Give Me 50 Minutes, I'll Give You 1000+ Hours Of Claude Code Knowledge (2026 Guide)

CChase AI
컴퓨터/소프트웨어AI/미래기술

스크립트

00:00:00This video is going to replace over 1,000 hours of trial and error inside of Cloud Code
00:00:04and tell you what you need to focus on, whether you're at the beginning stage,
00:00:08intermediate stage, or pro. I'm going to cover everything from the desktop app
00:00:12to how to prompt, MCPs, CLIs, and even more advanced topics like graph engineering and
00:00:18creating your own agentic OS. And by the end, you're going to have a complete roadmap on how
00:00:22to master the most powerful AI tool in the game today. There is a ton to cover, so let's get
00:00:27started. Now, the first thing we need to talk about in the beginner section is where do we run
00:00:32Cloud Code? Because this can actually be a confusing question. You technically can run Cloud Code from
00:00:37the web app in the cloud. We also have the desktop version of Cloud Code. And of course, we have the
00:00:43terminal. Which of these three should you be using, especially if you're just getting started? Well,
00:00:48if you'd asked me this question a few months ago, I would have pushed you to the terminal or something
00:00:51like VS Code with the Cloud Code extension. But these days, the desktop app has gotten a
00:00:56lot better. And if you are someone who is not coming from a technical background, this is all
00:01:01brand new to you, I would suggest using the Cloud Code desktop app. Yes, the terminal will always be a
00:01:07powerful option for those of you who are more technically inclined. But if that's not you, don't
00:01:11feel pressured into using it if you're not comfortable. You really aren't missing out these
00:01:16days. And the desktop app is getting features that just aren't available inside of the terminal.
00:01:21Things like voice mode, things like browser automations. And there's a lot of quality of life
00:01:25things when you work inside the desktop app, like inline artifacts, and generally a better user
00:01:31experience for those of you who have never touched the terminal before. Not to mention, we can also use
00:01:36the terminal when we're inside the desktop app. So it's not like it's an either or thing. Now to
00:01:40download and install the Cloud Code desktop app is super easy. Just search for Cloud desktop app,
00:01:44you'll hit the first link, and then you're going to download the installer and run it. So now you have
00:01:48the desktop app installed. And let's do a speed run through this thing and what you need to actually care
00:01:53about. Now in terms of settings, you come over here to the left and hit customize. Let's then go to the
00:01:57general tab. And I want you to take a look at instructions for Cloud. Take a look at mine. It's
00:02:01blank. Yours probably should be blank too. These are global instructions. By global, this means it applies
00:02:06to everything, all your projects and all of your prompts. So if I put something in there, like I want
00:02:11all of your responses to me to rhyme, well, guess what's going to happen? Every single response is going
00:02:16to rhyme. Stupid example. But the point is, do you actually have something that's relevant enough to
00:02:21every single chat you're ever going to have with Cloud that we should spell it out here? Maybe,
00:02:26but that's a very high bar. And if what you're thinking about adding here doesn't meet that bar,
00:02:31I would leave it blank. Next, go to capabilities. Make sure tool access mode is load tools when needed
00:02:36and turn everything else on this page on. In the Cloud Code tab, I would keep everything in general turned
00:02:41on. It has some settings that are purely personal preference. And in terms of local sessions, I would
00:02:47turn all of these on as well. Now, the exception is when we come down to pull requests. If you don't
00:02:52know what a pull request is, just leave this off. It's a little more advanced. And for Cloud in Chrome,
00:02:58which is a Google Chrome extension where Cloud can control things in your browser, I turn this on as
00:03:03well, but understand you need to download the Google extension if you want to do that. Now we'll go into more
00:03:08detail when it comes to skills, connectors and plugins and memory later. So don't worry about that just yet.
00:03:14Now over here, we have artifacts. You don't really need to worry about artifacts, honestly, inside of
00:03:18Cloud Code. Routines have to do with automations we run and we'll go deeper into automations later.
00:03:23And over here, this is just previous chats we've had with Cloud Code. Now if you go ahead and hit new,
00:03:29it should bring up a page that looks like this. And let's break down what we're looking at. So obviously,
00:03:33right here, we have our little chat window and then we have these four things up here.
00:03:37Local, something that says 1000, main, work tree, and then a little plus button.
00:03:42Local is just you telling Cloud where it's actually going to be running. If you don't know what any of
00:03:47these things are, Cloud, remote control, WSL, SSH, you should be on local 99.99% of the time.
00:03:53Everything else here besides WSL is all about being able to use Cloud when you're away from your computer.
00:03:58So if you're in the beginner stages, that's not going to be you. Next, right here where it says 1000,
00:04:03that is just the folder I am working in. So if I click on this, I can open a new folder and I can
00:04:08pick any one inside of Cloud, well, inside of my computer that I want to work within. So you could
00:04:13create a folder on your desktop that just says Cloud Code projects. And that's what you work in.
00:04:18Everything you do here inside of Cloud Code will live in that folder. So you just have to pick one.
00:04:23Next, we have main and work tree. This has to do with Git, a little bit more of an advanced
00:04:26topic. So if you're a beginner to oversimplify this completely, Git just has to do with saving your
00:04:32work. So if you don't know what Git is, I'm not going to turn this into a Git lesson.
00:04:36Just keep it on main and don't check work tree. And then we have this plus button. This allows you
00:04:40to add an additional folder. So your work essentially gets copied in two places.
00:04:44Next, we have permissions. If I click on auto, I will see five different modes. What are these modes?
00:04:48This has to do with us telling Cloud Code what it has permission to do with or without our consent. So
00:04:55on one end of the spectrum, we have manual. This means it's going to ask you all the time. Can I do
00:04:58this? Can I do that? Can I edit things? On the other end of the spectrum, we bypass permissions.
00:05:03It can do whatever it wants, download things, install things, delete things, edit things. Kind of scary.
00:05:07Now in between those, we have auto, which is essentially bypass permissions, except there is
00:05:12a classifier that takes a look at the commands that Cloud Code is running and decides if they're dangerous
00:05:17or not. And if they're dangerous, it will stop them. So this is the default for a reason. You should be
00:05:22sitting here all the time. The only other one we're going to play around with is plan and we'll go into
00:05:26that more later. This plus button lets you add things so you can add screenshots, that sort of
00:05:30thing. And then you have a microphone. And then over here on the right, we have the model, we have
00:05:34the effort level, and then we have our context window. What model should we be using? Well, it depends
00:05:39on what plan you're on. If you're on the $20 a month plan, you really aren't going to be able to
00:05:42use Fable. It's just going to burn too much usage. And so you're going to have to sit on Opus.
00:05:46If you are on a max plan, 5X, 20X, so $100, 200 bucks a month, I would suggest working in Fable
00:05:51most of the time. It is the best model by far. Now, the problem is the usage though,
00:05:58because if we click this little thing over here, you know, you see context window. We'll talk about
00:06:03that more later. We have a number of different limits. We have a five hour limit, a weekly limit,
00:06:07and a Fable limit. So only half of our usage per week can be dedicated to Fable. So we don't want to
00:06:14burn too much Fable too early. So what we need to really think about is our effort level, which can
00:06:21be anywhere from low all the way to ultra code. The more it thinks, the better it performs, but it's
00:06:27not linear. If I go from extra high to ultra code, does that mean I'm going to get like a 10X increase in
00:06:35the performance? No. It might be like a 1% increase, but you might pay 5X more. In fact,
00:06:39most of the problems you're trying to solve, especially if you're more in the beginner stage,
00:06:43don't require anything beyond medium. In fact, you could probably get away with low and just do fine.
00:06:48Me personally, I sit on Fable medium most of the time, unless I'm dealing with something rather
