WEBVTT 00:00:00.000 --> 00:01:18.225 What if AI did 80 to 90% of the work in your business and all you have to do is just review that output? I'm not talking about using a chatbot where you copy paste things. I'm talking about actual real AI systems that could run on your behalf. AI systems that read your files, connect your tools, and executes real work your way following your process in your voice. That's what Cloud Code makes possible and the crazy thing about it is that you do not need to write a single line of code. Now real quick if you notice this video is quite long so if you don't have the time to watch this video in its entirety I recommend for you to save it so you can learn to implement Cloud Code later because by the end of watching this video you'll know exactly how Cloud Code works and you'll know how to implement into workflow so it can get 80 to 90% of the work done in your business. So you can focus on the high leverage work as a business owner, you know, the work they'll pay you a thousand dollars an hour or more. First, I'll show you how Cloud Code works in a visual way so that you can understand the building blocks and how you could use it in your business. Then I'll walk you through step by step how to get started with Cloud Code with a starter kit on GitHub that you can get in the link in the description. This is the exact way I use Cloud Code to do real work in my business. Finally, I'll show you live how I build a real AI system on Cloud so you can see exactly how all the pieces connect and how you could potentially use it in your business to regain hours and contribute 00:01:18.305 --> 00:01:45.675 actual value in your business. Your first question as a business owner is probably why Claude code? How does it actually work? And how's it different from other AI and automation tools? So if we were to map on an axis, high autonomy as in the tool is actually able to do work for us with us just asking you to do it and how easy it is to set up, Cloud Code would be at the top right. That is where we want to be as a business owner. We want to be able to tell something to do it and it does the work the way that we want it done. 00:01:45.995 --> 00:01:49.275 Now if we compare it to other tools, let's say ChatGPT, 00:01:49.435 --> 00:02:03.500 ChatGPT is pretty easy to set up but it doesn't do the work for you. It just asks it just answers back, um, some questions. It might be able to write some text, but that's it. Now if you compare this to n a 10 and Zapier, there is improved agent capabilities, 00:02:03.980 --> 00:02:15.285 but the problem with these tools is that we need to build every path like node one, node two, node three, node four. While we could just ask Claude, it could figure out the plan, what are the things I need to get done, and it'll actually do the work for you. 00:02:15.605 --> 00:02:18.245 Then there's tools or SDKs like Landgraf, 00:02:18.245 --> 00:02:20.085 Google ADK, and Crew AI, 00:02:20.245 --> 00:03:04.095 which still allows to have high autonomy, but the problem is that it takes weeks of building. And as business owners, we do not want to spend weeks technically building and coding on Python and doing all that stuff. Um, Ideally, we wanna be using something that is out of the shelf that we could ask it to build something, and it's able to build an AI agent very quickly and start doing work as soon as possible for us so can regain back more time and leverage. Now the best thing about Cloud Code is that I know there's code in it. Can actually code for you but when you connect it to external tools and you give it, uh, specific agent skills it can actually do work for you. You do not need to build infrastructure for these agents to work the orchestration and so on. Cloud Code will do that all for us which is we can thank Anthropic for that for building all infrastructure. 00:03:04.335 --> 00:03:14.495 So simply describe in plain English what you want and Cloud will build it, whether it's coding or whatever digital assets you wanted to build for you. So you basically get developer level power 00:03:15.240 --> 00:03:21.160 without being a developer and that is a complete game changer as a business owner. So let me show you how it works exactly. 