one over here
4:08:20is saying X equals X. And this one is
4:08:22saying X is similar to X, Y, and Z. And
4:08:25so this all just goes back to vector
4:08:26databases. I've talked so so much about
4:08:29vector databases, so I'm not going to
4:08:30dive super deep in. And I've got so many
4:08:32resources on my channel, but basically
4:08:34what it is is we take a document. So
4:08:36let's just say YouTube transcript. We
4:08:38chunk it up and then each chunk is ran
4:08:40through an embeddings model. And the
4:08:41embeddings model puts that chunk of text
4:08:44onto like a three-dimensional space
4:08:46where space is related to meaning. And
4:08:49so it decides, okay, this chunk is about
4:08:50a company, so we're going to put it up
4:08:52here. This chunk is about finances, so
4:08:54it's going to go here. And we start to
4:08:55see these vectors form near other
4:08:57similar vectors. Now, do you guys
4:08:59remember how I said earlier like you
4:09:00want to think about how is the data
4:09:02going to be used? What type of questions
4:09:04are you going to ask? This is a reason
4:09:05why that's so important. So, think about
4:09:07this. Let's say I put my meeting
4:09:09transcript of March 5th meeting into my
4:09:12second brain and I put those in as you
4:09:14know vectorized chunks. So, let's say
4:09:16when I vectorize that meeting, we
4:09:18actually get, you know, like 20 chunks.
4:09:21It actually creates 20 chunks or however
4:09:23many that is. And then when I say, "Hey,
4:09:24Mr. AI agent, can you summarize the
4:09:27meeting on March 5th? It will basically
4:09:29search for March 5th meeting summary and
4:09:32it will pull chunks that are similar to
4:09:34March 5th meeting summary. And then even
4:09:36if it gets the right chunks, it's going
4:09:38to only summarize those five chunks.
4:09:40It's not able to look at the entire
4:09:42meeting summary or sorry like meeting
4:09:43transcript in entirety. So it doesn't
4:09:45really know a summary. It might be
4:09:47missing a lot of key information. Now,
4:09:48yes, there are things you can start to
4:09:49play with there like metadata and other
4:09:51things like that to make these results
4:09:52better. But at the end of the day,
4:09:55people kind of assume that a vector
4:09:57database was some magic solution where
4:09:59it could always pull back what you need.
4:10:00But that is very false. And I mean,
4:10:02think about it like this. Let's say we
4:10:03have a table and we say, "Hey, which
4:10:05week do we have the highest sales?"
4:10:07Okay, the agent looks for highest sales.
4:10:09It maybe grabs this chunk outlined in
4:10:11gray of data and then it looks at, okay,
4:10:13week six here was the highest sales, so
4:10:15that must be the answer. But in reality,
4:10:17you can see week 14 was higher, week 19
4:10:19was higher. So when you need something
4:10:21that has actual full context, then you
4:10:25can't do the vector database chunking.
4:10:27That's where you'd rather just have a
4:10:28markdown file of March 5th and then all
4:10:31this agent would have to do is read that
4:10:32entire markdown file and then give you a
4:10:35summary and that's just going to be more
4:10:37accurate. So in this project, if we open
4:10:39up level three, you can see it's very
4:10:40similar because you can still have
4:10:41context files, decision files, you can
4:10:44still have all that. And then you might
4:10:45identify, okay, actually this one
4:10:48specific unit of my business, maybe my
4:10:50YouTube transcripts, maybe I want just
4:10:52that to be a vector database, but I
4:10:54still want my context and my projects
4:10:55and my decisions to be markdown files.
4:10:58So another point I'm trying to make here
4:10:59is just because you have a second brain
4:11:01and just because you have a massive, you
4:11:03know, folder here with a bunch of
4:11:04folders and files doesn't mean that the
4:11:06whole folder needs to be one style.
4:11:09Doesn't mean that everything needs graph
4:11:11rag. Doesn't mean that everything is
4:11:12just LLM wiki. It means that you're able
4:11:14to decide based on the type of data and
4:11:16the way you use it, how can you
4:11:18structure this specific folder in the
4:11:20way you want it. So here we have a
4:11:21vector index folder and we click on the
4:11:22how search works. It works by chunking,
4:11:25embedding, search, hybrid, reranking.
4:11:28There's some things you can get really,
4:11:29really nitty-gritty on when it comes to
4:11:30semantic search. But what vector
4:11:32retrieval is really, really good at is
4:11:34looking at tons and tons of data,
4:11:35typically just like a lot of text, and
4:11:37when you need a very specific answer,
4:11:40something that's very similar. So, if
4:11:41you had a thousand rules that you needed
4:11:43to store, and you basically said, "Hey,
4:11:45um, can you remind me what rule 17 was?"
4:11:48That might be a really good use case for
4:11:50vector search because it's able to
4:11:51search for rule 17, pull in those
4:11:53chunks, and just give you a little
4:11:54snippet because it would be a waste of
4:11:56time and tokens for your agent to read
4:11:58the entire markdown file of all a
4:11:59thousand rules if you just needed rule
4:12:0217. So that's kind of the difference
4:12:03there. Like I said, I've got so many
4:12:05videos on vector stuff on my channel,
4:12:07but really you could say, hey to your
4:12:09cloud code agent, I have this data.
4:12:10Here's how I want to use it. Do you
4:12:12think this would be better for now as
4:12:13markdown files, or should I do semantic
4:12:15search? Like what would actually make
4:12:17more sense here? and it will help walk
4:12:18you through the way that you should
4:12:19actually set that up. So now I hope you
4:12:21guys are starting to understand why I
4:12:22said, you know, moving up on or I'm
4:12:25sorry, like moving up on levels, moving
4:12:27down doesn't necessarily mean better.
4:12:29It's all about figuring out what is the
4:12:30pain point with what you're currently
4:12:32doing and where would a different level
4:12:34help you out and fix that pain point.
4:12:36Okay, so now let's take a look at level
4:12:38four. This is where we start to get into
4:12:39like knowledge graphs and relationship
4:12:41graphs, which typically are going to be
4:12:43the most complex and sometimes the most
4:12:45expensive as well. If you're doing it on
4:12:46a certain platform, you could always use
4:12:48open source software. But anyways,
4:12:50knowledge graphs. And I also want to be
4:12:51upfront. I've played with these a lot,
4:12:53but I do not actually use these on the
4:12:55day-to-day because I found out just
4:12:57other ways to use routing files and wiks
4:12:59that fit my needs. Now, my work is very
4:13:02different than what a lot of your guys'
4:13:03work may be. Mine is very project based
4:13:05and it is very, you know, content heavy.
4:13:08I don't have a massive CRM to manage
4:13:10with a bunch of different businesses and
4:13:12clients, you know, and if I did, maybe a
4:13:14knowledge graph would make a lot more
4:13:15sense, and it probably would. But
4:13:16typically, the cool part about that is
4:13:18if you identified that you needed a
4:13:20knowledge graph, let's say for all your
4:13:22projects, you needed you wanted to put
4:13:23all of this in a knowledge graph, the
4:13:25data probably already exists here. And
4:13:27that's the thing about building out
4:13:29these relationships in your knowledge
4:13:31graph is that the system, whatever
4:13:33software you use, is typically going to
4:13:34be pretty good at embedding that and
4:13:36creating that. But the problem that you
4:13:38have to solve is you have to give it
4:13:39enough data. And so one thing that I
4:13:40really like to do is I like to have
4:13:42these brainstorm sessions as you can
4:13:43see. And what I do with these brainstorm
4:13:46sessions is I use a skill called grill
4:13:47me. So if you see here I have a skill
4:13:49called grill me which I originally got
4:13:50from Matt PCO. I customize it a little
4:13:52bit. I'll leave the skill for grill me
4:13:55in my free school community. The link
4:13:56for that is down in the description. All
4:13:57you have to do is hop in here, go to
4:13:59classroom, click on all YouTube
4:14:00resources, and you can find all the
4:14:02skills and everything like that. But the
4:14:04skill, what that does is it basically
4:14:05just grills me. It interviews me
4:14:07relentlessly about a certain topic and
4:14:09it creates a brainstorm file here. It
4:14:11only stops when it knows everything
4:14:12about it. So if you wanted to start
4:14:14building up a knowledge graph for all
4:14:15your clients and businesses, just say,
4:14:16"Grill me about client A, grill me about
4:14:18client B, grill me about business A."
4:14:21And it would just ask you questions and
4:14:22you can feed it files, you can give it
4:14:24stuff, you can feed it in transcripts,
4:14:25you can feed it in, you know, contracts,
4:14:27whatever it is. And that's how you can
4:14:28start to form a lot of data. Hey guys,
4:14:31me again. Real quick, I'm editing this
4:14:32video and I realized that I needed to
4:14:33throw out one thing here, which is that
4:14:36obviously if you're putting all of this
4:14:38data and you're sending it all to
4:14:40Enthropic to Claude models, then that's
4:14:42not [snorts] private. So, if you feel
4:14:44comfortable with that, that's fine. I am
4:14:46putting a lot of my data in there and it
4:14:47is my business stuff and that's what I'm
4:14:50doing. But if you don't feel comfortable
4:14:52with that or you, you know, don't want
4:14:53to send client data, of course you
4:14:54don't, then maybe you want to do that
4:14:56through open source models and maybe
4:14:58cloud code isn't where you have this
4:14:59second brain that has every single piece
4:15:01of information about you and your
4:15:03business and your client's business. So
4:15:04the point I'm trying to make here is
4:15:06just this is what I'm doing. I'm
4:15:07obviously aware of the fact that my data
4:15:10goes to Enthropic when I process it
4:15:12through Claude. And if you guys are
4:15:13doing that, then you should also be
4:15:14aware of that. But there are other
4:15:16options if you can't do that. So I had
4:15:17to throw that out there. I am planning
4:15:18to make a ton of videos here soon about
4:15:21local AI and open source models and all
4:15:22this stuff cuz it's a really really
4:15:24exciting space that I think is going to
4:15:25start becoming bigger and bigger. So
4:15:28yeah, keep that in mind. Back to the
4:15:29video. I think sometimes that's a
4:15:31misconception about how I got here and
4:15:34how people build their own AIOS or
4:15:36second brain is that they think the
4:15:39problem is the system not retrieving it.
4:15:40Great, which sometimes it is, but
4:15:42sometimes it seems like the bigger
4:15:43problem is getting everything out of
4:15:45your brain into the system. So before
4:15:47you blame AI, take a look at your
4:15:50folders and files and say, is this
4:15:52actually holistic? Is this does this
4:15:54have all the nuance that I have in my
4:15:56brain? Anyways, from there, when you
4:15:57open up level 4, you can see that it's
4:15:59it's, you know, very similar still.
4:16:01We're just adding on a few things. You
4:16:02can see here we've added an agent.mmd,
4:16:04which is the exact same as the cloud.MD.
4:16:06And what else is cool is you can
4:16:07literally just reference inside of your
4:16:08cloudmd at agents.mmd. And then you can
4:16:10delete all this because this basically
4:16:12just like injects that file into here.
4:16:14But I just wanted to show that. But
4:16:15anyways, you can see we're still
4:16:17following the same principles. We have a
4:16:19wiki. We've also added a knowledge graph
4:16:21layer. We've still got the same where
4:16:22things live with the routing with all
4:16:24these just regular folders and boring
4:16:25markdown. But boring is beautiful. You
4:16:27can see that our memory is still here.
4:16:29It's starting to grow and we just keep
4:16:31building on top of this. So, one thing
4:16:33we added here as you can see was our
4:16:34knowledge graph folder. And so, what
4:16:36happens here is we get different
4:16:37entities, right? So, like we can see,
4:16:38okay, Jordan is a person, Acme is a
4:16:40company. And then we can start to form
4:16:41relationships between all these things.
4:16:43So Jordan works at Acme. Acme is
4:16:46endorsed by Postpilot. Postpilot is a
4:16:48competitor of Cadently and it starts to
4:16:50build out not only these entities, but
4:16:52it shows you how they're all related.
4:16:54And so that's why when I said that I
4:16:56really like using, you know, this um
4:16:58what's it called? LM Wiki is because I
4:17:00have enough of that feel of all these
4:17:02relationships because I've put so much
4:17:04time and effort into ingesting these in
4:17:07the right way and giving it context. The
4:17:09thing about this one is that it has to
4:17:10read every single file it wants. Maybe
4:17:13it was looking at AI video production
4:17:15and all it needed to know was 11 labs.
4:17:17It still would have read this entire
4:17:19file first. And so that's where
4:17:21sometimes the knowledge graph is
4:17:22actually more lightweight in that sense.
4:17:24And this is the example I showed at the
4:17:26beginning of the video where we have
4:17:27light rag. And forgive me, I'm going to
4:17:29have to blur some of this stuff out cuz
4:17:30this is like legitimately my entire
4:17:31second brain and our business. But as I
4:17:33really zoom in here, and this kind of
4:17:35slows down my computer cuz there's so
4:17:36much. But what you'll notice is that we
4:17:39actually start to get relationships. I
4:17:40probably shouldn't have done this with
4:17:41so much data, but you can see like we
4:17:43have this collaborates with that. We
4:17:46have this builds that. And so if I
4:17:48really started to open up all of these
4:17:50little, you know, circles, we could see
4:17:53what was going on and how they're all
4:17:54related. We can see that our 7-day AIS
4:17:56challenge. It was provided from YouTube.
4:17:59It connects to the onboarding process of
4:18:01AIS Plus. It was developed by Aiden. And
4:18:04so we can basically follow around these
4:18:06relationships as you see. And even
4:18:08though it's pretty much the same data
4:18:09that you see here in Obsidian, we're not
4:18:10getting that same level of relationships
4:18:12between these different entities. So
4:18:13anyways, if you guys want to see, you
4:18:15know, a full breakdown video on
4:18:16something like light rag or um graphify
4:18:18or all the other solutions that there
4:18:20are out there for more of a knowledge
4:18:21graph, relationship graph, then let me
4:18:23know. But that is kind of the difference
4:18:24there. So if you don't need those sort
4:18:26of relationship chains and you're not
4:18:28worried about that semantic type of
4:18:30relationships, then you probably don't
4:18:31need to use something like a knowledge
4:18:33graph. And then level five, we have more
4:18:35of the always on brain OS and something
4:18:38like Gbrain. Gary Tan, CEO of Y
4:18:40Combinator, he created this thing called
4:18:42GBrain, which pairs really well with
4:18:44GStack. But GBrain is kind of the idea
4:18:46of everything we've talked about here,
4:18:48wikis, routing, relationships, tools.
4:18:50But GBrain has kind of that always
4:18:53element because it is like constantly
4:18:55syncing and refreshing memories and
4:18:57adding more stuff. So adding in Gbrain
4:18:59to something like a Hermes agent would
4:19:00be really, really good. You could still
4:19:02do it in cloud code, but you'd have to
4:19:03handle those crons and get all that
4:19:04stuff set up, which is why I don't
4:19:06currently run GBandra at the moment, but
4:19:08I have been playing around with it with
4:19:09my Hermes agent. So, anyways, the point
4:19:10here is that it's very similar to
4:19:12everything else we've just talked about.
4:19:14It's just having that auto updating
4:19:16feel, more of the autonomous always on
4:19:18feel. But I will say another thing that
4:19:20I kind that kind of scares me about that
4:19:22is you have this whole dilemma of, you
4:19:25know, when do you have too much context
4:19:27and when does it get to the point where
4:19:29it's actually doing more damage than
4:19:30it's doing good? And the reason I bring
4:19:32that up is because I am in complete
4:19:34control of what my second brain ingests.
4:19:36I will run a skill to go grab all of my
4:19:38meeting transfers from the week. I will
4:19:40say, "Hey, here's something. Help me
4:19:42figure out like help me brainstorm about
4:19:43this and then let's ingest it together."
4:19:44And for me, I really like being in that
4:19:46control because in my mind, there's a
4:19:48big difference between a few types of
4:19:50data. If you guys remember in my like
4:19:51AIOS videos, I've talked about the four
4:19:53C's. So, context, connections,
4:19:54capabilities, and cadence. And for the
4:19:56second brain, I mainly think about it as
4:19:58just these first two. So, context and
4:20:00connections. And so, when I think of
4:20:01context, that's stuff like, you know,
4:20:04what my business has done. So, if I come
4:20:05into here into my my second brain, and
4:20:07you can see here, if I go to um up at
4:20:10OTAAS, so OTAAS are basically just our
4:20:12projects for the quarter. And so here I
4:20:14can see all the Q1 ones, right? I can
4:20:15look at all those and I can click at
4:20:16them and see decisions that we've made
4:20:18in the statuses. And I can also see Q2
4:20:20OTAAS. So I can see what's going on
4:20:21here. And my second brain is able to see
4:20:23that because that has been basically
4:20:25those are locked in decisions. This is
4:20:26what we're doing this quarter. And then
4:20:28I'm updating the statuses of that stuff.
4:20:29So that's like context. That's what's
4:20:31going on in the business. But when it
4:20:32comes to connections, if I go back to
4:20:34this, this is more of like the real data
4:20:36that isn't as evergreen. This is stuff
4:20:38that changes. This is like Slack
4:20:40threads. This is emails. this is
4:20:42customer data and that type of data you
4:20:44don't want to ingest into a second brain
4:20:46because that's just noise then then you
4:20:48have to go back every month and like
4:20:49delete old stuff. So the way that I like
4:20:51to think about my actual second brain is
4:20:53stuff that I'm not going to delete. This
4:20:55is stuff that is like okay in a year
4:20:57will it be good for me to have this
4:20:58memory in here? Yes. Otherwise it's just
4:21:00adding noise. So when you're adding data
4:21:03into your project think about it like
4:21:05the context and connections. think about
4:21:07if this is kind of like more evergreen
4:21:08holistic data or if this is more things
4:21:10that are going to change next week. So,
4:21:12you probably shouldn't pull it in, but
4:21:14you should make sure that your second
4:21:15brain has access to go grab it. So, that
4:21:17way if I said to my second brain, "Hey,
4:21:20can you just take a look real quick at
4:21:21what John and I were talking about last
4:21:23week about, you know, OTAA number seven,
4:21:25it would first go to our OTAA file and
4:21:27it would search through there and it it
4:21:28would try to find it there. If it
4:21:30couldn't find it there, it would look
4:21:31through the wiki and it would look
4:21:32through meeting transcripts and see what
4:21:33we talked about there. And if it
4:21:34couldn't find it there, it would finally
4:21:36go to ClickUp itself, pull real data in
4:21:38from me and John's conversations and see
4:21:40if the answer lived there. And so that
4:21:41in my mind is still a second brain
4:21:43because I'm able to ask a vague question
4:21:45and the second brain knows exactly where
4:21:47to look in what order to find that
4:21:48real-time data and then give me back the
4:21:50answer that I need. That's the question
4:21:51I ask myself is does this thing
4:21:53understand where my data lives and where
4:21:55to look and can it give me accurate
4:21:56answers. So as far as finding your
4:21:58level, remember your whole project
4:22:00doesn't fit into one level. Maybe this
4:22:02folder is level two. Maybe this folder
4:22:03is level four. Maybe this folder is
4:22:04level three. Here are some things to
4:22:06think about. If you are reexplaining
4:22:07your setup and you need to find things
4:22:09by exact words or files, look at level
4:22:11one. If you have 30 plus notes and you
4:22:12keep forgetting what's in them, look at
4:22:14level two. That's where you sort of like
4:22:15ingest them and get that wiki with
4:22:17relationships. If your project is just
4:22:19completely whiffing on notes that you
4:22:20know exist and your routing isn't
4:22:22working, then maybe you want to look for
4:22:23something more like a semantic search
4:22:25that doesn't rely on an exact word level
4:22:27match. If you're looking for
4:22:29relationships and to be able to follow
4:22:30chains of questions and thoughts, then
4:22:32you probably want to look for something
4:22:33like a knowledge graph. And if you're
4:22:34running agents offline and you've got so
4:22:36much data and you want to sync up a
4:22:37bunch of Hermes agents together, then
4:22:39you probably are looking for something
4:22:40like Level Five, something like Gbrain.