00:06:52complex or I'm about to hit a usage reset and I know I can just burn it. So for today, we're going to
00:06:57sit on Fable 5 medium because it's a nice, happy middle ground. So lower effort means less usage,
00:07:03doesn't perform quite as well, but generally it does more than enough.
00:07:07Now let's talk about prompting. Anytime we are going to start prompting some new project,
00:07:12I highly, highly suggest you go into plan mode. Now, why is that? Well, plan mode is going to allow
00:07:18Claude and I to have a conversation to make sure we are on the same page before it goes and execute
00:07:22something. And more importantly, not only are we going to have a conversation, it's going to ask you
00:07:27questions. Because many times, and this is very true if you are coming, A, from a non-technical
00:07:33background, and B, if you're trying to do some sort of project that is not in your typical domain of
00:07:38expertise, the issue you're going to run into when using AI is that there's just so much that you don't
00:07:43know. And there is so much that you don't know that you don't know. These unknown unknowns are a real
00:07:49problem. And the only way you're going to be able to figure those out is by having Claude code pretty much
00:07:54bring them up for you to shine a light in these dark spaces that you didn't even know exist.
00:07:58And plan mode is the simplest way to solve that problem, because it's going to force it to ask
00:08:02us questions. So anytime you're like, have a fuzzy idea of I'm on step A, I want to get to step Z,
00:08:07and I don't know what to do, we're going to go into plan mode. Now, when I prompt Claude code,
00:08:12this used to be a big thing, kind of back in the day, the last like year or two, and this has kind
00:08:15of gone away a little bit. But you still have people out there who think there's some like magic prompt
00:08:19you need to do, it needs to go in this specific format. We're like, here's the goal,
00:08:22and here's the context. And here's how I want you to act. You don't need to do that.
00:08:25What you need to do is you need to buy yourself a microphone, you need to turn the microphone on,
00:08:29and you just need to give Claude code a stream of consciousness. So for this plan for this website,
00:08:33we're going to build, we're going to say it's just a website for this fake AI analytics company,
00:08:41and we'll call it Lighthouse. And so that's all I'm going to do. And I'm just going to ramble.
00:08:46Like there's no plan here. It's gonna sound like this. So I want to create a website for a fake AI
00:08:54analytics company called Lighthouse. I don't really know what I want to be on the website,
00:09:02although I know I want sort of the call to action at the end for them to book a call with us. So that's
00:09:07going to be kind of the call to action. I think in terms of who the target audience is, let's say it's
00:09:12going for like small startups. So other than that, I don't really know what I'm missing. Just go ahead
00:09:18and ask me whatever questions you think are relevant that I haven't thought of. Now what I want you to pay
00:09:22attention to is that last bit where I said pretty much, just start asking me questions and things I
00:09:27haven't thought about. Now, since it's in plan mode, it's kind of already going to do that. But anytime you
00:09:31prompt Claude, you can always add that at the end of your prompt. Like, what am I not thinking about?
00:09:37What sort of questions do you have for me? You know, and this is going to get you again,
00:09:41and sort of that back and forth with Claude code, because at any point it can bring up something like
00:09:44this where it asks you questions. So what's the purpose of this fake site? Well, we'll say it's for
00:09:49design slash dev practice. What does Lighthouse actually do? We're going to say product analytics and AI
00:09:56insights. How big should it be? Let's do a landing page like that. And then what does vibe should the
00:10:06design have? We'll go with clean light SAS. Next, it asks us what tech stack should we want the site
00:10:12to be built with plain HTML, CSS slash JS, next JS plus tailwind or ash plus tailwind? Do you know what
00:10:18any of those are? Genuinely, do you have any idea what those are? Do you even know what a tech stack
00:10:23is? If your answer is no, then what should we do? Should we just go with the recommended?
00:10:28Yes, but no. So here's a problem a lot of people run into is they're just going to be like, I don't
00:10:33even care what a tech stack is. Sure, plain HTML, go recommend it. And they just click this.
00:10:38And the thing is, these models are so good is that you're still going to get a pretty good output.
00:10:41The problem is, is when you repeat that behavior of not understanding the question and just hitting
00:10:45recommended again and again and again and again. The main problem is there's no differentiation
00:10:50between you and any guy on the street who I could put in front of your computer and have
00:10:54them do the exact same thing. Like where is any moat whatsoever for what you do? Like you're very
00:10:59replaceable. But even more importantly, you're not learning anything. And as good as these models are,
00:11:04eventually you're going to have a pet project that's very unique that maybe Claude doesn't know the
00:11:08best way to attack it. If you have spent all of your time learning Claude code by never actually
00:11:13learning anything and just hitting recommended, recommended, recommended, you're going to have
00:11:17zero idea of what's actually happening. Now you don't ever need to learn code again, but you do need to
00:11:23begin to learn AI software engineering fundamentals, like big picture stuff, like how these building
00:11:30blocks come together. And the only way you're going to do that is when you hit questions like this,
00:11:34you don't just hit recommended, you say something like this. Can you explain what a tech stack actually
00:11:40is? I don't really know. I don't really understand these options as well. So can you just give me a
00:11:44quick breakdown of what I'm looking at. And that's it, you're just going to tell it to explain the
00:11:49question a little more detail. And if you do that over and over and over again, for weeks, months and
00:11:53years, you're eventually going to build an actual foundation, you're not going to be this like,
00:11:58caricature of a vibe coder. This is extremely important. And this is the sort of mentality of the
00:12:02take with prompting and talking with Claude code if you actually want to like, learn anything,
00:12:07because it's going to make you better. So we'll do pricing product features. Sure.
00:12:12What should the book call CTA work? We'll do a fake booking form.
00:12:21It's off to the races. And so we can see here, it broke down the what is a tech stack question,
00:12:27and then goes into a little more detail about what everything is. Again, you don't need to become an
00:12:31expert at all these things. But as you do this again and again, and you kind of go down the rabbit
00:12:35holes when they make sense for you, you generally are going to start like putting the pieces together.
00:12:39Like, this isn't that complicated. Like coding is difficult. No one's asking you to become like a
00:12:47true software engineer in that sense. But you can learn what a tech stack is, you can learn what these
00:12:52different languages sort of are and like, what we should do in each sort of use case. Now once it
00:12:56proposes the plan, you will see it populate over here on the right hand side. Now what's cool about
00:13:01plan mode inside the desktop app is when it brings us up at any time, I can sort of select something.
00:13:07So let's say the audience, let's say I didn't just want small startups. Let's say I was also looking
00:13:13for add, I wanted to add medium sized companies as well. So if I do that and I hit comments, it now
00:13:22starts adding sort of like comments here. So I could add more and more comments and go down the line and
00:13:27kind of make these like almost little notes to the plan it's created. I also have the ability to add
00:13:31any prompts I want. And then I can at any point say, hey, let's go ahead and revise this. So now it's
00:13:37added medium sized companies to the audience as well. And once I like the plan, I can either hit accept
00:13:43or accept with auto mode. So make sure you hit accept in auto mode or else it's going to start running it
00:13:48inside of manual mode and we don't want to do that. So we'll do accept in auto mode. And now it has gone
00:13:52ahead and built a webpage for us. And it's at this point, we are going to move into the intermediate
00:13:57sort of skills and tips and tricks of when it comes to cloud code. But before we do that,
00:14:01a quick word from today's sponsor me. So just yesterday, I released a completely updated version
00:14:06of my cloud code masterclass. We go way deeper on all the topics we sort of touch on at a surface level
00:14:12here in today's video. And it is a perfect place for you if you're someone who is non technical and