00:03:21.720 --> 00:03:27.720 So if we were to take a normal AI chatbot which is how most people use Clont or Gemini 00:03:27.720 --> 00:03:29.000 or ChatGPT, 00:03:29.445 --> 00:03:43.900 it would look at you know, on a rudimentary level, it would be you as a user sending a prompt or a question or sending text to the language model, and then language model takes your text, it reasons, and then it gives you a response back. 00:03:44.300 --> 00:03:47.740 So this is how most business owners are using AI now. 00:03:47.980 --> 00:03:55.420 If you're using AI right now in this way, well, you're technically going to be doing all of the work in your business because AI is just being a consultant 00:03:55.525 --> 00:03:58.005 for you or a very smart adviser. 00:03:58.245 --> 00:04:12.430 In this case, you're probably just chatting with AI as a chatbot and you're just copy pasting all the time. Sam Altman predicted that soon we're gonna be seeing one person billion dollar companies and he's not wrong. Now is the best time to build a business based on leverage, 00:04:12.510 --> 00:04:24.350 specifically AI agents and digital asset. And if you're watching this, you probably think so too. That's exactly why I have a daily newsletter called the chief leverage officer. Every day, I'll share actionable insights on how I'm building a AI 00:04:24.565 --> 00:04:26.805 million dollar one person business. 00:04:26.885 --> 00:04:46.910 The thinking behind the systems, what's working, what's not, and where AI leverage is headed. If you wanna build a business that runs without you, not by hiring a team, but by building AI leverage, that's what I write about. Links in description and if you're watching this I highly recommend it. Now what makes Cloud Code different is that Cloud Code is not a chatbot. Now it has other capabilities which makes it agentic. 00:04:46.990 --> 00:04:57.015 So Cloud Code actually works under your local files and it can actually code for you. It can run commands for you and it could do actual work in your computer. 00:04:57.335 --> 00:05:04.295 Just as if you are interacting with your computer, Cloud Code can somewhat interact with your computer. While with a chatbot you're just getting text responses 00:05:04.380 --> 00:05:24.505 with Cloud Code, you're asking you to do something, it actually does things. So Cloud Code in its essence is an agent. It's a system that leverages an AI model to interact with its environment to achieve a user defined objective. That means you can give it a goal and AI can reason, plan, and execute on the actions to achieve that goal and fulfill tasks. 00:05:24.665 --> 00:05:31.400 And what makes it even better is that you can give Cloud Code access to the external world via MCPs and APIs. 00:05:31.560 --> 00:05:54.815 So this is why Cloud Code is not just a coding agent. It can code things. Yes. It can build tools like coding tools, Python scripts, but it can also interact with the external world, makes us a game changer as a business owner. Because now you can give AI some work and it can interact with your Notion, and it can do work for you as an example. And the best thing about this is that we do not need to code again. Cloud Code has very good orchestration. 00:05:54.975 --> 00:05:57.055 All we need to do is give it the right prompt, 00:05:57.295 --> 00:05:58.895 and it will do the work for us. 00:05:59.455 --> 00:06:18.275 Cloud will do the thinking. It will interact with our files. It will use MCP's APIs and track to the external world. It will do that work, come back, and then tell you what work it got done. So this is what an agent does, and this is how you want an employee to work in your business. You tell them what to do, and it does the work by itself. Now a quick caveat here, 00:06:18.595 --> 00:06:32.560 this can also extend to how other LLMs work. So this doesn't need to be only Cloud Code. This could be Gemini CLI or Codex. The only thing that we're swapping is a language model. So regardless if you're watching this video, if you use a different AI model, 00:06:32.720 --> 00:06:41.520 the concepts I'm sharing here can still be transferable to any other type of AI model that you have. And this is why it's very important to understand these concepts 00:06:41.845 --> 00:06:45.045 because if you understand how your files are structured 00:06:45.285 --> 00:06:58.340 and how MCPs work, APIs work, and how you an agent work, this gives you the power to use to swap out any language model in the future. Now the next thing I wanna go over is I wanna go over how Cloud Code works visually 00:06:58.420 --> 00:07:01.060 so that later on the walkthrough you understand 00:07:01.460 --> 00:07:28.860 what's going on, uh, with Cloud Code. So on the left side, we have you the user and you can interact with Cloud Code in two ways. You could either give it a prompt which could be like a question or the way or just ask it to do something or you can have it packaged into a slash command. I like to think of slash command kinda like a button and when you click this button it runs a prompt that you had previously or a workflow that you want Claude to do specifically. 