4:22:42And another topic that I get some
4:22:43questions about, which I'm not going to
4:22:45fully address in this video, but I will
4:22:46briefly bring up is the fact that you
4:22:49are building your own second brain OS.
4:22:51So are other people on your team. The
4:22:53next question is how do you actually
4:22:55make sure that everyone's data is
4:22:56syncing together and how do you have
4:22:58more of like your team second brain?
4:23:00There's a lot of different ways to solve
4:23:01that. I think once again it's not an
4:23:03issue of oh do we use Google Drive or
4:23:05notion or GitHub or cloud plugins. I
4:23:08think the issue to figure out with your
4:23:09team is how do we actually make sure
4:23:11that we all habit shift so that this
4:23:13stuff is actually useful and not just
4:23:15noise. How do we make sure that process
4:23:17owners are updating their docs and
4:23:19syncing their stuff there? How do we
4:23:20make sure that other people are pulling
4:23:22from that rather than always just
4:23:24pinging the same people for questions
4:23:25and answers all the time? I think the
4:23:27adoption and the change management
4:23:28question is the bigger one. The tech and
4:23:30the way it actually functionally rolls
4:23:32out is a little bit less. But what I do
4:23:34know is that you getting set up with
4:23:36your own first and understanding how it
4:23:38works, how you should route, how you
4:23:40should make the decisions of where the
4:23:41data should live, that's the first
4:23:43hurdle. You can only solve the teamwide
4:23:45problem once you feel comfortable about
4:23:46the way you run it every single day and
4:23:48that it works for you. All right, sweet.
4:23:51So, you guys have heard me earlier in
4:23:52this course talk about one of my skills
4:23:53called the grill me, which was inspired
4:23:55by Matt PCO's grill me skill as well.
4:23:58So, now that you're thinking about
4:24:00building out your second brain, I want
4:24:02to talk about this grill me skill that
4:24:04really helps me get more context into my
4:24:07systems, but also helps grill me on
4:24:10before I'm going to build an automation
4:24:11or before I'm going to build a new
4:24:13skill. Whenever I want to extract what's
4:24:14in my brain and get it into my second
4:24:16brain and get it into my AIOS, I like to
4:24:19use this grill me skill. So, let's talk
4:24:20about that real quick. The toughest part
4:24:22about building good skills and building
4:24:24a good operating system is trying to get
4:24:26everything from your brain into your
4:24:28system. So, for example, what you're
4:24:30looking at here is after months and
4:24:32months of me building up all of the
4:24:33knowledge that lives inside of my AIOS.
4:24:35It's basically just the idea that if
4:24:37everyone's using the same model, so if
4:24:38everyone's using Claude Opus 4.8,
4:24:40then everyone's going to be using the
4:24:42same prompts and getting the same output
4:24:44because the model is fundamentally the
4:24:45same for everybody. So what really makes
4:24:46the difference is when you add context
4:24:48into that model and you give it your
4:24:49taste, your voice, your decisions, and
4:24:51that's how you get outputs that actually
4:24:53sound like you. But once again, the real
4:24:54challenge is still the extraction,
4:24:56getting everything from your head into
4:24:58the AI system so that your skills can
4:25:01use it and that your context is better.
4:25:02And if you guys have been following me
4:25:03for a while and you've seen videos I've
4:25:05made about like discovery calls and and
4:25:07scoping out projects, that's the
4:25:09toughest part is especially if you're
4:25:11working with a client, asking them so
4:25:13many questions about this process to the
4:25:14point where they might even get annoyed
4:25:16because you're asking so many questions.
4:25:17But that's just what you have to do.
4:25:19It's the difference between a system
4:25:20that is successful 95% of the time and
4:25:22one that's only successful 80% of the
4:25:24time. So this one skill we're going to
4:25:25look at today is called grill me. It
4:25:27basically takes what's in your head into
4:25:29reusable context for your AI. So what
4:25:31happens is all of that knowledge that's
4:25:33in your head that you might think,
4:25:34"Okay, I'm just going to brain dump into
4:25:36clawed code for 5 minutes and it will be
4:25:37good enough, it's not ever good enough."
4:25:39So what this does is it basically
4:25:41relentlessly asks you questions. It
4:25:43grills you until it knows pretty much
4:25:45everything about the process. It'll ask
4:25:47you a question, you answer it, and then
4:25:49it basically will checkpoint and it will
4:25:50write everything back to a knowledge
4:25:52dock and it will just keep going over
4:25:54this loop endlessly until the knowledge
4:25:56dock is good enough and there's no gaps
4:25:58or holes in that knowledge. And so, like
4:26:00I said, this results to better skills,
4:26:01better context, and better projects. And
4:26:04originally, this skill was built by Matt
4:26:05PCO. And what's cool is if you look at
4:26:07it, it is a super simple prompt. It's
4:26:09like four to five sentences. Interview
4:26:11me relentlessly about every aspect of
4:26:13this plan until we reach a shared
4:26:15understanding. Walk down each branch of
4:26:17the design tree, resolving dependencies
4:26:19between decisions one by one. For each
4:26:21question, provide your recommended
4:26:22answer. Ask questions one at a time. If
4:26:24a question can be answered by exploring
4:26:26the codebase, explore the codebase
4:26:27instead. And I like to look at that
4:26:28because it makes you realize that a
4:26:30skill doesn't have to be super
4:26:31complicated automation. A skill can just
4:26:33be a prompt that you don't want to have
4:26:35to say every single time. And of course,
4:26:36naturally, what did I do? I destroyed
4:26:38that. I ruined the skill. I made it a
4:26:40little bit more complex, but I added
4:26:42something that I think makes it much
4:26:43better. So, if I go into mycloud, I go
4:26:45down to my skills and we look for the
4:26:46grill me right here and I open up the
4:26:48skill.md. You can see it's a little bit
4:26:50longer now. But basically what I did is
4:26:52I worked in that whole element of
4:26:54checkpointing after every single
4:26:55question because originally the skill
4:26:57doesn't do that and what happens is if
4:26:59you are talking you know if it's
4:27:01grilling you for an hour plus which
4:27:02sometimes it will and that's a good
4:27:03thing and as the condex window starts to
4:27:05fill up I started to get worried that it
4:27:07was going to misremember some of my
4:27:09answers from earlier. So I just found
4:27:10myself telling it manually hey write
4:27:12this to a doc write this to a doc
4:27:14checkpoint every time. And so I figured
4:27:15okay why not just work that into the
4:27:17skill. So now what the skill does is it
4:27:19creates a folder called brainstorms and
4:27:21it does this at the root of your
4:27:22project. So if I go down here, you can
4:27:24see I've got a brainstorm file or sorry
4:27:26a brainstorm folder right here with
4:27:27these four brainstorms. And so it will
4:27:29create that for you if you don't have
4:27:30it. But if you do have it, it will just
4:27:32chuck a doc in there, a markdown file
4:27:34right away. And so then if I open up
4:27:36like for example this packaging one
4:27:37which I was doing, it will find like the
4:27:40algorithm, the key decisions, but then
4:27:42it will also show you the step-by-step
4:27:44Q&A log of the questions that it asked
4:27:46and what I answered with and the key
4:27:47highlights. And then as soon as we
4:27:49finally got to the end of that packaging
4:27:50grill me session, it said, "Hey, I
4:27:52noticed you have this packaging guide
4:27:54and you have a packaging skill and
4:27:56there's a lot of nuance here that we
4:27:57talked about that's not in there. So do
4:27:59you want me to update both of those?"
4:28:00And then I said yes. And now those
4:28:02skills and docs are so much better. I
4:28:04also did one where I said, "Hey, I want
4:28:05you to understand everything about the
4:28:07business." And we walked through from
4:28:08beginning to end, all the decisions, all
4:28:10the processes, and now my OS just feels
4:28:12like it knows even more about the way
4:28:14the business works. And so, if you think
4:28:16about it like this, right, like nothing
4:28:17is going to be perfect on the first try.
4:28:19And so, let me just do a quick
4:28:21visualization. This is kind of the old
4:28:23way when you're building a skill, right?
4:28:24So, we've got iterations down here.
4:28:26Let's say by iteration one, after you've
4:28:28knowledge dumped in your brain and you
4:28:31want to build a skill, you maybe get
4:28:32somewhere. or let's just say around here
4:28:34where you're about like 70% successful
4:28:36on iteration one and then what happens
4:28:38is you run the skill and you make a
4:28:40small improvement and now you're up
4:28:41about I don't know right here like you
4:28:43go up from 70% to 75 and then every time
4:28:46you iterate you get a little bit better
4:28:48with each iteration until maybe you cap
4:28:50at the point where you're about like 95%
4:28:53good and this could be 10 iterations it
4:28:55could be 30 iterations it however many
4:28:56it takes for your skill to feel a bit
4:28:58more battle tested and honestly I don't
4:29:00think you ever get to 100% because as
4:29:02your business evolves and as you evolve,
4:29:04the skill keeps evolving. So like all my
4:29:06skills that I've been using for months
4:29:07and months, I'm pretty much still
4:29:10changing a lot. But the whole idea is
4:29:12what if on iteration one because you do
4:29:14this grill me and you spend extra time
4:29:16up front, you're able to jump right up
4:29:18here to like 90 at the beginning. And
4:29:20yes, it's not perfect. You're still
4:29:21going to iterate a little bit, but
4:29:23you're just there a lot quicker, which
4:29:24gives you more opportunity to find
4:29:27better ways to iterate on it. So that's
4:29:28my horrible visual of why I think this
4:29:31is valuable. It just goes back to that
4:29:32whole idea of if I had six hours to chop
4:29:34down a tree, I would spend the first
4:29:36four sharpening the axe where upfront,
4:29:38yes, maybe it feels boring or
4:29:40repetitive, but that's what you need to
4:29:42do is get all that context in there
4:29:43because it helps downstream so much
4:29:45more. So anyways, if you guys want to
4:29:47grab the grill me skill, you can look it
4:29:48up here from Matt PCO. Or if you want my
4:29:50version, you can come to my free school
4:29:52community. The link for that is down in
4:29:53the description. Just join the
4:29:54community, go to the classroom, click on
4:29:56all YouTube resources, and it will be
4:29:57right in there for you to find along
4:29:59with all my other free resources. And
4:30:00then it's as simple as saying, "Hey,
4:30:02grill me about this." Or, of course, you
4:30:03can invoke it with a slash command.
4:30:05Right there, you can see, "Grill me."
4:30:07But I could just say something as simple
4:30:08as, "Hey, I need you to grill me about
4:30:10the way that I think about applying AI
4:30:12to my own business internally in a safe
4:30:15way that won't damage the business." You
4:30:18can see it'll obviously load up that
4:30:19grill me skill. We're going to see in a
4:30:21second that it's going to create the
4:30:22capture file so nothing gets lost. Right
4:30:24there we have applying AI internally.
4:30:26And this is what it looks like. It's
4:30:28going to set up the discovery notes, the
4:30:30summary key decisions, um, Q&A log, and
4:30:32any open flags. And what's cool about
4:30:34this is it'll flag things that you need
4:30:36to go find. So, when I was running
4:30:38through this funnel map, you know, there
4:30:39were some things going on in the
4:30:40business that I don't actually know
4:30:42super well. Like, I can't explain the
4:30:44same way as the actual stakeholder or
4:30:45operator that does that process can
4:30:47explain it. So, it said, "Hey, here's
4:30:49some things to flag. Go reach out to
4:30:50this person and have them send you
4:30:52information and then come back and drop
4:30:54that into me and then we'll update this
4:30:56brainstorm." So that's basically how it
4:30:57works. It might ask you five questions.
4:30:59It might ask you 30. It's just going to
4:31:01go until you guys feel like you have the
4:31:02same shared knowledge and that it is a
4:31:04good stopping point. And the cool thing
4:31:06is because these are saved as docs. You
4:31:08can reference them later. But you could
4:31:09also come back to like for example
4:31:10packaging. Let's say I I find a major
4:31:13breakthrough in the way that I package
4:31:14my content. I would just come back to
4:31:16this doc and say, "Hey, grill me again.
4:31:19Here's some new things I found. Let's
4:31:20update all this information." All right.
4:31:22Let's talk about agent teams. What do I
4:31:24mean by this? I basically just mean how
4:31:26can you just leverage
4:31:28a bunch of cloud agents running at the
4:31:30same time creating a team and typically
4:31:33when I want to go for an agent team I've
4:31:35got one reason and that is because I
4:31:38want multiple perspectives. I want
4:31:41multiple different types of agents that
4:31:43have specialties in different areas to
4:31:45collaborate on something together. Now
4:31:47there is one important thing to
4:31:48understand about agent teams. There is a
4:31:51big difference between sub agents and
4:31:54agent teams. Take a look at this visual.
4:31:56On the lefth hand side, we have sub
4:31:57aents. On the right hand side, we have
4:31:58agent teams. Now, we've already talked
4:32:00about sub aents. You guys understand
4:32:01this. The main agent orchestrates sub
4:32:03aents with fresh context windows. And
4:32:05the sub aents cannot talk to each other.
4:32:07They can only talk back to the main
4:32:09agent. They can only report to the main
4:32:11agent. But in agent teams, look at this.
4:32:13The main agent spins up the team and
4:32:15there's a shared task list. And now all
4:32:18of these specialized teammates can
4:32:20communicate between each other and they
4:32:22can claim tasks off the list and they
4:32:24literally can work together and talk to
4:32:26each other and report back to the main
4:32:28agent. So it's just a lot more of a team
4:32:30rather than like individual workers in
4:32:33cubicles that don't talk. These are
4:32:35actual, you know, people sitting around
4:32:36one desk having a meeting talking and
4:32:38getting work done. So it's very cool,
4:32:40but it is a little bit more token
4:32:42intensive. So I don't use them all the
4:32:45time. Like I said, it's a very specific
4:32:46scenario when I want to have a bunch of
4:32:48different perspectives collaborating
4:32:50together in real time. Now, if you want
4:32:52to do stuff in parallel, you're probably
4:32:53better off just spinning up a dynamic
4:32:56workflow. So, a dynamic workflow is
4:32:57really cool because what it does is it
4:32:58takes the main session and it basically
4:33:02creates these dynamic workflows of
4:33:04agents that will do different phases. So
4:33:06maybe it'll spin up five for the
4:33:07planning phase and then maybe it'll spin
4:33:09up like 10 for the next phase and it
4:33:10will just dynamically spin up more
4:33:12phases of sub aents based on the results
4:33:15of the previous session. So it is very
4:33:18cool. It's it's also quite token
4:33:19intensive, but it's a great way to do
4:33:21like verification loops and to make sure
4:33:23that everything is working as it should.
4:33:25So um I do have videos on my channel
4:33:26about dynamic workflows. I'll tag one
4:33:28right up here if you want to check out
4:33:29like kind of I break down how it works.
4:33:31But those are all sub agents still. Now,
4:33:33what if we want agent teams that
4:33:35literally, as we know, collaborate with
4:33:37each other, talk to each other, and all
4:33:39report back to the main session, and
4:33:41maybe they have some sort of shared task
4:33:42list together. I mainly have one key use
4:33:46case when I make agent teams because
4:33:48typically I can I find that things can
4:33:49be done in parallel. But let me talk
4:33:52about my use case here. So, I like to do
4:33:56councils or war rooms or, you know,
4:33:58something what I call a roast. The idea
4:34:00is whenever I'm having a big decision or
4:34:03whenever I'm trying to analyze if
4:34:04something is going to resonate with the
4:34:06audience or you know if there's any
4:34:07holes in my plan, I like to spin up a
4:34:09debate team of different agents that
4:34:11literally talk to each other and debate
4:34:13and they go back and forth and chat and
4:34:15chat and chat until they reach some sort
4:34:17of consensus. And I just think that it's
4:34:19not only pretty fun, but it gives me
4:34:20different angles and then I'm able to
4:34:22read the debate and I'm able to read
4:34:24what did different perspectives bring
4:34:25up. Now, before we actually do that, I
4:34:28have to show you how you set up agent
4:34:29teams because if you go to the
4:34:30documentation, you can see that this is
4:34:32an experimental feature that is disabled
4:34:34by default, at least at the time of
4:34:35making this video. So, you have to
4:34:37enable them by adding a config to your
4:34:39settings. So, look at this. All you have
4:34:41to do is you have to add this
4:34:43environment variable cloud code
4:34:44experimental agent teams equals 1. So,
4:34:46what I'm going to do is copy this. I'm
4:34:48going to go into our knowledge work
4:34:49project that we've been working on
4:34:50together. You can see in here if I go to
4:34:52the files and I go to my.claude. We
4:34:54don't yet have a settings.json file. So
4:34:57we're going to have to create that. But
4:34:58here's what I'm going to say. Hey
4:34:59Claude, I want to test out agent teams
4:35:01which is currently an experient or
4:35:03experiential an experimental feature
4:35:05from cloud code. So take a look at this
4:35:08environment variable. And I'm just going
4:35:10to paste that in. The documentation told
4:35:13me that we have to put this inside of
4:35:14our local settings, the JSON file for
4:35:17this project if we want to use agent
4:35:19teams in this project. So inside of my
4:35:20do.claude create the settings file and
4:35:23add this to the config so that we can
4:35:25actually use agent teams. Okay. So I'm
4:35:27going to shoot that off and hopefully
4:35:29it's able to understand that request.