00:14:17really just wants to learn how to master this amazing tool. So if you want to get your hands
00:14:22on it, you can find it inside of chase AI plus there's a link in the pin comment.
00:14:27So let's build our website. But what I actually want to focus now is on this guy down here,
00:14:31this little circle. Remember this that showed our usage? Well, it also shows our context window now.
00:14:37And if we click on our context window, it gives us a very specific breakdown of what is actually
00:14:41filling it up. Now, the context window is a very important metric we always need to keep track of
00:14:47for a few reasons. Now, the first reason has to do with performance. But to understand this,
00:14:51you need to understand tokens and context. So to keep this oversimplified, every single word you send
00:14:57to cloud code and every single word you get back is considered a token. A token is the currency of large
00:15:03language models. And the context window is the budget. So we have a budget of 1 million tokens we can
00:15:09essentially spend in each session. And so far, we have spent 156k. Now, that sounds great. We've
00:15:16only used 15%. I can use another 844,000. Well, sort of. The issue is as this context window fills up,
00:15:25in fact, the performance of cloud code gets worse. Think of it as it just having too much stuff in its
00:15:31brain. If we are sitting at 800,000 tokens over here, out of 1 million, and I ask it questions about
00:15:38what's been going on in the last 800,000 tokens, it's going to struggle, especially if we're asking
00:15:43it questions that sort of happened in the middle. So because of this, we always want to keep an eye on
00:15:49our context window, because we don't want the performance to decrease. And this decrease is
00:15:53somewhat linear, and there isn't an exact science to it. So rule of thumb, as we sort of hit like 30%,
00:16:0040%, certainly 50%, 500k tokens, you definitely want to sit there and ask, do I need to continue
00:16:06the session? And really, for me, that's at like 30%. And while you can see here, the context window
00:16:11does get filled up with things besides our messages, including things like system tools and skills,
00:16:15the messages are the big thing. Let's say you hit 30% of your context window, or even 50%. What are your
00:16:22options? Well, we really have one option, and that's just to start a new chat. Okay, we're just going to
00:16:31start a new chat. Now, there's a couple ways you can do this, we can do commands like forward slash
00:16:39clear. If I do forward slash clear, this is going to get rid of everything, and we're going to start
00:16:43completely fresh, brand new context window, and we're going to have top performance. Now, the other option
00:16:48is to do slash compact. If I do slash compact, what Cloud Code is going to do is going to take a look
00:16:53at the entire conversation history we've had, it's going to create a new summary, and then it's going
00:16:57to start a new chat with that summary. Now, your other option is just to go over here and hit the
00:17:02plus button. And this will also start a new chat inside of that same folder. And then I can reference
00:17:06the old chat at any time. Now, the scary thing is, especially if you've come from primarily the web app,
00:17:11where the conversation you have is sort of all you have. And if you get rid of the conversation,
00:17:16forgets everything, remember, what are we doing here, we are working inside a specific folder,
00:17:20we're inside folder 1000 creating our website. So if I get rid of this entire conversation,
00:17:26it can still take a look at the files at all the code we wrote and understand what's happening.
00:17:31So starting a new chat isn't really starting from zero. And in that case, you really have nothing
00:17:36to fear. If you're filling up the context window, and you're scared of starting over, it's okay,
00:17:40just start over and worst case, just have it create that summary with slash compact. But at this point,
00:17:45we're only at 16%. So we are okay. Now let's talk a little bit about this website. This website
00:17:53is ugly. This website is actually pretty lame and generic. And by the way, I'm looking at this inside
00:18:00the browser pane inside of Claude Code desktop app. So I can actually do a lot of stuff here. Like if I
00:18:05open this up, I can select certain things, I can put comments on these just like in plan mode,
00:18:11which then go into a prompt, I can actually annotate things and say like, Hey, trash, you know, and it
00:18:17will also add it to a comment. So really easy to sort of do like micro edits here. If you want to, but big
00:18:24thing we need to solve for just like this website, why does this look so terrible? Well, this looks pretty
00:18:28awful. Because a we didn't give it enough context for what we wanted to look like, we didn't give it any sort of
00:18:35inspiration, we didn't give it screenshots. All we said was we wanted like a clean SaaS product. And
00:18:39one of the themes of sort of the intermediate section is context engineering, your ability to
00:18:44give Claude Code access, not just to sort of your thoughts and your vision, but for external tools and
00:18:50skills so that it can do a better job. So sort of a two part process here for solving this problem.
00:18:55This is where we're going to dive into the idea of skills. Skills are probably the most important
00:19:00thing you need to understand and master when it comes to leveling up your Claude Code performance.
00:19:06Now skills at their most reductive state and it gets a little bit more complicated are simply prompts
00:19:11that tell Claude Code to do a specific thing in a specific way. For example, there are a ton of front
00:19:17end design skills that are all about making better websites that are essentially just prompts that tell
00:19:22Claude Code, hey, when you're creating this website, avoid certain gradients, avoid things that look like
00:19:27AI slop, do this, do that, right? It's just giving it specific instructions. That's all skills are.
00:19:32So how do we actually get skills? Well, we can find them inside of the Claude app. So if I go to customize,
00:19:37move this over here, we go down to skills. And you can see some of the skills that I have right here,
00:19:44we also have plugins, which also can kind of be skills, it's sort of a gray area, especially app,
00:19:49we talk about skills versus plugins, you can kind of think of them as the same thing. Plugins can
00:19:54include multiple skills, but it's kind of arbitrary. So for example, if I go to plugins, and I go to
00:19:59browse, first of all, what am I going to see, I'm going to see the anthropic official plugins, this
00:20:04includes the front end design plugin, which is simply the front end design skill. So if I install this,
00:20:10mine's already installed, it will then add the front of design skill to Claude Code. And this is the
00:20:15actual prompt for the exact front of design skill. This is an official skill, like you can just go see this,
00:20:19this is on the official Claude Code GitHub. If I copy this whole thing, and I go back inside here,
00:20:24and I paste this into the prompt, that's just like I'm using the skill. But obviously, you wouldn't do
00:20:30that every single time you want to do something related to front of design. So instead, we simply
00:20:34add the skill, like I showed you in the UI, and we just do front end design, right, I can do forward
00:20:40slash, and now this is invoked. Now that's the same as if I just copied and pasted that whole thing.
00:20:44Now I don't only have to do the forward slash, I could just use natural language and say like use
00:20:49front end design skill. And it's smart enough to actually know it needs to call that. Now the
00:20:53confusion can happen is if you have multiple front end design related skills. And if you do that, then
00:20:58if you just say like, hey, I'm building a website that it might not know which one to pick. So if you
00:21:02have multiple skills that all kind of do the same thing, you need to nudge Claude in the right
00:21:05direction. Now, the most important skill you can actually add right now, and this goes beyond web design,
00:21:10is if you go back to plugins, and you go to browse, and you go to skill creator, this is the most
00:21:16important skill you can add, because this is a skill that allows you to create other skills and includes
00:21:21things like measuring skill performance, it runs tests, it runs evals, it does benchmarks, and we'll
00:21:27talk about this more in a little bit. But obviously, as we look through here, there isn't that many skills
00:21:31to choose from. And we all know there's a billion in one skills floating out there in the world. And so where
00:21:35you will normally find skills is GitHub, kind of like you see here. So let's say for example, I wanted
00:21:40to use another front end design related skill. And let's say I was looking for the impeccable skill,