00:07:29.180 --> 00:07:31.180 So instead of you asking 00:07:31.180 --> 00:07:32.060 Claude, 00:07:32.060 --> 00:07:33.340 uh, to do something 00:07:33.580 --> 00:07:46.995 over and over again, you could package that into a slash command, and Cloud can read that and then interact with your files and tools and do that work predictably. Now the thing about AI, AI doesn't come itself with memory. 00:07:47.075 --> 00:08:00.140 So AI is just a brain. So here's where we need to give, uh, Cloud the memory. So in Claude code, the memory is in a Claude dot m d file. So within the Claude dot m d file is where we store the information about our workspace 00:08:00.140 --> 00:08:09.245 and where Claude can find the tools and files. So all the routing rules memory is on here. So when when you give the prompt to Claude, 00:08:09.485 --> 00:08:19.165 it has this memory attached to it so that kinda like a map so it knows how to interact with your files and tools to execute that task for you. 00:08:19.565 --> 00:08:20.045 Now 00:08:20.460 --> 00:08:31.660 you might have some stuff in your files and tools or your code base or project base that you don't want Cloud to touch. So here's where you can have hooks, which is kinda like can act as guardrails. 00:08:31.740 --> 00:08:36.540 Or if you want Cloud to interact in a specific way, we can have hooks 00:08:36.315 --> 00:08:39.835 to tell Claude how to do something. So every time the Claude 00:08:40.075 --> 00:09:06.375 runs a tool or something, it has a hook to it. And then the best thing about this is that your files and tools, we can store, uh, skills or agent skills. Now if you don't know what agent skills are, I actually, uh, recorded a full video on that in my YouTube channel. You I put in the cards above and the link in description below if you wanna learn more about skills. But what skills does, it pretty much gives AI a playbook on how to do work in your business or like a standard operating procedure. 00:09:06.455 --> 00:09:13.335 Now this is game changer because AI right out of the box won't know how to do things in your business, but when you give it skills, 00:09:13.415 --> 00:09:19.200 it can actually do work predictably in the way that you want it to get done and can create value in your business. 00:09:19.680 --> 00:09:21.840 That in tandem with MCP, 00:09:21.840 --> 00:09:34.555 which is external connections to the external world, it can interact with external world and now can actually create value for you. You can have slash commands which can act as buttons if you want Cloud Code to work a specific workflow. 00:09:34.795 --> 00:10:01.635 You have a memory Cloud dot MD file that tells Cloud how to do things. You have hooks, so if you don't want Cloud Code to touch specific things, you have agent skills which teaches AI how to do work, and you have MCP connections which allows Cloud Code to interact with external world and deliver actual value in your business. Now the way Cloud Code works is that Cloud Code has a context window. There's a limit to how many messages you can send Cloud Code. So if you've ever used Cloud and you notice that you're out of messages and it compacts, 00:10:01.795 --> 00:10:06.515 there's a reason for that. It's because Cloud uses processing power to process text, 00:10:06.755 --> 00:10:30.635 and the more text you give Cloud Code, the more processing power it requires, and it exponentially increases after a certain point where at some point it's not financially logical to continue the same context window when it's much better to just compress the window by creating a summary and then spawning a new agent. Now this is important because when you work with Cloud Code, if you are working with one main agent 00:10:30.875 --> 00:10:32.875 and it always summarizes 00:10:32.875 --> 00:10:34.155 summarizes summarizes, 00:10:34.475 --> 00:10:40.560 the context that you've given at the first agent is going to be completely lost by the third or fourth cycle. 00:10:40.800 --> 00:10:46.640 So in order to maintain performance, our goal is to try to maintain a clean context window. 00:10:46.800 --> 00:10:51.920 K? Because over time, the performance of an AI model degrades as you give it more 00:10:52.515 --> 00:10:53.475 context. 00:10:53.795 --> 00:11:00.515 So there's research on this. It said that 11 out of 12 models dropped below 50% at 32 context. 