4:35:30It's able to create us that settings
4:35:32file and then we'll be up and running
4:35:33with agent teams. You can see it's even
4:35:35running a skill right here that's called
4:35:37update config. Okay, awesome. So it
4:35:39created that file. If we go in here and
4:35:41we just verify that in thecloud we have
4:35:43a settings.local.json and there it is
4:35:45right there. Perfect. Now, what we
4:35:47probably need to do is reset the
4:35:49session. So, I'm going to clear that
4:35:50out. A lot of times when you add a new
4:35:51config, you have to start a new session
4:35:53in order to be able to have those
4:35:54changes actually take effect. So, let's
4:35:58see what we can do real quick. I pulled
4:35:59up this report. It's a 2026 CEO study by
4:36:02IBM. They surveyed a bunch of CEOs and
4:36:04asked about some AI stuff, which I've
4:36:05broken this down in a few different
4:36:07videos. It's super super interesting.
4:36:08So, if you want to check that out, just
4:36:10Google this 2026 CEO study IBM and give
4:36:14it a read. Super interesting. Anyways,
4:36:16what I did though is I downloaded this
4:36:18PDF. So, I'm going to just pull up my
4:36:19downloads real quick. And what I like to
4:36:22do is I like to just copy the path to
4:36:24the file that I want my agents to look
4:36:27at because it can obviously access your
4:36:28downloads. So, I copied the path. I'm
4:36:30going to go back into Claude, paste that
4:36:31in there. Now, I'm going to say the
4:36:33following.
4:36:35All right. So, I just read this report
4:36:36and I think it's really interesting, but
4:36:38I'm having trouble understanding how to
4:36:40apply the insights to my day-to-day. So,
4:36:42what I would like you to do is I would
4:36:44like you to create an agent team using
4:36:46the team create function to create an
4:36:48agent team. I want this agent team to
4:36:50have different personas. So, I'm
4:36:51thinking maybe like a small business
4:36:53owner, a large enterprise business
4:36:56owner, a CEO, and
4:37:00maybe a entry-level employee. And then
4:37:02any other personas that you think would
4:37:04be interesting to have inside of this
4:37:05discussion because I want them to debate
4:37:08about this report. I want them to
4:37:09analyze what stats did they think were
4:37:11the most interesting and what was the
4:37:12most concerning and I want them to go on
4:37:15a couple rounds of debates so that they
4:37:16can basically give me an analysis of how
4:37:19I should use it for me and my business.
4:37:21Now I know you don't know too much about
4:37:23me at the moment. So go ahead and look
4:37:25through my Herk 2 project so that you
4:37:27can see who I am and what my business
4:37:29does and then make sure all of the
4:37:31insights and the debate are tailored
4:37:33towards me personally for me and my
4:37:34business. Okay. So, the reason why I
4:37:37said that so specifically up at the
4:37:39beginning where I said use the team
4:37:40create function in order to create an
4:37:42agent team is because sometimes if you
4:37:43say an agent team, it might just spin up
4:37:45a bunch of sub aents. So, you have to be
4:37:46specific about agent teams that can
4:37:48actually talk to each other. And the
4:37:49good news is you'll actually be able to
4:37:51see when it does that. Okay, so one
4:37:54thing I wanted to call out is that
4:37:55agents don't like to read PDFs and like
4:37:57HTML as much because there's a lot of
4:37:58metadata. So, what it did here is it
4:38:00converted that PDF to text. If you see
4:38:03right here, it converted it to text and
4:38:04then it read the text. So anyways, you
4:38:07can see here it says, "Before I spin up
4:38:08the team, let me stage a shared brief."
4:38:10So all personas debate the same facts
4:38:13and they all tailor to you specifically.
4:38:14So it's going to write up that brief and
4:38:16then it will spin up the agent team and
4:38:17I will point that out to you guys. Okay,
4:38:19here we go. So it is launching six
4:38:21distinct personas round one and it's
4:38:23going to run all six in parallel. So
4:38:25that word kind of scares me because I'm
4:38:27not sure if it fully understood that we
4:38:29want the team. Let's see what it does
4:38:31here. Now, as you can see these
4:38:32different personas being spun up. What's
4:38:33cool is you can click into them and you
4:38:35can see the actual prompt that was shot
4:38:37off. So, this is the prompt that was
4:38:38sent off to this specific agent, but it
4:38:42looks like this might not be a team. It
4:38:44looks like it's just spinning up six sub
4:38:46aents. So, I'm going to stop this real
4:38:47quick and just verify that it, you know,
4:38:49it understands me correctly. And that is
4:38:51something that you should do. Watch your
4:38:52agents as they're working and they're
4:38:54building things. And if it feels like
4:38:55they're going off the track that you're
4:38:57trying to put them on, then stop them
4:38:59and explain. So, it looks like you're
4:39:01creating just sub agents and you're just
4:39:03running those in parallel, but they
4:39:05can't actually talk to each other. And I
4:39:07wanted you to use the agent team
4:39:08function, team create function. So, you
4:39:11can actually create an agent team that
4:39:13can all talk to each other together. So,
4:39:14that's super important. Please make sure
4:39:16that you're doing it that way. Okay. So,
4:39:18I was correct. It was doing it wrong and
4:39:20it was able to do a little more research
4:39:21and correct itself. Now, quick quiz
4:39:23question. What would you do now? If you
4:39:27said build a skill around agent teams or
4:39:30if you said something like put this in
4:39:31the cloudmd, then I would say you're
4:39:33right. Whenever it misunderstands you,
4:39:35correct it and tell it to update
4:39:37instructions, update skills or create a
4:39:39skill so that that misunderstanding
4:39:40doesn't ever happen again. Remember? So
4:39:43that's what you should do after this
4:39:44message or even start up a new session
4:39:46and say, "Hey, here's what happened and
4:39:47here's what you did and we need to make
4:39:48sure that it doesn't happen again." So
4:39:50anyways, let's see what it's doing now.
4:39:52It has these six different personas that
4:39:54are being spun up and these are personas
4:39:57that are going to be able to actually
4:39:58talk to each other. You can see that
4:39:59it's sending off different messages to
4:40:01these different personas. So anyways,
4:40:03now that we know that they're able to
4:40:04actually talk to each other and debate
4:40:05the way we wanted them to, I'll check in
4:40:07with you guys when this is completely
4:40:09done. Now, the other thing to remember
4:40:10real quick about this is that this isn't
4:40:12going to fill up too too much of our
4:40:14actual context window, but this will eat
4:40:16at our 5-h hour limit because all of
4:40:17these agents, you know, they're on their
4:40:19own context windows and they're eating
4:40:20tokens and they're talking and agent
4:40:22teams are expensive. But just wanted to
4:40:23call out once again that these agents
4:40:26are talking not in our main sessions
4:40:28context window. All right, so that
4:40:29finished up. I'm not going to read this
4:40:31entire thing and every single round of
4:40:32debates because that would be boring and
4:40:33take forever. Let's just go over some of
4:40:35the key facts. So, six personas, three
4:40:37rounds, real cross talk. We had a small
4:40:40business owner, we had a chief AI
4:40:41officer, we had founder, CEO, we had a
4:40:43bunch of different roles, right? So, as
4:40:45far as the stats, each of the different
4:40:46personas pulled out one that was the
4:40:47most interesting and one that was the
4:40:49most concerning. We saw the team's
4:40:51collective verdict after they argued it
4:40:53was that the most interesting stat was
4:40:55the reality gap. 10% today versus 72% by
4:40:582030. It's the one honest number in the
4:41:00deck and it's your product market fit
4:41:02written in a footnote. The most
4:41:04concerning stat was the 25% use versus
4:41:0686% skills adoption gap. It's your own
4:41:09risk. The same disease as your 20% AI
4:41:11plus churn and your single most sellable
4:41:13lesson all at once. And the most
4:41:14overrated was the plus 17% headline. So
4:41:16anyways, they had to agree on this
4:41:18stuff. They had to all come to a
4:41:19consensus on these stats. And now it
4:41:22gave us basically actionable steps
4:41:24because it knows me. So audit your own
4:41:26agents before you sell a cert about
4:41:27auditing agents. Turn the Nate to John
4:41:29handoff into a one-page decision rights
4:41:31map. fix 20% churn, ship a free AI
4:41:34adoption at audit, and adopt Gordon's
4:41:35credibility guardrail on every citation.
4:41:37We could dig into this, right? And we
4:41:39can have a full report written up, but
4:41:41the point I'm trying to make here is
4:41:43getting different personas is really
4:41:44nice because you have subject matter
4:41:46expertise in one or maybe a few areas,
4:41:48but you don't you can't see everything.
4:41:50You can't see around every corner
4:41:51because of your experience and because
4:41:52of your knowledge, but other people can.
4:41:55And Stanford actually proved this. They
4:41:57have a research method called the storm
4:41:58method where they have different
4:41:59personas attack the angle and find the
4:42:02holes in it and then it has been proven
4:42:04that it's much better research and much
4:42:06more thorough research. So, I did a
4:42:08video breaking that one down as well.
4:42:09It's called the storm research method
4:42:11and I'll tag that video right up there
4:42:12if you want to check it out for next
4:42:14time you're doing research. But just
4:42:15thinking about the idea of using
4:42:17different agents as individual
4:42:19specialized experts that can help you
4:42:21create better plans, create better
4:42:23products, create, you know, have better
4:42:24ideas. What's really cool about this,
4:42:26this wasn't planned, but the next thing
4:42:27I wanted to talk about on here was
4:42:29artifacts. So, it's cool because at the
4:42:32end of this message, Claude says, "Do
4:42:33you want me to package this into a
4:42:34sharable artifact, a clean one page that
4:42:36you can revisit or send to other
4:42:38people?" So, basically, the way that you
4:42:40used to share information is you would
4:42:42create slide decks or you would create
4:42:44like a little memo or I don't know,
4:42:46however you like to send information to
4:42:48your team, that's what you would do. And
4:42:50a lot of times when you would send those
4:42:51artifacts, they would be static, meaning
4:42:52you would download it, you'd send it
4:42:54over, they would open it, and if you
4:42:55made changes on your end, it would kind
4:42:57of, you know, you'd have to send them an
4:42:58updated version. But Claude recently
4:43:00dropped these things called artifacts
4:43:01where it's able to package everything up
4:43:03into a pretty HTML style and then it
4:43:05will just give it to you on an actual
4:43:06URL, meaning it's live on the web and
4:43:08you don't have to host it. It's
4:43:09literally just Claude will host it for
4:43:11you. So that way whenever you have
4:43:12artifacts and whenever you've been
4:43:13brainstorming with Claude for a while,
4:43:15you can say, "Hey, put that into a quick
4:43:16artifact so I can send it to my whole
4:43:18team." And then what's cool is if you
4:43:19keep updating it on your side, as soon
4:43:21as you update the artifact, whoever is
4:43:23watching it, it will update on their
4:43:24side, too. So, it's a really nice thing
4:43:27to be able to utilize, especially if
4:43:28you're working on projects with your
4:43:30teams. I'm not going to spend too much
4:43:31time here because it's a pretty simple
4:43:33concept, but I definitely wanted to
4:43:34bring it up. So, on this lefth hand
4:43:36side, you can see if I click on
4:43:37artifacts, I only have one right now
4:43:38because it's a pretty new feature, but
4:43:40let me open up this artifact to show
4:43:41you. This is a URL. So, if I gave you
4:43:43guys this URL, you could actually see
4:43:45this as well. I guess actually you have
4:43:47to be in my organization to see it, but
4:43:48either way, the point is it's live. It's
4:43:50not a local host. It's a real URL. And
4:43:52so this one I had it spin up to prove to
4:43:54me that the spacing on our book was
4:43:56correct. So it was proving to me against
4:43:58like The Great Gatsby and it was showing
4:43:59me the indentation and the spacing and
4:44:01all that kind of stuff. And this is not
4:44:03itself a really impressive artifact. I
4:44:05just wanted to show you that Claude
4:44:07basically hosts these. You can switch
4:44:08between them. You can rename them. You
4:44:10can see the different versions. You can
4:44:11copy the actual prompt to edit the
4:44:13artifact. And like I said, if I go back
4:44:15into the app, you can see that you will
4:44:17be able to manage and view all of your
4:44:19artifacts right here. Copy the link,
4:44:20send them over to your team. All of
4:44:22these things over here, these are not
4:44:23artifacts. These are just different
4:44:24chats or different projects. These are
4:44:26the artifacts. All right. So, we have
4:44:28our artifact fully built out. As you can
4:44:30see, all I had to say was yes, turn that
4:44:32into an artifact. So, it used all the
4:44:34context and it used all of the rounds of
4:44:36debating to help us build this out. Um,
4:44:38let me just go ahead and open this up on
4:44:40an actual URL. So, here is the actual
4:44:43link. As you guys see, this is something
4:44:44that I could send to my team very
4:44:45easily. I'm going to open that up here.
4:44:47Rewiring the seauite translated for one
4:44:5015 person company. So, this was our
4:44:51agent team debate on this study. We can
4:44:53see the verdict. We can see the stats
4:44:55each seat fought over. We can see the
4:44:57round table with the different personas.
4:44:59We can look at the five moves. So, we
4:45:00could really just dig in here and I
4:45:02could obviously, like I said, if I
4:45:03wanted to show my team what I had worked
4:45:05on today or show them something
4:45:06interesting that I built, I would just
4:45:07send that over. Super super easy, super
4:45:09super quick. Awesome. So, that is
4:45:11artifacts. Very very cool. Let me go
4:45:13ahead and cross this out. We are really
4:45:15making some great progress on this
4:45:17course today. So the next thing you can
4:45:19see down here is routines. And routines
4:45:21are absolutely awesome in the desktop
4:45:23app. It's really, really easy to manage
4:45:25them. I click over here on routines. You
4:45:27can see different routines that I have.
4:45:29And what you'll notice over here is that
4:45:30some are paused. You'll notice that some
4:45:32are cloud. And then you can also spin up
4:45:34ones that are local. And so really the
4:45:36big difference is that the local
4:45:37routines, they have to be running
4:45:40locally. So your cloud code app has to
4:45:42be open and your computer has to be on.
4:45:44But if you put them in the cloud, then
4:45:46all of this can be turned off and
4:45:47they'll still be running on the cloud
4:45:49whenever you want them to. So if you
4:45:51want them triggered on a specific, you
4:45:52know, day and a certain time, it can be
4:45:54that. It can also be triggered on
4:45:56certain actions. Now, the one thing is
4:45:58you are limited on how many cloud
4:45:59routines you can have active. I believe
4:46:01it's like 15 per day can be triggered,
4:46:04but they're super easy to set up. And
4:46:05what's awesome is you get the full
4:46:06agentic loop. So, I'm about to shoot you
4:46:09guys over into a video where I break
4:46:10down these routines and how to set them
4:46:12up. It's completely awesome. Cloud Code
4:46:14has finally brought us routines, which
4:46:16basically means you can inject a prompt
4:46:18into Cloud Code, but it can be running
4:46:19on the web, so your laptop does not have
4:46:21to stay open. And I'm so excited about
4:46:23it. I've already been playing around
4:46:24with it. I've been migrating my
4:46:25automations over there, but there are a
4:46:26lot of little gotchas. So, I'm here to
4:46:28explain exactly how you can actually set
4:46:29up these automations so that they work.
4:46:31So, today, April 14th, Claude tweeted,
4:46:33"Now in research preview, routines and
4:46:35cloud code. You configure a routine
4:46:37once, which is basically like a prompt,
4:46:38and it can run on a schedule, from an
4:46:40API call, or in response to an event,
4:46:42and it runs on Anthropics web
4:46:44infrastructure. So, that's awesome. So,
4:46:46you can call a routine from an API, you
4:46:48can have GitHub events trigger it, or
4:46:50they can be scheduled, which are like
4:46:51the scheduled automations that we
4:46:52already have, but now they run on the
4:46:54web. So, you really can create these
4:46:55from anywhere. You can do it right here
4:46:57as a scheduled trigger to run scheduled
4:46:59remote agents, which is in the terminal.
4:47:01You could also go to cloud.ai/code. So,
4:47:03you could do it on the web. web. And
4:47:04right here, you see I have three web-
4:47:05based routines right here. Or what I'm
4:47:07going to be showing you guys today is
4:47:08just doing it in the desktop app.
4:47:10Because right here, if I go to my
4:47:11scheduled tasks, you can see that I've
4:47:13got some like these four that are local.
4:47:15And then I've got these four that are
4:47:16running inside of a GitHub repository.
4:47:18So these are the remote ones. If I go up
4:47:20here and click on new task, this is
4:47:22where we could set up a new local task
4:47:24or a new remote task. It's very similar.