00:21:45and I found it on GitHub. Well, how do I actually install this thing? So it will tell you in the
00:21:49description how to do it step by step. But oftentimes, it's kind of a pain in the butt. What you need to do
00:21:53is when you find a skill you like is just copy the URL from GitHub, go inside of Claude code,
00:22:01paste that skill in there and then say something like add this skill. From there, it will literally
00:22:07add the skill to your repertoire. And from there, you just invoke it like I showed you.
00:22:12Now you can get pretty advanced when it comes to skills, especially when you have the skill
00:22:16creator skill. For example, let's say we finish out this video, I add a bunch of new stuff to this
00:22:21web page, I could at the end of doing all this of like, prompting it, creating the website,
00:22:26doing the additions, adding whatever I want, I could then do something like,
00:22:30let's use the skill creator skill and then say, take a look at the entire message history of the
00:22:35session, take a look at everything we've done today and turn that into a skill. So if there's
00:22:39things that you do over and over and over again, you can turn those into skills. And later, I'll even
00:22:44show you how to then turn those skills into automations. So skills are very powerful because
00:22:48they allow you to codify things that AI does. You know, one of the issues with AI is that it's
00:22:53somewhat non deterministic, right? If you ask it to do something 10 times, it might do it 10 different
00:22:58ways. It's not deterministic. However, skills allow us to be somewhat deterministic and give us more
00:23:04control on how cloud does things. So that's why they're so important. Now, like I mentioned before,
00:23:08beyond skills, we also need to add more context here. So what I'm going to do is I'm just going
00:23:12to search for some screenshots that I can add to this website to make it look a bit better. So I went
00:23:16on Pinterest and I found this image when I put in SAS landing page, thought it looked kind of cool.
00:23:20So what we're going to do is we're just going to drop this screenshot into here. And we're going to say,
00:23:25use the front end design skill to redesign this webpage. And in fact, I want you to do three
00:23:30versions of it and show me all three versions inside the browser pane that I can choose from.
00:23:37All of them should kind of be in the style, but I want it to be divergent enough that I can kind
00:23:40of see some differences. So it did exactly what we asked. If we look at this here, we can now see the
00:23:45three different versions of the website. You can see what a departure these are from what we were just
00:23:50looking at. And that was with a single skill, which was just the generic front of design skill from
00:23:55Anthropic and a pretty basic prompt alongside the screenshot. So I can take a look at this one,
00:24:03V1 full size. This one looks pretty cool. We have this V2, not a huge fan of the colors,
00:24:11although the radar looks kind of cool. And then lastly, we have V3, which kind of looks like your
00:24:17typical AI slop, to be totally honest. I really like V1. I think this is something that looks
00:24:24pretty cool. So what we're going to do is we're going to say, hey, we're just going to go ahead
00:24:26with V1. What that's meant to demonstrate is really the power of just injecting context,
00:24:31literally one screenshot, one skill, infinitely different end result. So now let's talk about
00:24:36how we can supercharge cloud code by bringing in outside tools, by connecting outside applications
00:24:42to cloud code itself so that cloud can control them. And we never even have to really leave
00:24:47the app. Now there's really three different ways to do this. The first one is by going into customize
00:24:53and heading to connectors. Some of these are super easy to connect and you might have done already
00:24:57done it already. That's something like Gmail or Google Calendar or Google Drive. This allows cloud
00:25:02code to talk with these applications, control these applications, usually with some guardrails,
00:25:06and it just happens via prompting. So if I tell cloud code, hey, go read my Gmail, it does that because
00:25:11it's been connected. Now, most of the big apps out there have some sort of connector. So if you just
00:25:16go to add and you browse connectors, chances are you will find what you are looking for. The second way
00:25:22is via plugins. And like I said, there's very much like a fuzzy line between all these things. Similar to
00:25:28connectors, if it's big, if it's popular, there's probably a plugin for it. So if I go to browse,
00:25:34you're first going to see a bunch of anthropic plugins, which really are just skills. If I go to
00:25:39partners, though, I can find something like GitHub or Superbase. I simply click on it and you'll see
00:25:44what's going on under the hood. In this case, this is the GitHub MCP. Just like with connectors,
00:25:50if I add some sort of plugin, it's going to allow me to talk to control some sort of outside application
00:25:55via cloud code. But there is a third thing that is not an MCP or plugin or connector, and those are things
00:26:01like CLIs. So we have right here the GitHub CLI. So we have a GitHub CLI and we have a GitHub MCP.
00:26:08What is the difference? Well, the difference is a little bit technical in terms of the practical
00:26:13application, what you are going to care about. In many cases, there's not a huge difference. In
00:26:19general, the CLI tends to give you more functionality than an MCP. And oftentimes the CLIs also include
00:26:26skills. So when you are dealing with some sort of outside application, AKA, you're working inside of
00:26:31cloud code and you need to talk to something else, you need to figure out if you can add it through one
00:26:36of those three ways, connectors, plugins, or CLI. Like you saw connectors and plugins, we can do it through
00:26:42here via the CLI. It's also as simple as simply telling cloud code to add the CLI. So like the GitHub
00:26:49CLI, there's an actual command if you're inside the terminal, but I could just like copy the URL.
00:26:54I can go inside of cloud code and say like, "Hey, here's the CLI for GitHub. Go ahead
00:27:01and add this CLI." And it's going to do exactly that. From there, you've pretty much given cloud code.
00:27:08Think of it almost as a skill to call on that CLI and have it do whatever that CLI is of the ability
00:27:14to do. For GitHub, that means I can create a repository. I can upload all this code we just
00:27:19created to that repository. I can edit the repository. And if we take it a step further and we think of
00:27:25something like Vercel, and Vercel, if you are unaware, is a web application that allows us to host
00:27:31our actual website. So we just created a website and we want to put it on the web and have a true
00:27:35URL. Vercel will allow us to do that. Well, instead of going to a dashboard like this inside of Vercel
00:27:40and handling all ourselves, why don't I just search for the Vercel CLI? Oh, look at that. Vercel has a
00:27:46CLI as well. And so what we could do with just those two applications is I can take a look to see if
00:27:52there's a GitHub connector. There is. I can take a look to see if there's a GitHub plugin. There is. I can
00:27:57take a look to see if there's a GitHub CLI. There is. Add any one of those. Add it with Vercel. And I
00:28:04now have a pipeline where I can take the website I've created. I can create the repo for it inside
00:28:09of GitHub and then automatically connect that to Vercel. I've essentially created an entire deployment
00:28:14pipeline from Cloud Code that I just have to talk to in plain language. It'll do everything for me.
00:28:19And the big takeaway here is that anytime you are working with anything outside of Cloud, you need
00:28:24to ask yourself, can I actually just control it with Cloud? Because chances are Cloud can actually
00:28:28control it better than you can, especially if you aren't intimately familiar with how that application
00:28:33works. Whether it's a CLI, an MCP or a connector, it doesn't really matter. You just need to pick one.
00:28:38And you also don't even have to go out and find it. For example, if I asked Cloud Code a question like
00:28:44this, "Hey, so I'm thinking about deploying this website. I've heard of things like GitHub and Vercel.
00:28:51I'm not super familiar with those. I've also heard there's CLIs or perhaps MCPs we could use. So do you
00:28:58think you can go look and see if that would make sense for hosting our website? If they do, can you
00:29:03add those CLIs if we need them? And then once they're added, can you go ahead and just set up
00:29:08that deployment pipeline and get it all properly connected?" So we prompt like that where you're
00:29:12just saying like, "Hey, I heard there's some tools out there. Maybe they have CLIs. If they have them,
00:29:16add them, run them." That's all you got to do, right? You don't even really have to be a pro.