00:11:00.755 --> 00:11:11.930 And then at around 100 k is when things start getting worse and around one fifty k is where you probably will get unreliable outputs. So if you find Cloud Code not giving you good outputs, 00:11:12.490 --> 00:11:15.210 it's likely because you're giving it too much context. 00:11:15.450 --> 00:11:19.130 This is why it's very important that we use sub agents. 00:11:19.370 --> 00:11:24.785 So what Cloud Code can do is that ideally you want to use it as an orchestrator 00:11:24.785 --> 00:11:32.305 where you spawn sub agents because each sub agent will have a fresh context window. So instead of going through the cycle 00:11:32.545 --> 00:11:37.600 where you work parallel with one single agent, um, you probably will get 90% 00:11:37.600 --> 00:11:40.000 better results with a multi agent approach. 00:11:40.160 --> 00:11:43.440 Not only do you save context where the agent performs better, 00:11:43.600 --> 00:12:07.030 but you also get more work done because you have other sub agents doing smaller tasks. So this is why when you work with Cloud Code, if you're working on a big task or big project that has a lot of context, need to explain it. Ideally, you wanna use the main Cloud Agent that you're speaking with as a planning agent that you ask it to spawn sub agents to do the work for you and to have the main agent review the work that a sub agents, uh, does. 00:12:07.350 --> 00:12:13.590 Because if you were to use the main agent, the ones you're talking with to do all the work and plan and strategize, 00:12:13.670 --> 00:12:25.645 the chat window will compact and, like, you'll compact multiple times and you're gonna get frustrated because you by the requirements you shared with them in the start, they don't know what it is because they're working on a summarization of a summarization. 00:12:25.885 --> 00:12:34.605 So this is why it's important to understand how sub agents work and when to use them and when to spawn them when needed. This will also help you in terms of designing 00:12:34.940 --> 00:12:55.955 which sub agents to do specific types of tasks and how you could have them all connect together. And this is a very important concept because context engineering right now is gonna be very important in terms of getting, uh, AI agents to do predictable work that is accurate in your business, at least for now until they fix the context window and once AI has unlimited memory, 00:12:56.195 --> 00:13:13.220 then we're gonna be talking completely different game then. Now the way Cloud Code spawns sub agents is that you could ask it in your prompt or in a command to spawn the sub agent, and then it'll go ahead and spawn the sub agent and prompt that sub agent on what it needs to do. And because sub agents also have access, 00:13:13.380 --> 00:13:14.740 uh, to your files 00:13:14.900 --> 00:13:16.100 and MCP 00:13:16.260 --> 00:13:28.495 or the external world, it can actually do real work in your business as well. And Cloud Code can spawn as many sub agents in parallel as you want it to. Now big caveat here, if you're not on the max plan, 00:13:28.735 --> 00:13:45.260 if you're gonna spawn sub agents on the pro plan, you're likely gonna use up all your usages for the week. So I recommend using the Cloudmax plan if you are going to plan to spawn a lot of sub agents. And the great thing about sub agents is that if you know how to set them up, you probably get 00:13:45.660 --> 00:13:57.635 let's say, what would you get done in a week? Probably done in, ten minutes because sub agents can do all the work for you. Now a core pillar that allows agents to do real work in your business is agent skills, which is basically playbooks 00:13:57.635 --> 00:14:23.125 that your agents can run-in order to do specific tasks in your business. Now if you're again, if you're unfamiliar with agent skills, did record another video, a full breakdown. I'll put the link in description and the cards above. You can check that out. But, basically, in your files, you wanna have a set of skills that you could have that agents could quickly pull up like playbooks and run them. So the way this works is let's say you have a skill on YouTube research. 00:14:23.605 --> 00:14:27.365 You'd ask, hey. Can you research this video idea for me? Whatever the video is. 00:14:28.250 --> 00:14:39.050 Claude is gonna think, Video. It's gonna scan through this Claude skills folder. It's gonna find the one that matches which is YouTube research and then it's gonna load that. It's gonna load only the skill. 