4:47:26You set up the name, you set up what
4:47:28Claude should do, and this is the actual
4:47:30prompt. So I'll talk more about that in
4:47:31a sec. But then you would configure your
4:47:33model, your repository, and your cloud
4:47:36environment. You set the cadence hourly,
4:47:38daily, weekdays. I think the minimum is
4:47:40once an hour. Like you couldn't go like
4:47:42every 10 minutes or something, but still
4:47:44not bad at all. This is where you could
4:47:45configure all of your connectors. So if
4:47:47you need to connect Slack or Gmail or,
4:47:49you know, whatever it is, you can
4:47:50connect them right here. But you can
4:47:52also just do your regular API endpoints
4:47:54with your API keys. And then of course,
4:47:56you've got your permissions. So you can
4:47:57choose how Claude should be acting. Now,
4:48:00the one thing about these are these are
4:48:02meant to be a oneshot prompt. You're not
4:48:04around. So, you probably want to make
4:48:05sure that it doesn't ever have to stop
4:48:06and ask you questions. Otherwise, what's
4:48:08the point of the automation? So, like I
4:48:10said, there's tons of things to dive
4:48:11into here, and I'm not going to try to
4:48:13bore you guys, but some of this is
4:48:14really important because when I first
4:48:15got this set up, my automations weren't
4:48:18just migrating over and working. So, I'm
4:48:19going to tell you guys the issues that I
4:48:21ran into, and hopefully answer
4:48:22everything that you need to know so that
4:48:24you won't have to go into the comments
4:48:25and ask these common questions. I can
4:48:27just answer them right here for you. So,
4:48:28let me just first of all, real quick,
4:48:29show you guys what I tested out. The
4:48:31first thing I wanted to test out is if I
4:48:33came in here and I created a new routine
4:48:36for just shooting a message to my
4:48:38ClickUp. Obviously, that's not any
4:48:40value, but I just wanted to see how it
4:48:41worked because what I wanted to do is
4:48:42see if I could do this without adding my
4:48:44connector of ClickUp. And I was able to
4:48:47actually get this to fire off, but it
4:48:48didn't work right away. So, let me show
4:48:49you guys what I ran into. So, the way
4:48:51that this works is you need a GitHub
4:48:53repository to sync it to in order for
4:48:55this to actually run. So, it's going to
4:48:56clone my Herk 2 project right here in
4:48:58the web. It's going to be able to read
4:49:00my cloud.mmd. It's going to be able to
4:49:01read my scripts and my skills. And then
4:49:03after it finishes the job, it basically
4:49:05just destroys that little cloud GitHub
4:49:07clone. But as you guys know, you don't
4:49:09push your secrets into GitHub because if
4:49:12you see here, my my Herk 2 project, this
4:49:14is my ENV file with all of my API keys
4:49:17and this is listed in the git ignore,
4:49:18which basically says, hey, when you push
4:49:20to GitHub, you don't include these
4:49:22files. So what that means is in here if
4:49:24this is only looking at your GitHub
4:49:25repo, there's no ENV. So how do you get
4:49:28your API keys into this routine that
4:49:31runs on the web? Well, what you do is
4:49:33inside of this scheduled task, you have
4:49:36a cloud environment. So if I click on
4:49:38this one, you can see this one is called
4:49:39Nate Cloud. So if I open up the
4:49:41settings, what do you see? You have the
4:49:43name of this cloud environment, you have
4:49:44the network access, and you have
4:49:46environment variables. So right here is
4:49:48where I put in my YouTube API key, my
4:49:50ClickUp API key, any of the other API
4:49:52keys that I need to give this cloud
4:49:54environment access to. And then the
4:49:56other thing you have to do is you have
4:49:57to look at the access levels because
4:49:59right here you can see that this one is
4:50:00on full, but by default this will be on
4:50:03trusted, I believe. And that means you
4:50:05can only download packages from verified
4:50:07sources from Anthropic. And when we talk
4:50:09about this later, I'll have a link which
4:50:10you can go see all of them. You could
4:50:12even do custom if you wanted to allow
4:50:13specific domains that aren't on that
4:50:15list. But in order for ClickUp to work
4:50:16in this case, I had to go on full
4:50:18because when I went on trusted, it said,
4:50:19"Hey, we can't actually do that." But
4:50:21when I changed this to full, it let me
4:50:23send a message to my ClickUp. And that
4:50:24is how I got this message right here
4:50:25that says, "Just testing that the remote
4:50:27tasks work and the credentials work." So
4:50:29basically, when these run, whatever you
4:50:31have here as your instructions is what
4:50:33gets prompted. And that's exactly the
4:50:35same way that the scheduled tasks
4:50:36locally work. So right here, you can see
4:50:38I say send a message in the internal
4:50:39ClickUp channel. And right here, the
4:50:41actual thing that it says was send a
4:50:43message in the internal ClickUp channel.
4:50:45So, think of a scheduled task or a
4:50:47routine as you basically typing in a
4:50:49prompt and then someone coming in to
4:50:52your laptop and typing it in for you.
4:50:54So, it's the exact same type of
4:50:55interaction as you talking to cloud
4:50:57code. But that's why once again, you
4:50:59want to make sure it's specific enough
4:51:00so that it can basically oneshot it.
4:51:02Okay. So, let's take it a little deeper.
4:51:04Now, what I tried to do is I did another
4:51:05one which I wanted it to be able to use
4:51:07the YouTube data API in order to grab
4:51:10some YouTube comments for me and give me
4:51:12a little analysis in, you know, ClickUp
4:51:14or whatever. So, this is the prompt I
4:51:16said, right? Analyze 50 of my most
4:51:18recent comments from YouTube and give me
4:51:19a quick bullet rundown. My YouTube API
4:51:21key is available as an environment
4:51:23variable. Use it directly from the
4:51:25environment. Don't look for av because
4:51:28what happens is in your repo, right? So
4:51:30in this Herk 2 project, um, when I
4:51:32normally run this, it grabs all my API
4:51:34keys from thev and maybe it reads the
4:51:36claw.mmd and realizes that's where a lot
4:51:38of those live. So by default, it's maybe
4:51:41going to try to look in thev and it's
4:51:43not going to be smart enough to figure
4:51:44out. And so for ClickUp, it was fine. It
4:51:46figured it out. But for some reason with
4:51:48this YouTube one, it didn't. So I had to
4:51:49explicitly tell it, hey, look in the
4:51:51environment variable rather than in the
4:51:53env. So you can see this first time I
4:51:55ran it 1241, I didn't say that and it
4:51:58couldn't do it. It said like, "Hey, I
4:52:00can't find that. I'm getting an error."
4:52:01And I even tried to tell it here and it
4:52:03still didn't work. But then on this most
4:52:05recent run, when I updated the prompt a
4:52:06little bit, it was able to fetch it
4:52:08right away using the API key. And now I
4:52:11have a remote, you know, routine that
4:52:13would work. Obviously, I need to update
4:52:15this. I'm going to migrate over my other
4:52:17automations, but this was just for
4:52:18testing purposes. And then another one
4:52:20that I do is I have some automations
4:52:23here which basically opens up a browser
4:52:24using Playright CLI and it does some
4:52:27stuff in my school community because
4:52:28there's no publicly accessible API.
4:52:30We've kind of figured out a way to
4:52:32automate it without using browser. I'm
4:52:34not really going to dive into that right
4:52:35now. But what I wanted to tell you guys
4:52:36about that is I tried to basically move
4:52:39over this school wins engagement post or
4:52:41sorry automation into a remote session.
4:52:44So, I copied the exact same prompt that
4:52:46was in my regular scheduled task. And
4:52:48then I just added this little snippet at
4:52:49the end. But what happened is this
4:52:51wasn't working because it basically
4:52:53said, "Hey, you know, like when you do
4:52:54this, it spins up a browser, but there's
4:52:56no cookies because all of this is
4:52:57running remotely and all I have to look
4:52:59at is the GitHub repo. I can't look at
4:53:01the local, you know, cookies that we've
4:53:03used in the last couple sessions of this
4:53:05automation." And so, it doesn't seem
4:53:06like this would work because once again,
4:53:08it has no access to that stuff. So if I
4:53:11wanted to do an automation like this, I
4:53:13would have to use um an endpoint that
4:53:15takes authentication in the form of like
4:53:16actual cookies or a header or you know
4:53:19like an API key because every single one
4:53:21of these runs is going to be stateless
4:53:23and after the run the GitHub clone just
4:53:25gets deleted. Now the exception of that
4:53:27is if the automation is changing
4:53:29something in your codebase or doing a
4:53:30review. If it does do that it will
4:53:32create a new branch for you or it will
4:53:34give you some sort of output and not
4:53:35just delete everything that it just did.
4:53:37But for an automation like this it would
4:53:38just delete it. But hopefully after you
4:53:40guys have seen those examples, you now
4:53:42have the ability to come in and you know
4:53:44make some changes if you need in order
4:53:46to make sure that your automations are
4:53:48running. And what I mean by that is you
4:53:50understand this should be a very
4:53:52specific prompt. This is how you change
4:53:53the model. You have to have a GitHub
4:53:55repo. You can change the settings for
4:53:56your cloud environments right here. You
4:53:59set the schedule. You add any connectors
4:54:00you might need, which would honestly be
4:54:02a little easier if you added just like a
4:54:04Slack connector. And then you can set
4:54:05your permissions here. Now, the other
4:54:07thing to be aware of is you do have
4:54:09limits. So, if I come over here to my
4:54:11settings, you can see if I go to my
4:54:12usage, we have our regular session
4:54:14limits, our model limits, but for
4:54:16additional features, we have daily
4:54:17included routine runs. And I haven't run
4:54:19any yet on the actual schedule. I've
4:54:21just been testing them. Um, but we are
4:54:23at zero for 15. So, I could only have 15
4:54:26automations running with routines per
4:54:29day because I'm on the max $200 a month
4:54:31plan. Your limits would be less if
4:54:33you're on Pro. I think maybe three or
4:54:35maybe five. I'll I have that information
4:54:36later on, but just something to keep in
4:54:39mind. All right, so let's just dive into
4:54:40a little bit more of the details here
4:54:42that may answer some questions you guys
4:54:43have. I think it's pretty clear at this
4:54:45point what it is. Um, I'm going to give
4:54:47you guys this entire doc as well as
4:54:49anything else I've talked about in my
4:54:50free school community. The link for that
4:54:51is down in the description. So, some of
4:54:53the stuff I may not cover. If you want
4:54:55to read more about it, then just go
4:54:56ahead and grab that free resource. So,
4:54:58we know what it is. I think we know how
4:55:00it works, right? Like you define a
4:55:01routine, which is a prompt. You connect
4:55:03a GitHub repo. You could also trigger it
4:55:05by APIs or by a GitHub action and then
4:55:08you can connect your connectors and
4:55:10basically it acts as you talking to your
4:55:12own cloud code. Because of the fact that
4:55:14this is working off of a cloned repo,
4:55:16it's going to read the cloud.MD file
4:55:18automatically every time. So if you have
4:55:21a massive project like a Herk 2 project
4:55:23for example with tons of context and
4:55:24tons of stuff maybe you don't want to
4:55:26put that repo into the cloud to be a
4:55:29routine run because there's a lot of
4:55:32context in that cloudmd and in that
4:55:34whole GitHub repo that might not matter
4:55:36for this automation. So maybe you're
4:55:37better off setting up a specific GitHub
4:55:40repo per scheduled routine. But of
4:55:42course, cloud.md best practices putting
4:55:45in the information that's important
4:55:46because this stuff is going to drain
4:55:48your cloud code session limits the exact
4:55:51same way as it would if you were open up
4:55:54in cloud code just talking to it. So
4:55:56once again, three trigger types,
4:55:57schedule API, which I think is really
4:55:59cool. You could have a different
4:56:00automation make a post request to some
4:56:02sort of routine. And then of course
4:56:04GitHub so you can have it automatically
4:56:06fire off kind of on a web hook based on
4:56:09new PRs, new pushes, new issues, new
4:56:11releases, things like that. So how does
4:56:13this compare to what already exists? We
4:56:15have routines which is the new feature.
4:56:16We have desktop scheduled tasks and then
4:56:18we have something like just a /loop
4:56:20command. So routines run on anthropics
4:56:23cloud and these other two run on your
4:56:25machine. Do you need the machine on? No,
4:56:27for routines that's huge. But for
4:56:29desktop scheduled tasks and for loop,
4:56:31you need your machine on. Do you need a
4:56:32session open? No, that's the same across
4:56:34all three. Do they survive across
4:56:36restarts? The first two do, but loop
4:56:38does not. That has to live within a
4:56:40specific session. Local file access, no,
4:56:42for the routines because it works off of
4:56:44the GitHub repo. And for the next two,
4:56:47yes, you have local file access.
4:56:48Permission prompts with routines, it's
4:56:50fully autonomous. And for these two,
4:56:52they are configurable. And then the
4:56:53minimum interval routines is 1 hour. And
4:56:55these two are both could go every minute
4:56:57if you want. Okay, so let's talk about
4:56:59the environments. Obviously, your ENV is
4:57:01get ignored unless you push it into the
4:57:03GitHub repo. You know, ultimately, if
4:57:05you push it into a private repo, you're
4:57:07probably okay, but you want to be
4:57:08really, really, really careful because
4:57:10then, you know, there's history there
4:57:12and if other people, you know, end up
4:57:14collaborating on it, you just don't want
4:57:15to do that. So, you want to put your API
4:57:17keys in the environment variable like I
4:57:19showed you guys earlier. You want to
4:57:20look at the network access, whether that
4:57:22is full or trusted or none or custom,
4:57:24and potentially some setup scripts. So,
4:57:26that's not something I showed you guys
4:57:27yet. If you're creating a new remote
4:57:28task, you can do a setup script, which
4:57:31is basically just a script that will run
4:57:32when this new session fires up before
4:57:34cloud code launches. So if you need to
4:57:36install any packages or anything like
4:57:37that. Okay, so what's the difference
4:57:39between trusted and full? So trusted
4:57:42only reaches the known vetted services
4:57:44from Enthropic, which I thought I linked
4:57:46right here, but I just linked it there.
4:57:48This basically shows you all of the
4:57:49different domains that are allowed. So
4:57:51right here you can see we've got
4:57:52enthropic services, we've got version
4:57:54control, we've also got some cloud
4:57:56platforms like Google, stuff like this
4:57:58right here. These are the ones that are
4:57:59kind of already verified. So what is the
4:58:01risk of going on full? Well, if claude
4:58:04reads malicious content during a run,
4:58:06then it theoretically could be tricked
4:58:08into sending data to an external server
4:58:10and with trusted that outbound request
4:58:12would get blocked. Now practical risk
4:58:14for private repos where you control the
4:58:15inputs is very low, but I definitely
4:58:18just wanted to at least acknowledge
4:58:19that. So connectors, this is different
4:58:21than just adding your API key. This is
4:58:23more of like the connectors you would
4:58:24add to your actual claude chat or like
4:58:26claude co-workth into like Slack or
4:58:29ClickUp or stuff like that. Here are
4:58:31some security details. I'm not going to
4:58:32go super deep into this. You could also
4:58:34do some more research and download this
4:58:35doc. But of course, there are some
4:58:37things to be thinking about like your
4:58:38API triggers or what's going on with
4:58:40your GitHub repos and the branches
4:58:42because once again, everything is going
4:58:44to be running as you. So if you're not
4:58:46testing out these routines before you
4:58:47just kind of send them off every hour or
4:58:49something, you just have to be thinking
4:58:51about what could happen without
4:58:52permissions and you know stuff like
4:58:53that. Limits and quotas. So it looks
4:58:55like on pro you can have five runs a
4:58:57day. On max you can have 15 runs a day
4:58:59and on team and enterprise you can have
4:59:0125 routines a day. If you hit the cap
4:59:04the orgs with extra usage enabled can
4:59:06exceed it on metered overage. And then
4:59:08we have the minimum scheduled interval
4:59:10which is one hour. And there are also
4:59:12resource limits. So every one of these
4:59:14routines in the cloud runs on four
4:59:16vCPUs, 16 gigs of RAM, and 30 gigs of
4:59:19disk space. So once again, just be
4:59:21thinking about are you putting an
4:59:22absolutely massive GitHub repo up into
4:59:24the cloud right now to run. That could
4:59:26just be wasting resources for no reason.
4:59:28So what persists versus what gets
4:59:30destroyed, the cloud branches gets
4:59:31pushed to your GitHub repo and the
4:59:34session also stays. So as you saw, if I
4:59:36came into here and I looked at all of
4:59:38these tasks, I could see all of the past
4:59:40runs and I could go look at them to see
4:59:41if something's going wrong. but the
4:59:43actual cloud environment that gets
4:59:44cloned will be destroyed. Basically, the
4:59:47rule of thumb here is if something's
4:59:48local or if cloud code can't reach it in
4:59:51your GitHub repo or via an API, then it
4:59:54won't work. We already talked a little
4:59:56bit about writing good prompts, but you
4:59:57definitely want them to be more
4:59:58specific. For example, with my um
5:00:01scheduled automation here, this is much
5:00:03more specific, right? I have a skill
5:00:04that I want to run. I give it the order
5:00:06of operations, but something more like
5:00:08this YouTube comments one, this is not
5:00:10what you'd want to put in there unless
5:00:12you were defining a skill to just let it
5:00:14run because once again, this is supposed
5:00:16to be a oneshot prompt. So, you wanted
5:00:17to make sure it gets it right on the
5:00:18first try. Okay, so why is this so
5:00:20exciting and why does this beat normal
5:00:22automation? Because we are actually
5:00:24keeping the agentic framework. If you if
5:00:27you know when I talk about the WAT
5:00:28framework where we have workflows and
5:00:30agents and tools, when we actually push
5:00:32those automations to the cloud and it's
5:00:34just a you know sort of a Python script,
5:00:36we're losing the agentic piece. We're
5:00:38only sending off really the tools and
5:00:40the workflow. But in this case, we're
5:00:42keeping the WA and the T all running
5:00:44together because the agent is looking at
5:00:47the, you know, cloudm. It's looking at
5:00:48its scripts and it's figuring out what
5:00:50to do. And if it runs in errors midrun,
5:00:52it will selforrect. And if you configure
5:00:54it the right way, it will be able to
5:00:56sort of like leave a memory trail and it
5:00:58can leave like, you know, updates even
5:01:00though each run is stateless. You can
5:01:02still have them kind of continuously get
5:01:04better. And real quick, let's speedrun
5:01:06through these common questions. Do I
5:01:07need to know cron syntax? Nope. You just
5:01:09can schedule a natural language. Super
5:01:11easy. Can it access my local files?
5:01:13Nope. It only gets what's in your GitHub
5:01:14repo or your APIs. What model does it
5:01:17use? You can choose any of the models as
5:01:19you guys saw. Can you watch it work in
5:01:20real time? Yes, you can hit run now and
5:01:22then you can obviously watch it go right
5:01:24there. Same way you would in Claude. You
5:01:26can even talk to it after it's done or
5:01:28interrupt it and then continue going.
5:01:30Can it use my MCP service? Yes, that is
5:01:32what the connectors are. Can teammates
5:01:35use my routines? Nope. These belong to
5:01:37your individual account. You might be
5:01:38able to share those if you're on a team
5:01:40plan, but I haven't actually yet tested
5:01:41that myself. What's the cost? It's just
5:01:43your normal subscription usage. So, keep
5:01:45that in mind. What happens when a run
5:01:47fails? Every one of them will be stored
5:01:49in your history. So you can go see why
5:01:50they failed. You could maybe even have
5:01:52it at the end of every single routine,
5:01:54say, "Hey, if this does fail, just shoot
5:01:55me a Slack message to let me know."