00:29:21When in doubt, just ask Cloud Code what the best practice is and if it has a CLI. It'll go ahead and search for it,
00:29:26right? So if I run that, and by the way, I already have these installed, so it's probably going to come
00:29:29back and say, "Hey, I already installed it." All you have to do, if you've never used them before,
00:29:33is create an account and then we'll walk you through the setup. And as these agentic coding
00:29:37harnesses like Cloud Code only become more and more ubiquitous, you're going to see pretty much every
00:29:42app out there come up with some version of a CLI, a connector, and an MCP, which means again,
00:29:47Cloud runs the show. And so with that prompt, obviously it's going to walk you through how to log in.
00:29:51If you haven't done that before, it created the GitHub repo. So all the code from our website is
00:29:55essentially living in the cloud. And it also set up the Vercel connection. So I now have a live URL I
00:30:02can go to, and we can also see it right here. So this has a real URL, Lighthouse site to Vercel app,
00:30:07and I could share this with anybody. And again, did I have to go into GitHub? No. Did I have to go
00:30:11into Vercel? No. All control from Cloud Code. And because it is a GitHub and Vercel connection,
00:30:17any changes I make to the website inside of Cloud Code here, if I want those to be reflected on the
00:30:22live website, I just tell it and it does it. So this deployment pipeline is just one example of bringing
00:30:28outside tools into the Cloud Code fold. Now it's time to move on to some more advanced topics. We're
00:30:34going to touch on things like automations. We're going to talk about how we can best approach long
00:30:39horizon tasks. We'll do that by discussing things like slash goal, loop engineering, graph engineering.
00:30:44We're going to talk a little bit about model routing, how we can bring in other models like
00:30:48codecs and the GPT models into our workflow. And then we'll finish it all off by talking about
00:30:52some more custom type harnesses and UIs we can layer on top of Cloud Code, whether that's something like
00:30:57this or going with an obsidian command center type approach. So let's begin by talking about long
00:31:02horizon tasks and loop engineering. And this kind of bleeds into things like automations. So when we
00:31:07talk about long horizon tasks, we're talking about things like loop engineering and graph engineering.
00:31:11What are we really saying? What we're saying is we have some sort of task, some sort of goal we need
00:31:16to complete, but that this might be something we need Cloud Code to do an infinite amount of times.
00:31:21This might be something that runs every single day. And ideally it is something that
00:31:25not only is going to run every single day, but something that we want to be self improving.
00:31:29So for these long horizon tasks, there's really three parts. We are going to have a trigger,
00:31:37a task, and then some sort of success criteria. Now, not all long horizon tasks are loops. This
00:31:43could simply be some sort of task that you think is going to take two hours, four hours, 12 hours,
00:31:48days for Cloud Code to complete. And you don't want it to just stop every single time it fills up its
00:31:53context window. You want it to keep going and going and going until it completes that task.
00:31:57There is a built in command inside of Cloud Code that does this. It is called slash goal. Now,
00:32:03slash goal is perfect if you have a complex project that you want Cloud Code to complete and you don't
00:32:07want to babysit it throughout the entire process. However, there is one specific thing you need and that
00:32:12has to do with success. You need to be able to define success because slash goal isn't the only thing you
00:32:20pass. You also have to pass it a prompt. And that prompt needs to be what the success criteria is.
00:32:28Okay. Like you have a goal for Cloud Code to do. What do you want it to do? It's not enough for you
00:32:32to explain, Hey, I want you to do X, Y, and Z. No, like what is the end state? Because what's going to
00:32:38happen when we run forward slash goal is it's going to attempt to complete that goal with code or whatever
00:32:44it's going to do. It's going to run its first iteration. It's going to compare that iteration
00:32:49to the success criteria you defined. If it meets the success criteria, cool. It's all done. If not,
00:32:55it's going to pull up a second session, run it again. It's going to go ahead and check the success
00:33:00criteria. Did it work? No. Then it's going to run it again. Now, each time it runs, it's going to take
00:33:05a look at its previous iterations to see, Hey, what worked, what didn't, but it's going to continue
00:33:10sort of this internal loop until it reaches its goal. And this is extremely powerful. It's kind
00:33:16of similar to Ralph loops, if you know what that is. And so it's important that our success criteria is as
00:33:22objective as possible. If I'm just saying, Hey, the goal is to create a cool website
00:33:28that looks cool, right? That looks neat. How does it know what cool is? How does it know what neat is?
00:33:34How can it actually take a look at the end of every single run and say, did I complete this or not?
00:33:39So the more subjective your criteria is, the sort of worse it's going to get, not even necessarily the
00:33:44worst it's going to get, but the less likely that it's going to meet your needs. But with all that being
00:33:49said, goal is definitely a form of loop engineering, but it's one that pretty much has a definite end.
00:33:57We aren't going to do forward slash goal and expect this to run for eternity. But there are things we
00:34:03may want to do where we do want it to run for eternity. And we still want it to act in the same
00:34:07way as goal. We want it to trigger on demand or perhaps on a schedule. We have a task we want it to
00:34:12complete. We have success criteria and we want it to also be self-improving because again, this is
00:34:17self-improving in an aspect because it's always taking a look at its outputs and comparing it against
00:34:23some sort of success criteria. So what happens if we want to create something that does sort of loop
00:34:28forever, right? Perhaps it's an automation that runs every single day that we want to continuously
00:34:33improve upon itself. Now we will continue with these three steps, but next we're going to have to add
00:34:37some sort of logging phase. So now let's take a look at a custom loop for an example. We have clod code
00:34:45right here. What we want clod code to do every single day is we want it to create some sort of
00:34:50morning report for us, some sort of morning brief. I want it to go out on the web. I want it to find
00:34:54AI news for me. I also want it to check my Gmail and at the end, give me some sort of document. So clod code is
00:35:00going to take a look at YouTube. It's going to take a look at Twitter, Reddit, in my Gmail. It's going to
00:35:05grab all that information, scrape all that information, consolidate it, synthesize it,
00:35:09and give it to me in a report. And I want it to do it every single day. So how could we bring loop
00:35:16engineering fundamentals into this? Well, remember, we just have to set up these four things. So what
00:35:23is the trigger going to be? Well, let's say the trigger is every day it's going to run at 7:00 AM.
00:35:29What is the task? Well, the task is just what I described. Scrape these websites, consolidate,
00:35:34create a report. So that's set up. Well, what is the success criteria? Well, this is where it gets
00:35:39difficult, correct? Because this is something that's kind of objective, like what makes a good
00:35:45report a good report. We could add some sort of subjective material there. Like, hey,
00:35:49every single report must have at least five videos from YouTube and five Twitter posts and five Reddit
00:35:54posts. And you must call out X, Y, and Z on the Gmail. So there are some things we can do here,
00:35:59but it's not a simple saying like, hey, you're doing a loop on some Python application and your goal
00:36:03is to get it to this particular speed. And then lastly, we have logging. So every
00:36:09single one of these reports, we could put it in some sort of database, right? That way,
00:36:14Claude code can always take a look at its past work and compare its upcoming work to things we
00:36:19have already done. And this in essence, from a theoretical perspective is loop engineering. Now,
00:36:24what does that actually look like in terms of a practical sense inside of Claude code?
00:36:28Well, step number one would be creating some sort of skill, because what did I just describe here?
00:36:34Well, I just described a skill. We can create a skill that Claude code runs on command
00:36:39or on a trigger, where it does all this stuff where it scrapes information and turns it into report.
00:36:44And we can include in that skill, it's sending all that information to a particular database.