00:14:39.210 --> 00:14:50.065 Now the reason why we wanted package in the skill is because package skills allows Claude to only get what is relevant to the specific task and not pollute the context window. 00:14:50.225 --> 00:14:55.585 Because if Claude were to read all of this as, like, one single prompt, it would pollute the context window, 00:14:56.080 --> 00:15:07.360 and then it would compact, it wouldn't be as efficient with work. So pack so skills is kinda like packaged prompts, instructions, everything in one place to do a specific task, uh, predictably. 00:15:07.600 --> 00:15:13.615 So Claude will scan the YAML front matter, which is pretty much kinda like the title and the description. 00:15:14.255 --> 00:15:21.215 And then once it matches the keyword that we asked it for, which is researches video, it will pull up that skill. 00:15:21.455 --> 00:15:22.255 So for example, 00:15:23.310 --> 00:15:31.550 research this video will pull up the YouTube research skill. It'll have the skill that MD file over here with the name YouTube research, description analyze videos for insights, 00:15:31.790 --> 00:15:34.350 tools, YouTube MCP, Notion MCP. 00:15:34.510 --> 00:15:50.585 K. With the instructions. So now the agent will know how to do this task. It will know that it will use these two tools and it will have instructions on how to do it specifically and the prompt on how to think about it and how to do the analysis and so on. And then we give it access to Notion and YouTube. 00:15:51.710 --> 00:16:02.830 Now the agent can actually do real valuable work because you've given it access to tools, you've given it access to instructions on how to do things specifically in your business and what to do with that information. 00:16:03.150 --> 00:16:04.990 And then once it does all this work, 00:16:05.390 --> 00:16:06.910 it can then save that, 00:16:07.485 --> 00:16:11.405 um, let's say, this research plan or this research analysis 00:16:11.485 --> 00:16:31.700 into your folder. So now you can actually see the folder and interact with it. And if there's anything wrong with the folder or with the research, you can prompt it again to say, hey. Look up this keyword. Check this, run this with another skill that you want. So now skills stack up on each other and you could work, uh, with agents kinda like an employee 00:16:31.940 --> 00:16:41.015 doing work for you in your business. Previously, in order to run a business as a business owner, you'd call yourself a CEO, but I like to call that as chief everything officer. 00:16:41.255 --> 00:16:47.575 So you'd be in charge of everything. You need to do everything yourself or review everything yourself. Content, research reports, 00:16:47.815 --> 00:17:06.685 email scheduling, data entry, you'd probably be doing at least some of these tasks yourself if you don't have delegated. And even if you have delegated, you probably will have to do some of this stuff as a CEO because you need to manage things in your business. So this will likely take you forty hours a week. Now because Cloud Code exists, you can Cloud Code can be your CEO. 00:17:07.085 --> 00:17:29.410 So now Cloud Code does the things that you were doing before. So content, email, research reports, scheduling, data entry, Cloud can do that. So what would take you forty hours a week, now Cloud Code can probably do thirty six hours of that and you just take it just takes you four hours of your time where you just direct it and give it feedback on what to do. So you can pretty much get the same output 00:17:29.650 --> 00:17:32.050 of what you did before with 10 x less of your time. 00:17:32.735 --> 00:17:49.600 So this is why using Cloud Code is one of the highest leverage things, uh, you could do as a business owner and it starts by understanding how Cloud Code works. So with that being said let's dive into the walkthrough. Real quick if you find this video helpful I'm running a paid two week cohort called your first AI employee 00:17:49.600 --> 00:18:05.335 where our goal is to turn Cloud Code into an AI employee in your business that generates you at least $10,000 a year in value. We're going to x-ray your business to find where AI would fit, then we work to build your first AI employee that saves you at least ten hours a week and generates you $10,000 00:18:05.335 --> 00:18:25.560 a year. This isn't going to be a generic system. It's gonna be one built customized to you, your process, your standards, and your voice. If you'd like to join me and other business owners, links in the description and there's limited spots I accept every month. Okay. So the best way for us to get started to learn how Cloud Code works is for us to build something live. So we're gonna build a YouTube breakout analysis, 00:18:25.720 --> 00:18:27.800 which is our goal is to have Claude 00:18:28.355 --> 00:18:33.795 spawned sub agents that will scrape through the all of YouTube with specific keywords. 