5:01:57Things like that. And can I test a run
5:01:59before going live? Yes, in fact, you
5:02:00should test it multiple times before it
5:02:02goes live. You just go into the routine,
5:02:04you hit run now, and then it will pop up
5:02:06as running. And then you just watch it,
5:02:07you know, watch it go through its order
5:02:09of operations, and you can inject, and
5:02:10you can help it correct itself so that
5:02:13you have confidence that once it shoots
5:02:15off the prompt next time, you won't have
5:02:17to get in the way at all. But anyways,
5:02:18that's going to do it for this one. I
5:02:20hope that those tips and some of the
5:02:21examples that I showed you were helpful
5:02:23and now you can go off and try to
5:02:24migrate your scheduled tasks or any
5:02:26other automations that you've been
5:02:27meaning to build into these web- based
5:02:29routines and not have to keep your
5:02:31hardware on. Okay, so now that we've
5:02:34talked about routines, there are some
5:02:36other ways that you can actually go
5:02:37ahead and deploy automations, especially
5:02:39if they are more code-based, meaning
5:02:41they're more deterministic than
5:02:42nondeterministic. you only really want
5:02:44to go for a routine if you need that
5:02:46full agentic loop and you need like a
5:02:48legit clawed code. But sometimes you
5:02:50don't need clawed code. Sometimes you
5:02:51just need something that's really simple
5:02:53like a Python script that will execute a
5:02:54command or moving data from one platform
5:02:56to another on a web hook or something
5:02:58like that. So let's talk about other
5:03:00ways that you can deploy these sort of
5:03:01like clawed agents. All right, so
5:03:04remember how we talked about this AI
5:03:05systems pyramid a little bit earlier in
5:03:07this course. Where'd it go? Up here. And
5:03:10now that you understand routines, you
5:03:11know that that's basically us, you know,
5:03:13opening up Claude and actually sending
5:03:15it a prompt here and we get that full
5:03:17agentic loop inside of Claude code
5:03:19because we're using the harness and
5:03:20we're using the model because it can use
5:03:22our skills and our files and look
5:03:23through everything. So routines are
5:03:24insanely powerful and I love them. But
5:03:27the thing is you don't always need a
5:03:28routine because once again you want to
5:03:31have it as simple as possible. And
5:03:32routines are basically a mix of the
5:03:33chatbot but also the agent because but
5:03:36realistically what happens is we have
5:03:37this agentic loop where we don't exactly
5:03:40know what it's going to do and why. So
5:03:43if we can make these routines a little
5:03:45bit more deterministic then we probably
5:03:47would like to and on top of the fact
5:03:49that if you're doing cloud routines you
5:03:50have a limited amount of how many can go
5:03:52off per day. So it's always nice to try
5:03:54to find only use these routines when you
5:03:56actually do need them. So, let me talk
5:03:58about a real quick example of let's say
5:04:00we wanted to do a daily AI briefing.
5:04:02What would that look like? Well, we
5:04:05probably would have something like this.
5:04:06The trigger would be maybe 6 a.m. And
5:04:09then from there, what we would do is we
5:04:10would do research. So, research on like
5:04:13the AI space. Now, maybe we have
5:04:14multiple sources of research. Maybe we
5:04:16want to do, you know, X and we also want
5:04:19to research um just in general like
5:04:21Tavly. So, maybe there's two different
5:04:22sources, you know, just the web, maybe
5:04:24even LinkedIn. And then what we would do
5:04:26is we'd maybe consolidate that research
5:04:28and we'd send that into an AI model who
5:04:30would like write the report, right? So
5:04:32this would write the AI briefing and
5:04:36then from there what would happen is we
5:04:38would want to get that sent to us
5:04:39somehow. So send to user. This is
5:04:43basically the order of operations that
5:04:46we need for this briefing. In the
5:04:48morning wake up do the research write
5:04:51the brief send a user. So because this
5:04:53is such a linear predictable process,
5:04:56even though let me just actually um I
5:04:59mean this is like the main AI step,
5:05:00right? So I'll just call kind of call
5:05:02this blue. Even though there's AI inside
5:05:04of it, this is still very much a linear
5:05:06predictable workflow. So there's no need
5:05:10to go for a routine here where we would
5:05:12have, you know, an AI agent step and the
5:05:15agent would be responsible for doing all
5:05:17of the research and the writing and the
5:05:20sending. Right? This is what this would
5:05:23look like more as an agent because it
5:05:26has access to all these tools and it can
5:05:27do things in different orders where like
5:05:29this would totally work as a morning AI
5:05:32briefing, but it's not as efficient and
5:05:34we could definitely do it like this,
5:05:36right? Like we're not really going to be
5:05:37losing quality here. So, this is where
5:05:39we might want to do something like this
5:05:41and host it on modal or trigger.dev. Two
5:05:44tools that I do like to use. So, I'm
5:05:46going to show you guys a quick example
5:05:47of us building this out and it's going
5:05:48to be super simple and we're going to
5:05:50build this out in modal. So, real quick
5:05:52before we start doing this, if you don't
5:05:54have a modal account, go ahead and get
5:05:56signed up. modal.com. It's like I think
5:05:58you get five free bucks to start with
5:06:00and if you put in a credit card, you'll
5:06:01get an extra 25. So, it's super super
5:06:03cheap because it only actually charges
5:06:05you per run and it's pennies. So, you
5:06:07can see it's going to be pretty cheap.
5:06:09So, go to modal.com. You can sign up
5:06:11with the GitHub account that you created
5:06:13earlier if you built the websites out.
5:06:15And yeah, now we can actually build out
5:06:17this automation. So what I'm going to do
5:06:18here is I'm just going to start talking
5:06:20to our cloud code in a really simple
5:06:22way. I am looking to build an automation
5:06:25and we're going to ultimately have this
5:06:27deployed on modal. So it's going to be a
5:06:29Python script. All I want this to do is
5:06:31every morning at 6 a.m. Central time, I
5:06:34want this to fire. I want it to do
5:06:36research for me on the AI space. What
5:06:38I'm specifically looking for are any new
5:06:41AI news like announcements, any stories.
5:06:44So if the, you know, any companies have
5:06:46come out with a case study or if there
5:06:48was a a big failure or something like
5:06:50that or any acquisitions or IPOs, I also
5:06:52want to know if there's any new models
5:06:54or new tools or anything like that. So
5:06:56basically just a concise news briefing
5:06:59of what's going on in the AI space. And
5:07:01then I just want it to write that up in
5:07:02a nice way for me and shoot that over to
5:07:05me in ClickUp. So that is what the
5:07:07automation is supposed to do. I would
5:07:08like to use Opus 4.8 8 as the model that
5:07:12actually writes the email for us or the
5:07:14the daily brief. So with that
5:07:17information, let me know what questions
5:07:19you have. Let me know what API keys
5:07:21you're going to need for me. And then
5:07:22once we build this out, we'll go ahead
5:07:24and push that to modal, which I will
5:07:25also need your help with because I've
5:07:27never used it before. Okay, so that was
5:07:28my prompt. If you guys read this, you
5:07:30can understand exactly what I'm saying
5:07:31here, right? It's not technical at all.
5:07:33And now it's going to help us plan out
5:07:34this automation and then just actually
5:07:36go ahead and build it for us. I think
5:07:37the only thing that you guys might not
5:07:38have said without knowing is that it's
5:07:40going to be a Python script. But Claude
5:07:43would have done research on modal and it
5:07:44would have made sure that it built you
5:07:46the right type of file to send over to
5:07:48Modal. So here's what it's saying. It's
5:07:50going to need an anthropic API key
5:07:52because it's going to obviously need to
5:07:53call on Opus. It's going to need my
5:07:55ClickUp token so that it can send me um
5:07:58you know an actual message in ClickUp.
5:07:59And then it's going to need my Tavly key
5:08:01which I already have because it's going
5:08:03to use Tavi to do the research. And then
5:08:04it's going to need the modal account. So
5:08:06we'll need to give a modal token to
5:08:08authenticate. The first thing is how
5:08:10should the script gather the AI news
5:08:11each morning? We'll just say tavly plus
5:08:13opus. That works fine. How should the
5:08:15finished brief show up in ClickUp? I
5:08:17would like it to be a DM sent to the
5:08:20Nate Herk user in ClickUp. Now this is a
5:08:23permissioning thing. So we'll talk about
5:08:25that to make sure that we're more
5:08:26confident that it's going to be sent to
5:08:28only me in ClickUp. But I'll hit next
5:08:29for now. And then how long in detail
5:08:31should the brief be? I want it to be
5:08:33very very scannable. So we'll do this
5:08:35one. So, while all of that's happening,
5:08:37let's go ahead and start collecting
5:08:38these API keys. So, actually, instead of
5:08:39using an anthropic API key, let's use
5:08:42Open Router. And the reason why is
5:08:43because Open Router is way more
5:08:44flexible. So, if you haven't got an
5:08:46account here, go to open router.ai. And
5:08:48what you'll see here is that it has tons
5:08:50of different models, right? It has one
5:08:52API for all the models, higher
5:08:54availability, and it's just really cool.
5:08:56Basically, what that means is you can
5:08:58have a credit card in here and then you
5:09:00don't have to create an account for API
5:09:02usage, for anthropic, for Google, for
5:09:04OpenAI, for whatever one comes next
5:09:06because Open Router has basically all of
5:09:08the models. So, it's just nice for me to
5:09:10be able to stay organized and watch all
5:09:11of my activity in one spot rather than
5:09:13managing multiple different dashboards.
5:09:15So, when you get in Open Router, you're
5:09:17going to go to your credits. You're then
5:09:19in here going to go to your API keys,
5:09:22and then this is where you're going to
5:09:23create a new key. And we're going to
5:09:25call this our knowledge work demo. Here
5:09:29is where you can have a credit limit as
5:09:31far as price. You can have this expire.
5:09:33And I'm not going to touch any of that
5:09:35right now. I'm just going to go ahead
5:09:36and create this API key. And then, of
5:09:38course, it's going to give us one that
5:09:39we can go ahead and copy. So, I'm going
5:09:41to copy this API key. I'm going to go
5:09:42into Claude. We're going to open up ourv
5:09:45file. And then in here, what I'm going
5:09:47to do is I'm going to make this a little
5:09:48bigger so you guys can see better. I'm
5:09:50basically just going to add manually
5:09:52open router. So, I'm just going to do in
5:09:54all caps open router_appi
5:09:59key equals and I'm going to paste in
5:10:01that key. The next thing I'm going to do
5:10:02is in my ClickUp, I'm going to get the
5:10:04API key. So, I'm going to go into my
5:10:06settings and I have to just find
5:10:08somewhere right here, ClickUp API. And
5:10:10now I can copy this key and paste that
5:10:12in. So, pretty much whatever platform
5:10:13you need, it's always going to be in
5:10:15your settings. and you might have to
5:10:16look for developer settings or somewhere
5:10:18you'll have to find the API
5:10:18documentation or sorry the API key area
5:10:22but if your software has API
5:10:24documentation then you should be able to
5:10:26find your API key in the settings okay
5:10:28and this is pretty cool as I was adding
5:10:30those keys cloud code was actually
5:10:32seeing that and it started testing
5:10:34things out so I didn't even tell it by
5:10:35the way that I was going to switch from
5:10:36enthropic API to open router it just saw
5:10:38that and it used it so tavly is working
5:10:41open router's working and clickup I just
5:10:43added so hopefully it will test that in
5:10:44just a check or I might have to prompt
5:10:46it to but it looks like right now it's
5:10:48building the actual automation. So you
5:10:49can see it created this Python file
5:10:52which says AI daily brief a modal
5:10:54automation every morning central time.
5:10:56You gather fresh AI news with Tavi. You
5:10:58write a concise scannable brief with
5:10:59Opus via open router and then you
5:11:01deliver the brief to Nate as a ClickUp
5:11:02DM selfdm by default. Okay, so there we
5:11:05go. It says the code is written. Before
5:11:07we touch modal, let me prove that this
5:11:09works and show you the output quality.
5:11:11So I can run the research and then Opus
5:11:13will write and then we will skip the
5:11:15clickup send for now because it hasn't
5:11:17seen my token yet. But we'll just
5:11:18basically see the output that we will be
5:11:20getting inside of our ClickUp when this
5:11:23automation is ultimately done. And the
5:11:25other thing to think about here is what
5:11:27is our risk with this automation? That's
5:11:28a really important thing to think about
5:11:30is what is your risk? In this case, our
5:11:32risk is basically that it spends too
5:11:34much money, right? Our risk is that this
5:11:35somehow starts firing off a lot or it
5:11:37gets somehow stuck in a research loop
5:11:39with Tavi and we spend a lot of credits.
5:11:41So maybe we could mitigate that risk by
5:11:43scoping in those API keys so that they
5:11:45can only spend like a dollar a day.
5:11:47Another risk is that open router is down
5:11:49or tavly is down and in that case what
5:11:51do we do? Maybe we route from Tavly to
5:11:54like perplexity to do research if Tavi
5:11:56is down and maybe if open router is down
5:11:58then we would route to Enthropic. So,
5:12:00those are some of the other like what if
5:12:02edge case scenarios to be thinking about
5:12:04and to be protecting against. Anyways,
5:12:06here is what the brief would look like.
5:12:07This would be our chance to give
5:12:08feedback and iterate if we don't like
5:12:10how it's how it's formatted. But this
5:12:11one looks good. It gathered 30 stories.
5:12:13It dduped. It grouped and it's sourced.
5:12:15And we have it said it's written in my
5:12:17tone of voice as well. So, here is our
5:12:19daily brief for Friday, July 10th. Here
5:12:21are new models. So, GBD 5.6. Space X
5:12:23releases Grock 4.5. We're also going to
5:12:25get a clickable link with the actual
5:12:27source and an AI dominated H1 venture
5:12:30blah blah blah blah blah. Okay, cool. So
5:12:32now it's asking for my ClickUp token
5:12:34which I have pasted. So I'll say hey go
5:12:35ahead and test it out and then we'll set
5:12:37up modal. So awesome. So I've given you
5:12:39the ClickUp API key. Now one thing I
5:12:42noticed is in ClickUp I wasn't able to
5:12:44set up like a scoped API key. It seemed
5:12:46pretty general. So, I want you to help
5:12:48me figure out how do we make sure that
5:12:51this automation can only send to my
5:12:54ClickUp channel or my personal Nate Herk
5:12:57DM. I don't want it to ever be able to
5:12:58accidentally send into like the general
5:13:00channel or any public channels. It
5:13:02should only send to this one specific
5:13:04place. How can you prove that to me? And
5:13:06in this case, it's going to be pretty
5:13:07simple because the Python script is
5:13:09going to basically hardcode the endpoint
5:13:11to hit. We're not going to give modal
5:13:13the ability to change where it gets
5:13:15sent. So, we're going to send it to my
5:13:17ClickUp DM once and then it will never
5:13:19change because that step of the process
5:13:21is fully deterministic. This is where if
5:13:24we made this step non-deterministic and
5:13:26we let the agent choose where to send
5:13:27it, that's where you can get
5:13:28variability. But because we're writing
5:13:30an actual Python script, it's not going
5:13:32to change. So, boring is beautiful.
5:13:34Predictability is beautiful when it
5:13:36comes to self-firing automations. Okay,
5:13:39so it hardened up the code. Basically,
5:13:41what it did is it made sure that it's
5:13:43only able to send to this specific DM,
5:13:45which is perfect. It's blocking out all
5:13:47these public channels and blocking out
5:13:49all the other channel paths. And you can
5:13:51see here, this is the test delivery that
5:13:53it sent. This is a one-time test
5:13:54confirming the automation can post to
5:13:56this private DM. If you can read this,
5:13:57then delivery works. So, let's go ahead
5:13:59and get this pushed to modal. So, after
5:14:01all that, it wants us to run these four
5:14:03commands in our terminal, which we could
5:14:04easily just copy and paste in and it
5:14:06would not be difficult at all. But, I
5:14:07said, can't you just run all these for
5:14:08me? can't you just set everything up?
5:14:10And it said, "Yeah, mostly I can. I can
5:14:12set it all up, but you're still going to
5:14:14have to authenticate in." So, do you
5:14:15remember earlier when I was showing you
5:14:17how like sometimes it'll give you a link
5:14:19and you click on the link and then you
5:14:20log in to modal or to claude or to
5:14:22notion or whatever it is, and that's how
5:14:24it authenticates. That's what it needs
5:14:26to do. And it's going to basically kick
5:14:27that off. Now, we'll sign into modal and
5:14:29then claude code will come back here and
5:14:31say, "Okay, cool. I got your modal
5:14:32information. I got that modal token or
5:14:34cookie. And now I can set everything up
5:14:36for you." pretty much the same exact way
5:14:37that I use GitHub. You know, when I have
5:14:39Cloud Code doing things for me inside of
5:14:41GitHub, it can do everything. It can
5:14:43push commits, it can create repos, it
5:14:44can do everything. But I just needed to
5:14:46authenticate once and now Claude Code
5:14:49can do everything. So, as you saw, this
5:14:51page opened. I just have to authorize
5:14:52Cloud Code to use my workspace right
5:14:55here. And then it says, "Your client has
5:14:56been granted an API token and ready to
5:14:58use modal." And now Claude, if we switch
5:15:00back in here, is going to be able to
5:15:02actually make all that stuff. Okay, look
5:15:04at this. In my modal, we can now see
5:15:05that we have this AI news brief and it
5:15:08just got taken away because Claude is
5:15:10right now running a test. So, it tested
5:15:12it out and now maybe it found a problem
5:15:14or maybe it needs to refresh. But
5:15:16anyways, it looks like it's deploying it
5:15:18again. So, I'll go back into modal.
5:15:19We'll hit a refresh. We see if we get
5:15:21anything. This should pop up in a sec.
5:15:23But also, what I realized is I got a
5:15:24ClickUp message. So, this ClickUp
5:15:26message just came in 11:42. And this was
5:15:28our daily brief. So we saw the new
5:15:29models, we saw new tools, we see Google
5:15:32photo ads video remix, we got meta rolls
5:15:34out AI room visualization, we got
5:15:36funding, we have business adoption, and
5:15:38we have some failures and controversies.
5:15:40So this is an example output. If we
5:15:42wanted to change this, we easily could
5:15:43instruct Cloud Code to change up the
5:15:45script a little bit, but right now I'm
5:15:47happy with that. And it looks like in
5:15:48here, we just got our AI news brief
5:15:50back. Let's see what Claude is saying.
5:15:52It says that it's live. It says that
5:15:53this is the app name, and this is the
5:15:54link. Every day at 6:00 a.m. Central, it
5:15:56will go ahead and write that. And take a
5:15:58look at this. So as far as managing it
5:16:00later on modal, we can see everything.
5:16:02We can see the runs, the logs, and the
5:16:03errors. We can trigger it on demand. And
5:16:05we can also rotate keys. So what happens
5:16:07is remember how our secret keys, our API
5:16:11keys live inside of thev file in this
5:16:13project. Modal needed to access those
5:16:15somehow. And what it does is it stores
5:16:17those in modal as secrets right here. AI
5:16:20news brief secrets. You can see that
5:16:21it's able to run this script and that's
5:16:23how it's able to grab different
5:16:24environment variables that we need and
5:16:27it stores them securely in here. Right
5:16:29here, if I click on edit, you can see we
5:16:30have the ClickUp token. Here's the
5:16:32token. Open router and Tavi. All of
5:16:34those tokens are in here. And if we
5:16:35needed to change them, we could change
5:16:37them. And if we wanted to add more, we
5:16:38could add more. So that's where modal
5:16:40stores our secrets. Now, as far as the
5:16:42app, let's take a look in here. So you
5:16:44can see next run is in 18 hours, which
5:16:46would be 6:00 a.m. tomorrow morning. But
5:16:48we could also hit run now. And when I
5:16:49click run now, it schedules the run
5:16:51right now. That doesn't mess with any of
5:16:52the scheduled ones. But what I wanted to
5:16:54show you guys is what it looks like when
5:16:56something is actually running and how we
5:16:58are able to use modal here for our sort
5:17:00of like visibility. And that's really
5:17:02really important observability and
5:17:04visibility. So right now you can see
5:17:05it's running. I'll click into it and we
5:17:07can see in the execution that we're
5:17:09actually seeing what it's doing in real
5:17:11time. So gathering news for Friday, July
5:17:1310th, collected 30 unique stories,
5:17:15writing brief with Opus 4.8. You can
5:17:17look at the execution and you can look
5:17:18at the call graph if you really want to
5:17:20see like the timing of what's going on,
5:17:21which is pretty cool. But there you go.