00:36:48So step one to do sort of loop engineering for real would be to invoke the skill creator skill. Remember,
00:36:54I showed you how to do that earlier. And then you would just describe the skills I just did.
00:36:59From there, you would then run the skill manually over and over and over until you got it to a pretty
00:37:03good spot. Once you are happy with that skill, you would make sure you add a language about,
00:37:09hey, I want this to be logged in a database. And every time we run the skill, I want you to take a look
00:37:15at the previous iterations and see if we can do better. That's sort of where the self improving
00:37:20aspect comes in. Ideally, you would be able to score every single one of your previous reports. So it has
00:37:26some sort of objective measure it can base its outputs on, but simply you're first going to turn
00:37:31it into a skill. Once you've turned it into a skill, all you have to do now is turn it into an automation
00:37:36that runs all the time. And this is really simple inside of cloud code because we're just going to
00:37:41turn it into a routine. So if I come here to the left and hit routines and I go to new routine and I go to
00:37:46local, guess what I'm going to do? I'm just going to tell this to run that skill every day at a particular
00:37:53time. So this would be like loop skill, we are running the loop skill.
00:38:02And then the instructions would literally just be run the slash loop skill. Okay, when you're doing
00:38:11this, and you're creating this inside of cloud code with the skill creator, you would also say like,
00:38:14hey, I'm trying to do this to so it's like self improving, I want to the database loop engineering,
00:38:18fundamentals, etc, etc. The great thing about the skill creator skill inside of cloud code is that
00:38:22it's going to do all the heavy lifting for you, it understands sort of like the goal here. And then
00:38:27from there, you just schedule it, right? Ideally, it's probably just gonna be something that's daily,
00:38:31but you can do it hourly, weekdays, custom, whatever you want. And that is the practical application of
00:38:37loop engineering. And that's pretty much all you need to know. Because beyond loop engineering,
00:38:42we then start talking about graph engineering and graph engineering can get a little more complicated.
00:38:48But what's really happening is remember before, when we looked at this, we had all this loop
00:38:53engineering stuff going on. Let's actually just let's undo all this. We are all this loop engineering
00:38:59stuff going on, right? Like, hey, this one guy is scraping everything it's looking, it's creating the PDF,
00:39:04and it's being graded. Well, what if we did all that loop of trigger, of task success criteria,
00:39:12and then grading it, you know, logging the information? What if we did that at every step
00:39:16of the journey? So we had one agent that does this with scraping YouTube. So it has a trigger,
00:39:20it grabs YouTube, it scrapes the data, and then it judges how well it scraped it by looking at its
00:39:25past iterations itself improves. And then we did that with the Twitter poll, and the Reddit poll,
00:39:30and the Gmail poll, and this poll as well. So instead of having a loop for the entire thing,
00:39:36just like one huge loop, we instead had a bunch of like micro loops nested together inside of
00:39:41one run. That's graph engineering. So I did a whole video on that. It can be a little complicated. But
00:39:46at its core, that's all graph engineering really is. It's just a bunch of looping agents that also talk
00:39:52to one another. That's all it is. And for most people, this is total overkill. You don't usually need
00:39:57this. But conceptually, that's how it's working. Now shift the discussion over to dynamic workflows
00:40:02and ultra code. So what exactly is ultra code? How does this different from max effort? Well,
00:40:08ultra code, what it's going to do is it's essentially going to create a custom harness for whatever problem
00:40:14you're trying to solve. What that means in practical terms is it's probably going to spin up a bunch of
00:40:19sub agents to deal with whatever issue you have. This can be extremely effective, but this also can be
00:40:25extremely expensive. One example of a dynamic workflow is forward slash deep research. This is essentially
00:40:34like a prebuilt dynamic workflow. And it is very similar to how deep research works. If you're just
00:40:39on the web app and do deep research. So if I run deep research, what's going to happen is it's going to
00:40:44spawn a ton of sub agents and these sub agents are going to do a bunch of different tasks. So if I ask a
00:40:49question like let's deep research best use cases for dynamic workflows inside of cloud code,
00:40:59it's going to now instead of doing a standard web search where it spawns, you know, maybe like five
00:41:03sub agents to basically do a Google search, it's going to spawn several more than that. I've had it
00:41:08spawn well over 100 sub agents. And these are going to do a number of tasks, it's going to actually go out
00:41:12on the web and scrape data, it's then going to create adversarial agents that take a look at the data we
00:41:18found, and then compare and contrast it to see what actually holds up the scrutiny, then it's going to do
00:41:22synthesis, and then it's going to give me a final report. And so you can see over here, it decided that for the
00:41:28scope, it's going to decompose the question into five search angles, it says it's only going to need five
00:41:33parallel web search agents, which is nice for us since we're on fable. And then it's going to pull
00:41:38the top 15 sources, verify everything with a three vote adversarial check on each claim, and then finally
00:41:45synthesize it. And so we can see it working here over on the right since we spawned six agents and every
00:41:50single agent has like a pretty much a token cost right off the bat, we've already burned 314,000 tokens.
00:41:57So I wasn't joking about that. Now, when you run ultra code, if you don't tell it specifically, like
00:42:02let's say I'm on fable five and running ultra code, it's going to use fable for these sub-agents, which
00:42:07can be a problem. Because what if it did say, hey, I'm going to spawn 100 web search agents. Now when I
00:42:13give it the prompt to get away from something like that, you can specifically say, hey, limit it to
00:42:1720 sub-agents, limit it to 50 sub-agents. Or you can say something like, I want you to use Sonnet
00:42:22for the sub-agents or Opus for the sub-agents. So you aren't necessarily a slave to whatever model
00:42:26you're using at that time. Now Anthropic put out a pretty good blog explaining dynamic
00:42:31workflows. And so what's going on under the hood is that it's writing orchestration scripts that run
00:42:3610 to hundreds of parallel sub-agents in a single session, checking its work before anything reaches
00:42:40you. And here's some examples of different sorts of dynamic workflows. Remember, when you run dynamic
00:42:45workflows in ultra code, cloud code is going to figure out the best one that fits your problem. It might be
00:42:50one of these, it might be something completely different. So classify an act, you give it some sort of task,
00:42:54we have a classifier agent that then chooses the best sub-agent for you. Fan out and synthesize
00:42:59an adversary review. If we kind of combine these two, that's what we're doing with deep research,
00:43:02right? We have some sort of task, find this information, it fans out on the web, gets all
00:43:07the information, and then it also does adversary review to see what actually makes sense before
00:43:12synthesizing it for us. Then we got stuff like generate and filter, we have a tournament style
00:43:16thing where we try different attempts to solve some sort of problem and includes judges. And then we have
00:43:21loop until done, which again, very similar to loop engineering. If we jump back inside here,
00:43:25we can see this deep research run we started earlier included 103 agents and burned 6 million tokens.
00:43:32And this was all on Fable. So you can imagine how expensive this is. Here's a look at the actual report
00:43:38it gave us. And as you can see, pretty deep and also includes 21 different sources. And out of all
00:43:43the examples, I think deep research is the one you're going to use the most. And if you're someone who's
00:43:47about to embark on a pretty complicated project and you really want to get all your ducks in a row before
00:43:52you start even going into plan mode, I highly suggest using deep research. So Claude Co can go out there
00:43:57and see what's what before it starts building. Let's talk a little bit about model routing. And really,
00:44:01that's just how can we bring in outside models into the Claude Code fold? Because some of the big
00:44:07players like ChatGPT, their models are great. Sol 5.6 is awesome. Luna and Terra are extremely token efficient.
00:44:13And one sort of thing you need to keep in mind is AI systems in general, these models in general,
00:44:20don't do a great job of grading themselves. So if I ask Claude Code to grade its own work,
00:44:25I ask Opus to grade its own work, be able to grade its own work, it's going to pretty much always say,
00:44:30I did a great job. So how do we solve this problem, especially when we're having a great work that we
00:44:34ourselves can't grade? It's like outside our domain of expertise. I don't really know if that code is
00:44:38great. Well, why don't we bring in another frontier agent to take a look at our work? And the way we're
00:44:43going to do this is through different skills and plugins. There's actually an official Codex plugin
00:44:47for Claude Code. This is from OpenAI themselves. And this allows you to call on Codex from Claude Code.
00:44:53You simply take this URL, you paste it in Claude Code and say, I want to install this. You can.
00:44:58And from here, you can have Codex do adversary review of the code you've already created. Or you can even
00:45:04use this to have Codex work on specific features of your product. And when it comes specifically to the
00:45:09planning stage, I created a skill called Grill Me Codex, which combines Matt Pocock's Grill Me
00:45:13skill with adversarial review from Codex. What happens is you and Fable talk to one another,
00:45:20you come up with a plan, and then that plan is routed to Codex. And Codex and Fable kind of have
00:45:25a back and forth up to five rounds where Codex takes a look at what Fable created. Codex says,
00:45:29this is wrong, this is why. Claude reacts to it, says, okay, I'll fix that, or I don't agree. And
00:45:35they keep going back and forth until they reach an agreement. And so this kind of solves the problem
00:45:39of these models struggling to grade their own work and gets us a second set of eyes. And so we can feel
00:45:46pretty confident moving forward in complex cases. And this sort of model can be taken even a step
00:45:51further if you're someone who doesn't want to use something like Codex and instead want to rely on much
00:45:55cheaper models or even local models. The whole point of it is you aren't stuck just using Opus, Haiku,
00:46:02Sonnet, and Fable. We can bring in whatever we want, especially if we use it as a skill. And again,
00:46:07you would just use the skill creator skill to do that. And last but not least, we have our custom
00:46:11agentic OS structures like you see here and here. These are all about building a custom wrapper on top
00:46:17of Claude and giving us sort of a visual interface that we can't get elsewhere. And it's ultimately custom.
00:46:24Now, the real value of these things isn't the visual sort of wrapper, even though this is,
00:46:27you know, useful for me. This gives me my social media metrics. I can easily click and get a deeper
00:46:33dive on that. I get my research done every single day all in one place, showing me what's going on
00:46:37in GitHub, Hacker News, et cetera, et cetera. All these buttons are related to specific skills and
00:46:41automations that I can run on demand. And over here, I pretty much have the exact same thing,
00:46:45but it also has a voice mode. But like I said, the true value isn't in these cool visual layers.
00:46:51The real value is in the skill architecture behind it all. The whole idea of these things is turning Claude
00:46:58into essentially your personal assistant or an actual worker in your organization that can do
00:47:03everything you do. So for me, what you're looking at here is essentially all the skills I use in my
00:47:09day to day mapped across all the different domains that are important to me. So, you know,
00:47:13I have things related to memory, things like productivity, which is like Gmail, my calendar,
00:47:18all of that. I have stuff related to research, content, my community, my AI agency, sales,
00:47:22et cetera, et cetera. Every single one of these things are a task I would do on my own manually.
00:47:27Instead of doing them on my own manually, I have now mapped them to specific skills.
00:47:31And if they make sense, I turn those skills into automation. And when we look at that in a big picture
00:47:36sense, that is really what an agentic OS is. It's a series of skills and automations that you have
00:47:41mapped to your daily tasks. Now, how do you build something like this? Well, we've already kind of
00:47:44talked about it, right? This is just building skills. It's just using the skill creator skill,
00:47:49turning on your microphone and giving it a stream of consciousness about what it is you do day to day,
00:47:53week to week, and then asking Claude code, can we turn those into skills if they make sense?
00:47:58And if they do, you do it. And you do that over and over again. And eventually you create this corpus
00:48:02of skills that allow you to automate huge swaths of your life. And behind all that is Obsidian. And
00:48:08Obsidian allows us to easily track everything we do in terms of a markdown file system. Now, Obsidian
00:48:15in itself isn't giving you some like crazy upgrade to what Claude code can do, but allows us to actually
00:48:20track everything and allows it to give it some memory. And this is sort of the basis of like the
00:48:24Carpathi Obsidian RAG system you've probably heard before. Now, as a very quick review of what I mean
00:48:28when I'm talking about an Obsidian memory system, this is just a file structure. This is just a
00:48:33coherent file structure that Claude code lives in. So you understand where things go. This is again,
00:48:38often referred to as the Carpathi Obsidian thing. And it's very simple. You have some sort of folder
00:48:42that your agentic OS lives in. Mine is called the vault. Inside the vault, you have some sort of file
00:48:48structure that looks like this. It doesn't have to be exact, but the idea is in one of those folders,
00:48:52we have the raw section. This is sort of where raw data goes is where research goes.
00:48:56We then have a wiki section and we have an output section. The wiki section is where we take all this
00:49:01raw data and we essentially turn it into different reports or wiki style articles. So imagine
00:49:06I had Claude code go do a bunch of research on AI agents. We'll dump the raw information here and then
00:49:11it creates an article about AI agents under the AI agents subfolder. Now, let's say I wanted to create
00:49:17something from that AI agent type wiki. I wanted to turn it into a slide deck, for example. Well, then that
00:49:23slide deck would go to the output section. And the idea is with this simple sort of model, I am able to
00:49:29deal with a ton of files, potentially hundreds of thousands. And it is set up in a clear way that I as
00:49:35a human being can easily navigate and B, Claude can easily navigate. And if it can easily navigate this
00:49:41file structure, it's going to make it more accurate and ultimately cost less tokens. The key of this
00:49:45whole thing are these sort of like index files at every step of the journey. Every time I go deeper
00:49:50into the file structure, there's essentially an index.markdown file telling me what is going
00:49:55on inside of there. So it's essentially like a table of contents. Now, I have a ton of content that goes a
00:49:58lot deeper into this if that was confusing. But the idea of these agentic OS systems is that this visual
00:50:05wrapper sits on top of that custom skill architecture I just talked about and is sort of buoyed by that
00:50:10obsidian memory layer, which helps you navigate everything you've done and helps Claude code navigate
00:50:15it just a little bit better. And the way these systems work, because it's like, hey, I'm not actually
00:50:19using Claude code here, am I is we use Atlas Claude, which uses dash P. So essentially, a command is sent
00:50:28to the terminal, instead of running like just Claude, it uses Claude dash P, which means Claude
00:50:34essentially runs invisibly in the background. There was some drama about this for a while where Anthropic
00:50:38was saying they were going to charge different rates for that and wouldn't be usage, but that's no
00:50:41longer the case. So these sort of systems are actually just as money efficient as anything else.
00:50:47Now creating agentic or Claude OS like this isn't 100% necessary. But I think it is necessary to sort
00:50:53of create these sort of skill architectures that really is the backbone of it all. And of course,
00:50:59if you want to get my exact setup, I have that inside of Chase AI plus as well. So that is where
00:51:03I'm going to leave you for today, we covered a ton of different topics. And these are the ones that I
00:51:07felt kind of can give you the most bang for your buck in terms of what you should focus on from everywhere
00:51:12from the beginning, all the way to a more advanced level. So as always, let me know what you
00:51:17thought. Make sure to check check out Chase AI plus if you want to get your hands on my Claude
00:51:20code masterclass, which kind of covers a lot of the same stuff we did here, just in much greater detail.
00:51:26Besides that, I'll see you around.