5:17:23This one just finished up and it only
5:17:25took about 30.36 seconds. You can see if
5:17:28I go to the log, everything here was
5:17:30good. It delivered to ClickUp, sent
5:17:31brief to ClickUp to a private DM
5:17:33channel. If I open up ClickUp, you can
5:17:34see we once again just got another one,
5:17:36which is obviously very similar to the
5:17:37one that it just ran up here. But that's
5:17:39just proving that end to end this thing
5:17:41worked. And now we have logging
5:17:43available inside of Modal. in case
5:17:45anything ever fails, we can figure out
5:17:47why. We have the ID number for each of
5:17:49our executions. So, we can feed it back
5:17:50into cloud code later and say, "Hey, so
5:17:53these 10 runs were good, but if you look
5:17:54at run number 12 or you know, whatever
5:17:57number, this errored right here and
5:17:59figure out why and figure out what we
5:18:01can do to fix this so that this sort of
5:18:03error doesn't ever happen again. All
5:18:04right, so that is like a modal
5:18:06deployment that is cronbased." Cron just
5:18:09basically meaning it is on some sort of
5:18:11schedule. Now, what happens if we wanted
5:18:13to have something that's more web hook
5:18:15based? If you guys don't know what a web
5:18:16hook is, very simple concept. It's
5:18:19basically just think about it like a
5:18:20doorork knob. So, right now, what we're
5:18:22doing is we are scheduling off these
5:18:24automations based on time. So, 6 a.m.
5:18:27central, right? But what if we wanted to
5:18:29schedule an automation where I don't
5:18:31know, a good example would be every
5:18:32single time you get a new form
5:18:33submission. That would be a web hook
5:18:36trigger. Because if you're looking for
5:18:37an event, you can do something called
5:18:39polling, which is basically just
5:18:40constantly checking. It's basically just
5:18:42a loop of seeing, hey, is there a form?
5:18:45If yes, I will process it. If no, I will
5:18:47just wait 5 minutes and pull again or
5:18:49check again. And it's just this constant
5:18:51loop of polling every 5 minutes or every
5:18:5310 minutes or whatever interval you
5:18:55decide to pull. But instead of polling,
5:18:57what you can do is you can set up a web
5:18:59hook, which means basically if you think
5:19:01about it like a door, I did I say
5:19:03earlier doororknob? I did not mean to
5:19:04say doorork knob. I meant to say
5:19:05doorbell. If you think about it like a
5:19:07door, rather than saying, "Oh, I wonder
5:19:09if there's any guests at my door. I'm
5:19:10going to go open it every 10 minutes.
5:19:12And rather than checking every 10
5:19:13minutes, you could instead just say,
5:19:14"Oh, okay. I'm going to put this
5:19:16doorbell here so that whenever someone
5:19:18comes, they ring the doorbell and then I
5:19:19know to go open the door." So that's
5:19:21basically all a web hook is. It's one
5:19:23system sending a request to another
5:19:26system that tells it, "Hey, it's your
5:19:28turn to do some sort of work." So in the
5:19:30case of a form submission, as soon as
5:19:32someone fills out a form and hits the
5:19:33button submit, that should send an API
5:19:36call to our web hook. So, let me show
5:19:38you a super super simple example of what
5:19:40that could look like. So, I'm going to
5:19:41go ahead and do a session handoff here
5:19:43because I want it to sort of have the
5:19:45context of what we just did with modal,
5:19:46but I want to have a fresh context
5:19:48window because right now we are at 133.
5:19:51So, just a good place to sort of clear
5:19:53it out and restart because we're
5:19:54starting a new build. If we were still
5:19:57editing on this one and improving this
5:19:58one, I wouldn't have done this. But
5:20:00because we're starting a completely
5:20:01different like project, I'm going to go
5:20:04ahead and get a fresh session to work
5:20:05off of. Okay, so I'm going to copy this.
5:20:07I'm going to clear out this context,
5:20:09paste in the handoff message, and then
5:20:12I'll show you guys what we're about to
5:20:14do. Okay, so now I'm going to do a /goal
5:20:16just to make this more fun. And here's
5:20:17what I'm going to say. All right, so we
5:20:19just deployed a Python script to modal
5:20:22and that was a cronbased automation. Now
5:20:24what I'm trying to do is create a web
5:20:26hook based automation. So here's what
5:20:28I'm imagining. Create me an HTML
5:20:30document. Super super simple. That's a
5:20:32form submission. And when the user hits
5:20:35submit on that form, it's going to send
5:20:37that to our modal web hook. Now, all
5:20:40modal is going to do that Python script
5:20:43needs to read the form submission and
5:20:46then send me the business owner a
5:20:48notification saying, "Hey, you know,
5:20:50this user submitted a notification. Here
5:20:52is what his business does and here's
5:20:54what he's looking for." And then it will
5:20:56basically just send that to me once
5:20:57again in the exact same ClickUp private
5:21:00DM that the previous automation that we
5:21:02just built did. So you shouldn't really
5:21:04need anything new as far as API keys,
5:21:06but you do need to create me both of
5:21:08these deliverables, both of you know all
5:21:10these scripts and then go ahead and just
5:21:11do as much as you can here and then stop
5:21:14once you've tested and proven that this
5:21:16endtoend pipeline works and is deployed
5:21:19on modal. Okay, so that is my slash
5:21:21goal. I will basically just check in
5:21:23with you guys once this is done so I can
5:21:25show you what that web hook based
5:21:26automation looks like. All right, so
5:21:29this just finished up. That was insanely
5:21:31quick. If I go to my apps now, you can
5:21:33see that we have lead web hook as a
5:21:35separate app. You can see that there's
5:21:36two different functions. So this was a
5:21:38test submit which has zero calls and
5:21:39then we have the actual web function. If
5:21:41you look at this, you can see this isn't
5:21:44triggered on a schedule. If you guys
5:21:46remember in the other app, there was a
5:21:48timer right here and then there was a
5:21:49button to execute now. But the reason
5:21:51why that doesn't happen here is because
5:21:53when we send data over, it has to accept
5:21:55some sort of information, right? Because
5:21:57the form submitted data over to this web
5:21:59hook. So if I go back into the um cloud
5:22:02real quick, you can see that this is the
5:22:04form that it built. So if I put in my
5:22:06name, if I put in my email, and if I put
5:22:09in my business name, which is we'll just
5:22:11put in uppit. What does our business do?
5:22:13Sells shoes. And we're looking for AI
5:22:17implementation. And if I hit submit,
5:22:20let's see if this HTML is working
5:22:21properly. This should trigger a request
5:22:24to modal. So, thanks, your submission
5:22:26was received. If I go back in here, we
5:22:29see that we just got a submission to
5:22:31come through. We got an options and we
5:22:33got a post. So, if I click on this one,
5:22:35we should see sent lead notification to
5:22:37ClickUp. We can go to execution. We can
5:22:39see we got all of this. Now, let's make
5:22:41sure that it's actually capturing the
5:22:43right information. Boom. I go over to my
5:22:45ClickUp and we can see new lead Nate. We
5:22:47have the contact as nateest.com. Their
5:22:50business sells shoes and they're looking
5:22:52for AI implementation. Now, one thing I
5:22:54want you guys to take note of here, this
5:22:57has zero AI involved. It's literally as
5:23:00simple as whatever data is submitted in
5:23:03here. The modal script or the Python
5:23:06script uses a template and basically
5:23:09just fills out the information exactly
5:23:12as the user typed it. If this was AI, it
5:23:14probably would have fixed my typo. I
5:23:15didn't even realize that I put two A's
5:23:16in there and it probably would have just
5:23:18like formatted this a little bit
5:23:19different. But that would be a waste.
5:23:21Build the simplest solution that you
5:23:22actually need for the problem. And in
5:23:24this case, the problem was we want form
5:23:27submissions to instantly come to our
5:23:29ClickUp. And this is getting the job
5:23:31done. I don't really need AI here at
5:23:33all. All that would do is it would add a
5:23:34little bit of risk and it would increase
5:23:36our cost. So great example of a
5:23:38situation where a simple modal script,
5:23:41this took probably four minutes in
5:23:43total. And now anytime someone would
5:23:45submit a form on our website, we would
5:23:47instantly get their submission. All
5:23:49right. So now that we understand how to
5:23:51deploy stuff, I want to talk about
5:23:53another way that you can actually use
5:23:55your own cloud stuff. Because what
5:23:57happens is as we're building out these
5:23:58folders and files and skills and all
5:24:01these connections, it's really valuable.
5:24:03And I don't know about you guys, but I
5:24:04get in this situation where sometimes I
5:24:06get anxious about like leaving my house
5:24:08or not not leaving my house. That makes
5:24:09me sound like I'm a freak. I mean, like
5:24:12I get worried about how much work I'll
5:24:14be able to do if I'm maybe like on a
5:24:16vacation or if I'm away from my PC setup
5:24:18cuz I like my monitors, I like my
5:24:20standing desk, I like all this, right?
5:24:21But what we can do is we can use a thing
5:24:23called remote control, which basically
5:24:25lets us use our phones to actually
5:24:27continue to work on our sessions. And
5:24:29what I love about remote control is
5:24:30because it's it's so so easy where I can
5:24:33go down to the gym and before I do, I'll
5:24:34just start a remote control. So, as I'm,
5:24:36you know, working out, I can keep
5:24:37sending off prompts and I can keep
5:24:39building stuff. or if I know I'm going
5:24:40to go on a walk or if I know I'm going
5:24:41to go to lunch or even if I know that
5:24:43I'm going to be away from my home setup
5:24:44for a few days, obviously I bring my
5:24:46laptop and I'll still have my AIOS on my
5:24:48laptop, but I can still start a session
5:24:50from my phone and just check in on
5:24:51things. So, remote control, super cool
5:24:54and it's super easy because all you have
5:24:55to do here is you're on a certain
5:24:57account, right? So, pretend this is, you
5:24:59know, nateis.com
5:25:00is my email that I have this account on
5:25:02on my phone on the Claude app. I would
5:25:04also just need to be signed in
5:25:06nateis.com.
5:25:07And then let's say you see this session,
5:25:08right? This was our agent team debate
5:25:10session. All I have to do is come in
5:25:12here and do a slash remote control. And
5:25:14when I shoot that off, that basically
5:25:16just like exposes this session. It is
5:25:19local, but it exposes it for my phone to
5:25:21be able to control it. So, I'm going to
5:25:23open up my phone here and I'm going to
5:25:24go into the clawed app on my phone. Now,
5:25:26when I'm in here, what I can do is I can
5:25:28go to the code section. And now, you can
5:25:30see I have an idle session. So, you guys
5:25:32probably can't see this too well, but if
5:25:34I hold this up to the camera, this is
5:25:36our session, right? So you can see here
5:25:38it says like the verdict, the stats
5:25:40table, the round table. If you look
5:25:42really hard, you can actually see that's
5:25:43exactly what it says right here. The
5:25:45verdict, the stats table. Anyways, let
5:25:47me just prove it to you by shooting off
5:25:48a prompt. So on my phone, I'm saying,
5:25:50"Hi, this is from Nate's phone." And I
5:25:55will hit enter. And now you see that
5:25:56this is going to come through on our
5:25:58laptop right here, or sorry, not laptop,
5:26:00on the desktop. And it's processing that
5:26:01message. So it saw, hi, this is from
5:26:03Nate's phone, blah blah blah. And I'm
5:26:05getting all of this response on my phone
5:26:07as well. So basically the point is we
5:26:09now have two different ways to control
5:26:10this session. I can do it from my phone
5:26:12and I can also do it from right here. So
5:26:15what also is cool is I can come in here
5:26:17and I can do a slashclear. So I sent
5:26:19slash clear from my phone and what's
5:26:21going to happen is it's still able to
5:26:22recognize that that was a slash command
5:26:24and you can still have all that full
5:26:26functionality of what you typically are
5:26:28doing when you are driving this from
5:26:30your computer. Okay, you know what?
5:26:32That's actually really weird, but I'm
5:26:34glad that we found this out. So, I'm not
5:26:35sure if this is a bug because I have to
5:26:36relaunch or, you know, the desktop app
5:26:39is always being improved on, but both of
5:26:41my slashclear commands came through as
5:26:43no content, which is very strange
5:26:45because I do this a lot from my VS Code
5:26:47terminal and it works. So, that's just
5:26:49another quick example of some of the
5:26:51tiny tiny little things where the VS
5:26:53Code terminal gets you the full
5:26:55functionality. And for some reason, that
5:26:57didn't work when I did this here, but
5:26:58that does work when I'm using the
5:27:00terminal. But anyways, that is remote
5:27:02control. Like I said, it's super handy
5:27:03if you know you're just going to be
5:27:04stepping out for a bit, but you still
5:27:05want to keep being able to check in on
5:27:07something and keep working on something.
5:27:08So hopefully you guys are able to find
5:27:10some good use cases for that. Okay, so
5:27:11that was remote control. Now, token
5:27:14management, such a big conversation,
5:27:17right? It's always going to be important
5:27:19even when we get into the future and
5:27:20models are getting cheaper potentially
5:27:22and you're able to run them locally, but
5:27:24token management is so so important
5:27:25because not only from a cost
5:27:27perspective, but from a performance
5:27:29perspective. We all know about context
5:27:30rot and we all know about things like
5:27:33confusion or bloating when it comes to
5:27:35the context window of your AI agents.
5:27:37So, we're going to spend some time here
5:27:39talking about tokens, what they really
5:27:41are, how they work, how Claude works
5:27:43with them, and how you can actually use
5:27:45these little tricks and use these things
5:27:46and keep them in mind to maximize your
5:27:49sessions and your tokens. In the past
5:27:51week or so, so many people have been
5:27:52complaining about hitting their claude
5:27:54code limit insanely fast. claims like
5:27:56one prompt that is about 1% of the limit
5:27:58is now around 10%. You could go through
5:28:00X and find tons and tons of threads
5:28:02about this topic. Even on a $200 per
5:28:04month plan, people are reaching the
5:28:06session limit way too fast. And then we
5:28:08got this post from an anthropic employee
5:28:10that basically said that they are
5:28:11working on a little change with peak
5:28:13hours and off peak hours. But even after
5:28:15that, some people were saying they were
5:28:17still hitting it really quick even
5:28:18during off- peak hours. So anyways, I've
5:28:20been playing around a ton, trying
5:28:21different things, doing research, and I
5:28:23have 18 token management hacks for you
5:28:25guys that I've organized from tier one
5:28:27all the way up to tier three, so they
5:28:28get more advanced as we go. I'm very
5:28:30confident that by the end of this video,
5:28:31you will feel like your Claude code
5:28:33usage has doubled, tripled, maybe even
5:28:355xed. So, let's not waste any time and
5:28:37just get straight into the video. So, as
5:28:39I've been optimizing my own token
5:28:40management, I think that what's really
5:28:42important to realize first is how tokens
5:28:45actually work. Because once you realize
5:28:46how Claude uses tokens, it makes it very
5:28:49clear how you should actually reverse
5:28:51engineer the way that you work in order
5:28:52to use less tokens. So a token is the
5:28:55smallest unit of text that an AI model
5:28:57reads and charges you for. It's roughly
5:28:59one token is one word, but that's not
5:29:01explicitly true. Kind of just a good
5:29:03baseline. So every time that you send a
5:29:04message, Claude rereads the entire
5:29:06conversation from the beginning. And all
5:29:08of those are tokens that it's charging
5:29:10you for. So, message one, it will read
5:29:12it, then it will read its reply, and
5:29:13then message two, and then the reply all
5:29:15the way up to your latest prompt. And it
5:29:17does that every single time. And I think
5:29:19that alone is a huge light bulb moment
5:29:21for a lot of people. This means as
5:29:23you're having a conversation with
5:29:24Claude, your cost is compounding, not
5:29:26just adding, it's exponentially growing.
5:29:29Meaning, message one might cost 500
5:29:30tokens, message 30 costs 15,000 because
5:29:33it's rereading everything before it. One
5:29:36developer actually tracked a 100 plus
5:29:38message chat and found that 98.5% of all
5:29:41the tokens were just spent rereading the
5:29:43old chat history in the session. Like
5:29:45that's a huge waste. Now yes, the
5:29:46argument has to be made that well it
5:29:48needs the context and it needs to
5:29:50understand what we're doing. But still
5:29:5198.5% is crazy. So take a quick look at
5:29:54this graphic here. Along the x-axis we
5:29:56have message number and as it increases
5:29:58you can see that we have our per message
5:30:00cost and our cumulative tokens
5:30:02increasing. But it's not linear. It's
5:30:04basically each message is rereading all
5:30:06of the past ones and it has to count
5:30:08that in. So message one could be 500,
5:30:10message 30 could be 15,500 which is 31
5:30:13times more. And then after 30 messages
5:30:14you might already be at almost a quarter
5:30:16million cumulative tokens. Now on top of
5:30:18all of your own messages, Claude will
5:30:20also reload your cloud.MD, your MCP
5:30:23servers, your system prompts, your
5:30:24skills, your files on every single turn.
5:30:26And this is invisible overhead, but it
5:30:28is constantly dripping into your context
5:30:30and your tokens. And a really important
5:30:32thing to realize is that bloated context
5:30:33doesn't just cost you more money, but it
5:30:35also produces worse output. So you're
5:30:37paying more and you're getting less.
5:30:38There's this phenomenon called loss in
5:30:40the middle, which basically says that
5:30:42models are paying the most attention in
5:30:44the beginning of your session and kind
5:30:45of at the end. So all that stuff in the
5:30:47middle of your session, kind of in this
5:30:48dip is getting ignored. All right, so
5:30:51now that we kind of understand a little
5:30:52bit more about how cloud code works and
5:30:53how tokens work, let's move into the
5:30:55hacks. We're going to start here with
5:30:56tier one hacks. These are the ones that
5:30:58are going to be super easy to implement
5:31:00and everyone should be able to
5:31:01understand. So, we've got nine of these.
5:31:03Number one is to start fresh
5:31:05conversations. Use slashclear between
5:31:08unrelated tasks. Don't carry context
5:31:10about topic A into a conversation about
5:31:12topic B. So, every single message in a
5:31:14long chat is exponentially more
5:31:16expensive than the same message in a
5:31:17fresh chat. So, this one habit is the
5:31:20number one thing that extends your
5:31:21session life. And it's pretty obvious
5:31:23based on what we just talked about. So,
5:31:25that's why this was number one. Okay,
5:31:27number two is to disconnect MCP servers.