핵심 요약

Mastering Claude Code requires leveraging the desktop application, utilizing plan mode to uncover unknown constraints, and building modular skills and CLI connections to automate complex workflows.

하이라이트

  • The Claude Code desktop app provides built-in browser automations, voice mode, and inline artifacts that are unavailable in the terminal version.

  • Global instructions in the Claude Code settings apply universally to all projects and prompts, requiring a high threshold of relevance to avoid unintended formatting.

  • Claude Code maintains a budget of 1 million tokens per session, and performance degrades as the context window fills up.

  • Connecting external tools like GitHub and Vercel CLIs to Claude Code enables the creation of an automated deployment pipeline controlled entirely through natural language.

  • The slash goal command executes an internal loop that repeatedly tests outputs against defined success criteria until the objective is met.

타임라인

Choosing and Configuring the Claude Code Interface

  • The desktop application offers features like voice mode and browser automations that do not exist in the terminal.
  • Global instructions should remain blank unless the rule applies to every single prompt and project.
  • Default permissions use an auto mode that runs a classifier to block dangerous commands while bypassing manual approval.

Non-technical users benefit significantly from starting with the desktop app rather than the terminal. Users can select specific folders for projects and manage token limits across models like Fable and Opus. Configuring effort levels helps balance output quality with weekly usage restrictions.

Prompting Strategies and Plan Mode Execution

  • Plan mode forces Claude to ask clarifying questions before executing commands to expose unknown constraints.
  • Micro-edits and comments can be directly applied to generated plans within the desktop application interface.
  • Understanding foundational software engineering concepts prevents reliance on default recommendations.

Using voice input for a stream of consciousness prompting style eliminates the need for rigid prompt templates. Plan mode enables a conversational back-and-forth that uncovers requirements the user did not initially consider. Accepting the plan in auto mode initiates the code generation process.

Managing Context Windows and Utilizing Skills

  • Model performance degrades as the 1 million token context window fills up during long sessions.
  • The slash clear and slash compact commands reset or summarize conversation history to restore top performance.
  • Skills function as specialized prompts that instruct Claude Code to execute specific tasks in a precise manner.

Monitoring token consumption prevents performance drops caused by bloated session memory. Adding visual context, such as design screenshots alongside frontend design skills, dramatically improves the visual quality of generated websites. The skill creator tool allows users to build and test custom skills.

Connecting External Tools and Deployment Pipelines

  • Command-line interfaces, model context protocols, and connectors allow Claude to control external applications.
  • GitHub and Vercel CLIs enable automated repository creation and web hosting directly from natural language prompts.

External integrations expand the capabilities of Claude Code beyond local editing. Users can establish complete deployment pipelines without manually navigating external dashboards. Claude manages authentication steps and links code repositories directly to live hosting services.

Advanced Loop Engineering and Dynamic Workflows

  • The slash goal command runs iterative sessions against defined success criteria until a complex task is completed.
  • Routine automations combine custom skills with scheduled triggers to execute recurring operational tasks.
  • Deep research workflows spawn numerous parallel sub-agents to analyze data and perform adversarial reviews.

Long horizon tasks utilize loop engineering to self-improve through iterative grading and logging. Model routing integrates external frontier models like Codex for adversarial code reviews. Custom agentic operating systems built on markdown file structures help organize thousands of files for efficient retrieval.

커뮤니티 글

모든 글 보기