5:31:30Every single connected MCP server loads
5:31:32all of its tool definitions into your
5:31:34context on every message. This is
5:31:36another source of completely invisible
5:31:38tokens that might just be eating up and
5:31:39eating away. So, one server alone might
5:31:42be something like 18,000 tokens per
5:31:44message. So, run MCP at the start of
5:31:47each session and disconnect the ones
5:31:48that you don't need. And better yet, if
5:31:50you're able to find CLIs for something,
5:31:52so for example, rather than having the
5:31:53Google Workspace or Google Calendar MCP
5:31:56server, which eats a lot of tokens, just
5:31:58use the Google Workspace CLI. It's
5:32:00faster, it's cheaper, and I think the
5:32:02future is moving towards having our
5:32:04agents use CLIs rather than MCPs. All
5:32:07right, number three, batch prompts into
5:32:09one message. Three separate messages
5:32:12cost three times what one combined
5:32:13message costs because of the way the
5:32:15tokens work, right? Instead of summarize
5:32:18this as one message and then now extract
5:32:20the issues, now suggest a fix, send it
5:32:22all in one prompt. If clog gets
5:32:24something slightly wrong, edit your
5:32:25original message and regenerate instead
5:32:27of sending a full follow-up correction.
5:32:29Follow-ups stack onto history
5:32:31permanently while edits replace the bad
5:32:33exchange entirely. Now, I will say there
5:32:35is an argument to be made here that
5:32:37potentially doing it this way where
5:32:39you're doing task one, task two, then
5:32:41task three might actually be better
5:32:43output quality. I think it depends on
5:32:45the actual use case. Basically, the idea
5:32:47would be if you can give AI one specific
5:32:49task at a time, it's going to do better
5:32:51because it's more specialized and it's
5:32:52more focused. But this is definitely
5:32:54something that you should be aware of.
5:32:55Okay, number four is to use plan mode
5:32:57before any real task. This lets Claude
5:32:59map out the approach, ask you the right
5:33:01questions, and it prevents the single
5:33:03biggest source of token waste, which is
5:33:04just having Claude go down the wrong
5:33:06path, writing code, and then basically
5:33:08everything that it just did, you have to
5:33:10basically like scrap and redo. It's just
5:33:11a huge waste of time and tokens. So, you
5:33:13can add something like this to your
5:33:14cloudmd. Do not make any changes until
5:33:16you have 95% confidence in what you need
5:33:18to build. Ask me follow-up questions
5:33:20until you reach that confidence level.
5:33:22This is something that I'm putting into
5:33:23all of my cloudmds when I am having it
5:33:26help me build things. Number five, we
5:33:28have run/context and /cost. /context
5:33:31shows you exactly what's eating your
5:33:32tokens right now. So, your conversation
5:33:34history, your MCP overhead, loaded
5:33:36files, stuff like that. And /cost shows
5:33:39you your actual token usage and
5:33:41estimated spend for that current
5:33:42session. Most people have no idea where
5:33:44their tokens are going. And these two
5:33:46commands make the invisible visible
5:33:48because if you don't actually know that
5:33:50you're bleeding because of MCPS, then
5:33:51how would you be able to fix that? So,
5:33:53when you run /context, this is what it
5:33:55will look like. It'll basically give you
5:33:56a screenshot of how many tokens you're
5:33:59at, what is the cap, and it will
5:34:01estimate based on the different
5:34:02categories. And what I did here is this
5:34:04was ran in a completely fresh session,
5:34:06no chats. So, what that tells me is,
5:34:08okay, before I even talk to Claude, I'm
5:34:10already down 51,000 tokens because of
5:34:13things like the system prompt, the
5:34:14system tools, my custom agents, my
5:34:16skills, memory files, and here I've
5:34:19actually cleared out all the MCPS, so
5:34:20there wasn't anything in there, but
5:34:22those can, like I said, completely blow
5:34:24up your tokens right from the get- go.
5:34:26Okay, number six is to set up a status
5:34:28line. This kind of goes handinhand with
5:34:30having more visibility. You only
5:34:32actually see this in your terminal,
5:34:33though, so you will have to do it there.
5:34:34Um, and it basically lets you see what's
5:34:37going on. So, right here, you can see in
5:34:38my terminal, I've got this set up so
5:34:39that I can see the model I'm using. I
5:34:41can see a visual kind of progress bar of
5:34:44my usage. And then I can see uh 5% of my
5:34:47whole 1 million context window. And I
5:34:50can see 52,000 tokens out of a,000,000,
5:34:53which is a million. And just to clarify,
5:34:54this isn't my session, like my 5 hour
5:34:56session. This is basically just
5:34:58indicating that I'm 5% of the way or 52K
5:35:01out of a,000K. So all you have to do is
5:35:03include code in the terminal do /st
5:35:05status line and explain that you want to
5:35:07replicate this setup. Number seven is
5:35:09just super simple but keep your
5:35:11dashboard open. Same thing with
5:35:12visibility. You might run into issues
5:35:14with your limit and just get hit out of
5:35:16nowhere. But if you have it pulled up
5:35:18next to you or you have it ready so that
5:35:19you can switch into that tab and you
5:35:20know check every 20 40 minutes then
5:35:23you're going to be able to pace yourself
5:35:24a little bit better. You could even set
5:35:26up a automation to basically check in on
5:35:28it every 30 minutes and send you like a
5:35:30text or a Slack message and say, "Hey,
5:35:32by the way, you're getting near your
5:35:35usage." All right, so number eight, we
5:35:37have be smart with pasting. Before you
5:35:39drop a document or a file or something
5:35:41large, just ask yourself, does Claude
5:35:44need to read this whole thing? Sometimes
5:35:45it does. Sometimes it needs that full
5:35:46context, but sometimes it just needs one
5:35:48section or one page. So if the bug is in
5:35:51one function, then paste just that
5:35:52function. or if it just needs the
5:35:54context of one little paragraph, just
5:35:55paste that. Claude needs to be precise
5:35:57about what it reads, but you also need
5:35:58to be precise about what you feed it.
5:36:00And number nine, our last tier one hack
5:36:02is to actually watch Claude code work.
5:36:05Don't just fire off a prompt and walk
5:36:06away or switch tabs. Watch what Claude
5:36:08is doing, especially on longer tasks.
5:36:10And this is because if you actually sit
5:36:12and watch it, sometimes you'll realize
5:36:13it's going down the wrong path.
5:36:14Sometimes it gets stuck in its own
5:36:16loops, rereads the same files, things
5:36:18like that. So, if it's doing that, you
5:36:20might as well just stop it right there.
5:36:22Kind of the same idea as plan mode. Why
5:36:24would you let it go down the wrong path,
5:36:25waste all your tokens, and then just
5:36:27have to scrap it all in a bad loop? 80%
5:36:29of the tokens are being used, producing
5:36:32zero value. So, if you're able to just
5:36:34watch your session run until you know
5:36:36it's going down the right path, it could
5:36:37save you thousands of tokens. All right,
5:36:39let's kick it up a little bit. Let's
5:36:40move into our tier 2 hacks. And for
5:36:42these ones, we have five of them. So,
5:36:44number one is to keep your claw.md file
5:36:47lean. Place it in your project route,
5:36:49whether that is globally or in in local
5:36:51project. Claude auto reads it at the
5:36:53start of every single chat as system
5:36:55context. So keep it under 200 lines.
5:36:58Include things like your text stack,
5:36:59your coding conventions, your build
5:37:00commands, the 95% confidence rule, only
5:37:03the most important things. And you need
5:37:05to treat this like an index route to
5:37:07where more data lives. And it's a
5:37:09complete mindset shift. This file
5:37:10basically just tells Cloud Code where is
5:37:12everything that it needs and what to do
5:37:14every single time. So it can point to
5:37:16files that are huge, but that way it
5:37:17just says, "Okay, I don't need this
5:37:18right now, but if I do need this, I know
5:37:20exactly where to look." And because it
5:37:22knows exactly where to look, it's not
5:37:23going to waste time or tokens searching
5:37:25through and reading other files. It's
5:37:26just able to grab it right there by the
5:37:28file name. And the reason I say this is
5:37:29a mindset shift because you should be
5:37:31doing this with other things, not just
5:37:32your cloudmd, with your skills or with
5:37:34your um, you know, master reference
5:37:36guide sheets. I saw someone talking
5:37:38about how they created an index that's
5:37:39super super lean and it shows cloud code
5:37:42exactly where to go in the cloud code
5:37:44documentation. So if it needs help with
5:37:46something related to cloud code, it
5:37:47doesn't have to search through the whole
5:37:48database. It can just say, "Okay, here's
5:37:50my index file. I know exactly which URL
5:37:52to look up at." Super simple. You want
5:37:54to keep this lean and trim it all the
5:37:55time. It's always a work in progress
5:37:57because every single chat, not just like
5:37:59your session, every single message,
5:38:00CloudMD gets read. So if your CloudMD
5:38:03file is a thousand lines, every single
5:38:05time you shoot off a message, even if
5:38:07you just say hi, the whole thing's going
5:38:09to get read. Okay. Number two here is to
5:38:11be surgical with file references. Don't
5:38:13just say something like here's my whole
5:38:14repo, go find the bug. Say something
5:38:16more like check the verify user function
5:38:18inside the off.js file. Or you can also
5:38:22use at file name to point at specific
5:38:24files instead of once again letting
5:38:26claude explore freely. The whole idea of
5:38:28being specific and routing. All right.
5:38:30So number three, I'm saying to compact
5:38:32at around 60% capacity. Autocompact
5:38:35triggers at like 95% by which point your
5:38:37context is already pretty degraded. So
5:38:39run /context to check your capacity
5:38:41percentage or you should have the status
5:38:42line set up and at about 60% just run
5:38:45the slash compact with specific
5:38:47instructions on what it should actually
5:38:48be preserving. After three to four
5:38:50compacts in a row the quality does start
5:38:52to degrade. So at that point once you've
5:38:54done three or four just get a session
5:38:56summary/clear
5:38:58give the session summary back and then
5:38:59keep going. All right so number four
5:39:01short breaks are actually costing you.
5:39:04Cloud code uses prompt chaining to avoid
5:39:06reprocessing unchanged context, but the
5:39:08cache has a fiveminute timeout. So if
5:39:10you step away and you come back and it's
5:39:12been longer than 5 minutes, your next
5:39:14message reprocesses everything from
5:39:15scratch at full cost. And that is why
5:39:18some people feel like their usage just
5:39:19randomly spikes if they might have, you
5:39:21know, stepped away and came back. So if
5:39:23you're going to do that, just consider
5:39:24doing a /compact or a sl before you step
5:39:27away. All right, number five, command
5:39:29output bloat. When Claude runs shell
5:39:32commands, the full output enters your
5:39:33context window. So if you have a command
5:39:36that it comes back with 200 commits or
5:39:38you know just tons and tons of data,
5:39:40then all of that is tokens that gets
5:39:43sent to your model. So really the
5:39:45takeaway here is to be intentional about
5:39:46what you let Claude run. If you know in
5:39:48a certain project that it doesn't need
5:39:50to use certain commands, then you can go
5:39:51ahead and in that project deny those
5:39:54permissions. And this is another one
5:39:56that seems like it's invisible because
5:39:57when it runs like a bash or, you know,
5:39:59certain commands, it basically just has
5:40:01like one line and you don't actually see
5:40:03all the tokens that it has, you know,
5:40:05sent there. All right, so I'm sitting
5:40:07here editing this video and there's just
5:40:09one more thing that I wanted to get off
5:40:10my chest and it's basically about
5:40:13hitting your limit. And you know, the
5:40:15goal of this video and your goal should
5:40:16be to optimize so that you don't hit
5:40:18your limit. But I don't think that you
5:40:20should associate hitting your limit with
5:40:21like it shouldn't be a negative
5:40:23connotation because ultimately if you're
5:40:25doing a lot of these hacks and you are
5:40:28not just like being wasteful with tokens
5:40:31then hitting your limit is actually a
5:40:32good thing if you think about it because
5:40:33it means that you are using this tool so
5:40:36much and I think that's what you want to
5:40:37be. I think you want to be a power user
5:40:40of this tool to the point where it's
5:40:41like got to wait again and you know
5:40:44waiting sucks but people that are using
5:40:47it so much are going to be so much more
5:40:49productive and so much farther ahead
5:40:51than people who are never hitting their
5:40:53limits you know not make not getting
5:40:55their their money's worth and not truly
5:40:59getting the leverage that you are now
5:41:00getting. So anyways, quick little raw
5:41:02rant there, but I think it's an
5:41:04important mindset shift to have. You
5:41:05know, just something to think about. All
5:41:07right, so we're moving on to tier three
5:41:09now. I hope you guys feel like you
5:41:11already have a lot of things that you
5:41:12want to implement and these ones are
5:41:14getting a little crazier as well. So
5:41:15we've got four of these here and I've
5:41:16got a few bonus ones also, but number
5:41:18one is to pick the right model. So
5:41:20sonnet for your default most coding
5:41:22work, haiku for sub aents, formatting,
5:41:24simple tasks, opus for deep
5:41:26architectural planning, and only when
5:41:27sonnet wasn't enough. Try to keep this
5:41:29under 20% of usage or unless you just
5:41:31really really need it for that project.
5:41:33Now, a little tip here is when you have
5:41:34a huge code base and you want to do
5:41:36certain things like maybe a review, then
5:41:38try bringing in codecs. There is an
5:41:40official plugin now and I made a video
5:41:41about this. I'll tag that right up here.
5:41:43But you could basically have, you know,
5:41:44Opus and Sonnet working together to
5:41:46build you, you know, a project or a
5:41:48codebase and then you could just bring
5:41:49in codeex to actually review everything
5:41:50and that way you're saving yourself on
5:41:53the clawed tokens. The next one, number
5:41:55two here, is the cost of sub agents.
5:41:58Agent workflows use roughly seven to 10
5:42:00times more tokens than a standard single
5:42:01agent session. Now, why is that? Because
5:42:04they wake up with their own full context
5:42:06and it's a separate instance. So, they
5:42:08basically have to reload everything when
5:42:10you start up the new session. All of
5:42:12those files, all of the system tools,
5:42:14like everything like that. Now, what you
5:42:15can do though, which is helpful, is to
5:42:17delegate to sub agents for one-off
5:42:18tasks, especially if you want that
5:42:20one-off task to use Haiku. So maybe you
5:42:22need to process a lot of information or
5:42:24maybe you need to do a ton of research
5:42:25and get just like a summary back. Now
5:42:27yes, tokens are still tokens no matter
5:42:29what at the end of the day, but if you
5:42:31can make 80% of your tokens a cheaper
5:42:33model rather than 80% of your tokens an
5:42:35expensive model, then you're going to be
5:42:37saving money. And then of course agent
5:42:39teams are cool. Um sometimes I really do
5:42:41actually like them and it helps me get
5:42:43more higher quality outputs, but they're
5:42:46very very expensive. So try to use them
5:42:48very sparingly. All right, so number
5:42:50three is to understand peak hours. So we
5:42:53just talked about at the beginning how
5:42:54they've adjusted how fast your 5 hour
5:42:56session window drains based on demand
5:42:58during the peak hours, which are 8:00
5:43:00a.m. to 2:00 p.m. Eastern time on
5:43:02weekdays. But off peak, this is when
5:43:05your usage is kind of either normal or
5:43:07it lasts a little longer. And these are
5:43:09afternoons, evenings, weekends. So, if
5:43:11you actually think about this
5:43:12strategically, maybe you want to make
5:43:14sure that you're running big refactors
5:43:16or multi- aent sessions or big projects
5:43:18during off- peak hours only. Otherwise,
5:43:20you're going to, you know, drain right
5:43:22through that peak session. And on top of
5:43:24this, we'll call this a little hack 3.5,
5:43:27which is the one I kind of alluded to
5:43:28earlier when I said, hey, just keep open
5:43:30your clot account so you can see your
5:43:31usage at all times. If you're near a
5:43:33reset and you have room left in your
5:43:35allocation, then go heavy. Try to hit
5:43:36that usage limit before it resets. Get
5:43:39your money's worth. Let your agents go
5:43:40loose at that point. And on the other
5:43:42side, if you're getting near your limit,
5:43:44but you still have lots of time, then
5:43:46step away. This is your time to take a
5:43:48break, take a walk, make some lunch,
5:43:50come back with a full budget instead of
5:43:52burning the last 5% on something small
5:43:54and getting stuck mid task and having to
5:43:56just kind of, you know, lose that flow
5:43:58state that you might have been in. Okay.
5:44:00Number four, your systems constitution,
5:44:03which is claw.md. This should contain
5:44:05stable decisions, architecture rules,
5:44:07and progress summaries. Think of it like
5:44:09the source of truth that makes every
5:44:10prompt shorter and shorter. Save
5:44:12decisions, not conversations. Every
5:44:14architectural call that you store there
5:44:16is a paragraph that you never have to
5:44:17type again. So, this builds on top of
5:44:19the way that you were thinking about it
5:44:20back in tier 1. You can add rules in
5:44:23there that basically tell it, "Hey, I
5:44:24want you to help me make sure I'm being
5:44:26smart about tokens. Use sub agents for
5:44:28any exploration or research. If a task
5:44:30needs three plus files or multi-file
5:44:32analysis, spawn a sub aent and only
5:44:34return summarized insights. Spawn that
5:44:36sub aent in Haiku. And here's a little
5:44:38prompt that I have at the bottom of
5:44:39mycloud.mmd. And I will say before I
5:44:42read this out, you have to be careful
5:44:43because when you make a file like this,
5:44:46um, kind of self-learning or
5:44:47self-evolving, you have to check on it
5:44:49frequently because you don't want it to
5:44:50accidentally get too bloated. But here I
5:44:52said applied learning. When something
5:44:54fails repeatedly, when Nate has to
5:44:56reexplain, or when a workaround is found
5:44:57for a platform tool or limitation, add a
5:45:00oneline bullet here. Keep each bullet
5:45:02under 15 words. No explanations. Only
5:45:04add things that will save time in future
5:45:05sessions. And then it's got some
5:45:06bullets. Now, I'm not saying this is the
5:45:08most optimal prompt, but I think this
5:45:10sort of system of having your Claudetm
5:45:13MD actually learn and continuously think
5:45:15about how it can save you time and
5:45:17tokens is a good idea to play with. All
5:45:19right, so I know that we just went
5:45:20through a ton of stuff. This whole slide
5:45:22deck will be available for download for
5:45:24free in my free school community. The
5:45:26link for that will be down in the
5:45:27description. But right now, what you
5:45:28should go do are these things. Go
5:45:30run/context, see what it looks like. Go
5:45:33to some of your active sessions.
5:45:34Run/cost status line. Make sure it's
5:45:36showing your model, your context
5:45:38percentage, and your token count. Make
5:45:39sure you pull up your clog usage
5:45:41dashboard so you can see your remaining
5:45:42allocation and what time it resets.
5:45:45Disconnect unused MCP servers. Start
5:45:47complex tasks in plan mode. Use/clear
5:45:50when you're switching to an unrelated
5:45:51task. Manually compact at 60% context.
5:45:55Batch your multi-step instructions into
5:45:57single messages. And schedule heavy
5:45:59sessions for off- peak hours. and really
5:46:01just be mindful about the actual timing.
5:46:03So, I wanted to kind of leave you guys
5:46:04with one maybe two messages. The first
5:46:06thing is just the idea that there is a
5:46:09balance between quality and cost. And
5:46:12so, that's kind of a game that you have
5:46:13to play a little bit. And sometimes you
5:46:14do have to go for the higher quality,
5:46:15which ultimately is going to cost you
5:46:16more money. And that's just the way it
5:46:18works. But the other thing is just to
5:46:19keep it simple and think about what we
5:46:22talked about at the beginning of this
5:46:23video, how tokens actually work, how
5:46:25Claude Code actually charges you. Most
5:46:27people don't need a bigger plan. they
5:46:28need to stop resending their entire
5:46:30conversation history 30 times when you
5:46:32could just send it, you know, five
5:46:33times. It's not a limits problem. It's a
5:46:35context hygiene problem. So, look at
5:46:37this. On this day, I saved 91 million
5:46:39tokens because of cache read. And in the
5:46:41past week, I've saved over 300 million
5:46:43tokens because of it. Now, don't freak
5:46:44out. This isn't anything that you have
5:46:46to go change. This is happening
5:46:47automatically if you are using Claude
5:46:49Code or Claude. And I know that the
5:46:50concept of prompt caching might seem a
5:46:52little bit overwhelming, but today I'm
5:46:53going to make it as simple as possible
5:46:54and only really tell you what you need
5:46:56to know in order to make sure that you
5:46:57are saving your session limits and
5:46:59saving tokens. I'll also give you guys
5:47:00this entire token dashboard for free so
5:47:02you can actually start tracking your
5:47:03tokens a little bit better. Anyways, so
5:47:04let's talk about prompt caching, why
5:47:06your sessions burn out, and how to stop
5:47:08it. So what does caching actually cost
5:47:10you? Well, cached tokens only cost you
5:47:1210% of normal input. So all the tokens
5:47:15that are getting cached are saving you a
5:47:16ton of money. So, if we go back to this
5:47:18example, on this day when I had 91
5:47:20million tokens cached, that costed me
5:47:23only as if I was processing about 9
5:47:25million of those tokens. The cache
5:47:26window on a cloud subscription is an
5:47:28hour. Meaning, if you're working with
5:47:30cloud code and you don't touch it for an
5:47:31hour and then you send another message,
5:47:33everything in that session gets
5:47:35uncashed. So, if you leave a session
5:47:37sitting for an hour or longer, then
5:47:39you're going to pay more for it. And if
5:47:40you're using Claude via API or sub
5:47:42agents, then the TTL or the time to live
5:47:44is only 5 minutes. You can change that,
5:47:46but it's just a little bit more
5:47:47expensive. You can bump it up to an hour
5:47:48if you want. But for claude code inside
5:47:50of your terminal or your extension,
5:47:51whatever it is, that's an hour. And now
5:47:53here's a quote from Thoric from
5:47:54Enthropic. He said that we actually run
5:47:56alerts on our prompt cache hit rate and
5:47:58declare SUVs if they're too low. So
5:48:00basically them saying, "We take this
5:48:01stuff really, really seriously and if we
5:48:03see that the hit rate isn't very high
5:48:05for users cloud code caching, then we do
5:48:08something about it immediately." And
5:48:09that's very nice of them. But also, of
5:48:11course, it benefits themselves because
5:48:13with a high cash hit rate, cloud code
5:48:15feels faster. Their serving cost is
5:48:17lower. Subscription limits feel more
5:48:19generous, you know, because you're using
5:48:20less. And long coding sessions stay
5:48:23practical. And then if you have low cash
5:48:25hit rate, this is what happens. And
5:48:26obviously, it's just a lose-lose for
5:48:28everybody. And that's why I said like
5:48:30prompt caching can get very, very
5:48:32complex. And if you want to check out
5:48:34more, then I'll link this article right
5:48:35here, which Thor really goes into some
5:48:38depth here. But if you read this, at
5:48:39least when I did, I was like, "Okay,
5:48:40this is a little bit overwhelming." I
5:48:42have a feeling I don't actually need to
5:48:43know all of this, but I do need to know
5:48:45at least a little bit, at least, you
5:48:46know, the 8020 of prompt caching so that
5:48:48I can get the most out of my session
5:48:50limits. And that's what I'm going to
5:48:51break down today. So, let's take a look
5:48:53at an example of how this actually
5:48:55grows. So, by default, when you shoot
5:48:58off a message to Claude, there's going
5:49:00to be some information that needs to be
5:49:01cached right away. And actually, let me
5:49:02just switch back to one of Thor's
5:49:04graphics real quick. You can see here
5:49:05that we have the base system
5:49:07instructions get globally cached. We
5:49:09have tools like read, write, bash, grap,
5:49:11glob globally cached. We have per memory
5:49:13or sorry per project things like
5:49:15cloud.mmd in memory and that gets cached
5:49:17per project. We've got session state and
5:49:19then we have user messages which grow
5:49:21each turn. So now that we take this into
5:49:25context and we flip back over here, this
5:49:26is what it looks like. This is an
5:49:28example where we have four turns. So on
5:49:30turn one there's no cache. Basically
5:49:31we're matching on a prefix. So don't
5:49:33really have to worry about what that
5:49:35means, but I might mention that later.
5:49:37So anyways, on turn one, there's
5:49:39nothing, right? We're opening up a fresh
5:49:40session. We load in the system prompt,
5:49:42the project context, and we shoot off
5:49:44our first message. And all of this is
5:49:46kind of in this like brown highlight
5:49:48border, which means that this is new,
5:49:50and it has to be fully processed, and
5:49:52it's being written to the cache here. So
5:49:55before I continue down this graphic in
5:49:57this dashboard, you can see that we have
5:49:59the difference between cache create and
5:50:01cache read. So on these days you can see
5:50:03what are my input tokens, my output
5:50:05tokens and my cash create. And then over
5:50:07here you can see my daily cash reads.
5:50:08And just a quick explanation a cash
5:50:10create is writing something into cache
5:50:13for the first time. It's a one-time cost
5:50:15and it pays off the next turn unless of
5:50:17course everything gets uncashed. And the
5:50:18cash read is tokens that claude reused
5:50:20from a cache like your claw.mmd or some
5:50:23of the files or some of the global
5:50:24system instructions. And these are the
5:50:26things that are 10 times cheaper than
5:50:28fresh input. So anyways, on turn two,
5:50:31given that we're within that 1 hour TTL
5:50:33window, everything here is already in
5:50:36context. So it's cached and then all
5:50:37that Claude actually has to process for
5:50:39the first time is reply one and message
5:50:42two and it caches that. So then down
5:50:44here in turn three, all of that's cached
5:50:46and we are bumping up a reply and a
5:50:48message and those are the things that
5:50:49only get processed each time. But if we
5:50:51waited an hour and then we sent another
5:50:53message or if we change the system
5:50:55prompt then everything from the very
5:50:57beginning has to get fully reached. So
5:51:00imagine if you were on message like you
5:51:02know 16 and you're way way way over here
5:51:04on the right and you change the system
5:51:05prompt or you wait an hour then
5:51:07everything getting reached is going to
5:51:10be a pretty expensive move that you just
5:51:11made. So anyways once again we have the
5:51:13system layer, the project layer and the
5:51:15conversation layer. The system layer has
5:51:16instructions, tool definitions, output
5:51:18style, and here's where it might break.
5:51:21The project level or the project layer
5:51:23has cloudmd, memory, and rules, and then
5:51:25here's when that might break. And then
5:51:27we have, of course, the conversation,
5:51:28which is just like the replies and the
5:51:30messages, which gets reached every time,
5:51:32but that's how it should be. So, here's
5:51:34where there's been some confusion among
5:51:36the community. So, how long does the
5:51:39cache snapshot live, which is kind of
5:51:41called the TTL, the time to live. So on
5:51:43your cloud subscription, you have an
5:51:46hour by default because it uses your
5:51:47subscription. But if you go over that
5:51:49weekly limit and you are now playing in
5:51:51your extra usage territory where you are
5:51:53paying per token API, then by default
5:51:56that will be 5 minutes, which is very
5:51:58dangerous. If you're managing multiple
5:52:00sessions and you're constantly reaching
5:52:01everything because five minutes is
5:52:03passing, you got to be careful about
5:52:05that. And people were kind of
5:52:06suspicious. I don't know if you remember
5:52:07like a month or so ago when everyone was
5:52:09complaining about their Clawude uh
5:52:10subscriptions how quick they were eating
5:52:12it up. People thought maybe that they
5:52:14switched the cash TTL from an hour to 5
5:52:17minutes without like saying anything to
5:52:18anybody. It turns out they didn't. So it
5:52:21is an hour but that's just like you know
5:52:23there was a lot of confusion around that
5:52:24and I get why because honestly it's not
5:52:26super clear. Like if you're on an API,
5:52:28you have 5 minutes by default, but you
5:52:30can increase the cost and you can do an
5:52:33hour and then your sub agents on any
5:52:34plan are going to be 5 minutes. And for
5:52:36some reason, all of this is documented
5:52:37about cloud code and the API, which are
5:52:39two very different things. But the
5:52:41claw.ai like on the web, we don't know
5:52:44exactly how that works. At least I
5:52:45haven't found documentation on that
5:52:47exact. I'm assuming it's the same as
5:52:49your subscription, but I don't know 100%
5:52:51for for fact. Anyways, three habits that
5:52:54cover 95% of people. Don't pause too
5:52:56long. So if you've gone over an hour on
5:52:59a session, just hand it off to a new
5:53:01session. Obviously, start fresh when you
5:53:03switch tasks. So do a slash compact,
5:53:05which will break the cache, or do a
5:53:07slash clear. Or you can also use my
5:53:08session handoff skill, which I will
5:53:10include as well for free. So both the
5:53:13token dashboard, GitHub repo, and the
5:53:14skill will be in my free school
5:53:16community. The link for that down in the
5:53:17description. But basically what that
5:53:18means is, let's say right here, I've got
5:53:20this project which helps me build this
5:53:21HTML file you guys are looking at. It's
5:53:23got 205,000 tokens in here. And if I
5:53:25come in here and just do a session
5:53:26handoff, this basically summarizes
5:53:28everything we've done, all the important
5:53:30files that we've built, all of the open
5:53:31decisions, exactly where to pick back
5:53:33up. And then I basically am able to just
5:53:35copy that summary, do a slash clear, and
5:53:38then keep going. And it feels like I
5:53:39haven't actually lost anything. So that
5:53:41has been basically my replacement for
5:53:42doing /compact. I've just enjoyed doing
5:53:44this better. And sometimes the compact
5:53:46takes a long time. This typically
5:53:47doesn't take anywhere over a minute.
5:53:49There you go. So that is my session
5:53:51handoff. I do a /copy and then I just go
5:53:53ahead and clear that. Paste it in, hit
5:53:55enter, and now I'm basically right back
5:53:56where I was. And then this last one is
5:53:58for if you're using Claude chat
5:54:00specifically. If you're going to be
5:54:01pasting big documents in there, you're
5:54:03probably better off doing a project
5:54:04because like I said, I don't know
5:54:05exactly how the caching works in Cloud
5:54:07Chat. But we do have some confidence in
5:54:10saying that projects, those files are
5:54:13cached a little bit differently and
5:54:14probably more optimized for storing a
5:54:16bunch of documents compared to just
5:54:17dropping them into your cloud chat. So
5:54:19keep it alive, keep it focused, and
5:54:21start fresh when you switch. Now,
5:54:23there's a few other things that were a
5:54:24little bit confusing to me as far as
5:54:26like what breaks the cache. So the first
5:54:28one is if you switch the model. So you
5:54:31know, if you're in here and you're
5:54:32talking to Claude, hello, hello, hello.
5:54:34And then you go in here and you do a
5:54:36/model and you actually switch the
5:54:37model, that's going to reach everything.
5:54:39Because if you remember earlier, I said
5:54:40it's prefix matching, which I'm not
5:54:42going to dive into right now. But if you
5:54:44switch the model, then you are switching
5:54:46essentially the prefix and it can't
5:54:47match on that same cache. So if you
5:54:49switch the model, you are reaching
5:54:51everything. Now I do want to apologize
5:54:53for something here because if you do
5:54:54model opus plan, which is something that
5:54:57I've shown before in like token hacks
5:54:59videos, this basically means it uses
5:55:01opus for plan mode and then it switches
5:55:03to sonnet for the execution. But if you
5:55:06do that, just keep in mind that's
5:55:07actually going to break the cache
5:55:08because you're switching model halfway
5:55:10through. So right here you can see each
5:55:12model has its own cache. Switching with
5:55:13model means the next request reads the
5:55:15entire conversation history with no
5:55:17cache hits even though the context is
5:55:18identical. The opus plan model setting
5:55:20resolves to opus during plan mode and
5:55:22sonnet during execution. So each plan
5:55:24toggle is a model switch and starts a
5:55:26fresh cache. So it's very interesting
5:55:28because typically the point of that is
5:55:30to save your session limit and I think
5:55:31ultimately long run it does but it is
5:55:33important to understand that doing that
5:55:34does reset the cache. Now what you can
5:55:37do is you can edit your cloud. MD and
5:55:40you can do that mid session because the
5:55:41edit actually doesn't apply until you
5:55:43restart that session. So the cache stays
5:55:45safe. And then of course the cloud.AI
5:55:47projects caching. It's not exactly
5:55:48documented but pretty confident that it
5:55:51does help to drop docs in projects
5:55:52rather than in the chat. But anyways
5:55:54this token dashboard like I said is very
5:55:56helpful to just be able to understand
5:55:59get a little bit more visibility into
5:56:00your tokens. This does track your tokens
5:56:03on a local device. So, if you switch
5:56:05over to a laptop, then your dashboard is
5:56:07going to look different than on your
5:56:09main like PC or whatever you use, but
5:56:11it's very, very simple. It is a GitHub
5:56:12repo. You will go to my free school
5:56:13community. The link is in the
5:56:14description. You'll click on classroom.
5:56:16You'll click on all YouTube resources,
5:56:17and then you'll be able to find it right
5:56:19in there. And once you get that GitHub
5:56:20repo, all you have to do is give the
5:56:21link to cloud code and say, "Hey, this
5:56:23is a token dashboard. Set this up on a
5:56:25local host." Boom. You've got it open.
5:56:27And it will pull in all of your past
5:56:29sessions. So, it's not like you're going
5:56:30to start fresh. As soon as you uh, you
5:56:33know, link in this repo, it will read
5:56:35your past files and it will pull in your
5:56:36tokens. And then, of course, I will also
5:56:38include that session handoff skill that
5:56:40I just mentioned to you guys. So, I know
5:56:41this one was super quick. Hopefully,
5:56:42this one was helpful, though. Um, it's
5:56:44just important. Like I said, when I hear
5:56:47about stuff like this, I love to
5:56:48understand it to the point where I know
5:56:49how to use it and I know what's going on
5:56:51under the hood. But truthfully, if I
5:56:53looked at some of these other articles,
5:56:55like how in-depth they go and how much
5:56:57nuance there is, most of the stuff right
5:56:59now, I just don't need to know because
5:57:01I'm not using the the API in this way
5:57:03super heavily. So, the reason I wanted
5:57:04to throw that out there is because it's
5:57:06important to stay updated and follow
5:57:07things, but just understand what do you
5:57:09really need to know at its core. Okay.
5:57:13So, we've covered a lot in this course
5:57:14and I hope that you guys are kind of
5:57:16proud of how far you've come since the
5:57:18beginning where maybe you didn't even
5:57:19know what cloud code was up at the
5:57:21front. We've covered a lot of
5:57:22fundamentals. We've gone over a lot of
5:57:24different kind of skills and different
5:57:26things that you need to be aware of. And
5:57:27now we're really getting into the part
5:57:28where it's time to build your AI
5:57:30operating system. You've already kind of
5:57:32been building up your context and your
5:57:33second brain and your knowledge, but now
5:57:35you need to have an operating system
5:57:36where you actually are able to leverage
5:57:39all of that knowledge and become 10
5:57:41times more productive. And all of this
5:57:42is built on basically the four C's of
5:57:45building an AIOS as I like to call them.
5:57:47We have kind of these two, which is way
5:57:49more about the second brain element,
5:57:51context and connections. And then we
5:57:52have the capabilities and cadence. This
5:57:54is basically like the skills, the
5:57:56automations, the agents, and the
5:57:57routines and the deployments that
5:57:59actually make this thing work while you
5:58:00sleep. So, we've already spent multiple
5:58:02hours together in this course. So, I'm
5:58:03going to send you guys over to a
5:58:04different one. It's still completely
5:58:05free, and that is in my free school
5:58:07community. So, once you guys join this,
5:58:08or if you already have to grab all the
5:58:10skills and stuff, you're going to go to
5:58:11the classroom, and you're going to do
5:58:12the build your own AIOS course. It's a
5:58:152-hour course. It's going to walk you
5:58:17through the different mindsets around
5:58:19the AIOS and how you go from where you
5:58:20are right now to where you want to get
5:58:22to, which is probably something similar
5:58:24to me. But mine is still a work in
5:58:26progress. Every day, every week, every
5:58:27month, I'm adding so much stuff and I'm
5:58:29building it out. And this will never be
5:58:30a finished product. It's just going to
5:58:31be the way that I always work. And from
5:58:34there, if you guys enjoyed this course
5:58:36and you want to learn more and you want
5:58:37to connect with people who are building
5:58:39businesses or, you know, trying to get
5:58:40AI roles in companies, then definitely
5:58:42consider checking out our plus
5:58:44community. But anyways, that is going to
5:58:46do it for today. So, if you guys
5:58:48enjoyed, if you learned something new,
5:58:49please give it a like. It helps me out a
5:58:50ton. And I really, really appreciate you
5:58:51guys making it to the end of this course
5:58:53with me. And I hope to see you guys
5:58:54around in my communities or in future
5:58:56videos. Thanks so much, guys. Take care.

About this transcript

This page contains the full transcript of Claude Code for Non-Coders (6 Hour Course) by Nate Herk | AI Automation, generated from the public captions YouTube serves with the video. The transcript has 85,931 words across 11,730 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.