00:18:33.955 --> 00:18:50.460 We'll find the breakout videos which is specific videos that have a ton of views for the channel that has a smaller amount of subscribers. That's why they're called breakout views. So we can take that data and we could then use that to position our videos to get the most amount of views possible. 00:18:50.780 --> 00:18:57.420 So here's how it's gonna work. So I'm gonna ask Claude, hey. Research this video idea for me, which whatever video idea it is. 00:18:57.740 --> 00:19:01.455 And then Claude is gonna be able to go through cloud.md. 00:19:01.455 --> 00:19:02.815 It's gonna understand 00:19:02.975 --> 00:19:08.975 what skills we have, and then it's gonna run through the YouTube research skill, which we're gonna build together, 00:19:09.215 --> 00:19:12.335 and then it's gonna connect to the YouTube MCP, 00:19:12.575 --> 00:19:13.215 uh, as well. 00:19:14.190 --> 00:19:17.710 And then it's gonna spawn sub agents called YouTube researchers 00:19:17.790 --> 00:19:20.030 which with specific keywords 00:19:20.190 --> 00:19:32.295 that will, um, then look up YouTube. They'll get their own results and then they'll give it back to the main Claude or the Orchestrator agent and then let us know, uh, which videos, uh, performed best. 00:19:32.615 --> 00:19:47.780 So that's what we're gonna build. So let's dive in. Alright. So in order for us to get started, the first thing we need to do is we need to download Visual Studio Code. So we can go ahead and download it. It's completely for free. Or if you prefer flexibility and you'd potentially would like to use Gemini, 00:19:47.940 --> 00:20:09.720 um, you can download Google Anti Gravity. So both of these are really good options. I'm gonna go ahead with Visual Studio Code because a lot of the extensions that I use work perfectly on Visual Studio Code and some of them don't quite work that well on anti gravity. So once you have Versus Code, uh, installed, the first thing we need to do is we need to install the Cloud Code extension. We go to extensions on the left side 00:20:10.040 --> 00:20:13.240 and then what we're going to do here is we're going to search Cloud and 00:20:15.320 --> 00:20:17.400 then we're going to go ahead and install it over here. 00:20:19.605 --> 00:20:33.045 K. So you're gonna get a prompt over here that's gonna show you how to use Cloud Code. You can go ahead and read this. I'm just gonna click on mark done because I'm already familiar with Cloud Code. Now the next thing I'm gonna do is I need to clone a git repository 00:20:33.230 --> 00:20:36.030 because if I just start working Cloud Code 00:20:36.430 --> 00:21:15.405 or Versus Code, I don't have a folder open here. So and if you open up a empty folder, Cloud will still work, but it won't have the basic infrastructure that Cloud works with. So I actually created a Cloud Code business starter kit which has all of the infrastructure in order to get Cloud Code to work, uh, right away instead of you needing to do go through the trial and error, trying to get things set up, reading the developer documents, I put it all neatly over here. So you can go ahead and grab this from the link in the description below and you can get started. There the instructions on how to set it up are over here, but I'm gonna go through step by step right now. So I'm gonna grab this link. 00:21:17.885 --> 00:21:24.125 I'm gonna grab this link, and I'm gonna go to, uh, Visual Studio Code. I'm gonna click clone GitHub repository, 00:21:24.285 --> 00:21:33.360 clone from URL. I'm gonna go ahead and open it up to where I have my YouTube walkthroughs. I'm gonna select repository as destination. 00:21:34.720 --> 00:21:38.800 Okay. So I'm gonna click open and boom. It opened up 00:21:39.265 --> 00:21:44.305 the folder. So if you look in the left side, if we open this up, we can see the dot clot folder. 00:21:44.545 --> 00:21:51.025 So the dot clot folder is where, uh, we have the things that clot works with. So agents, 00:21:51.105 --> 00:21:56.160 so if we have custom agents, commands like I explained, uh, previously, and then skills. 00:21:56.400 --> 00:22:04.080 So I created some skills already so that if you want to create an agent, if you wanna build anything, you can ask Claude and we'll already know, 00:22:04.320 --> 00:22:05.840 uh, the correct format 00:22:05.955 --> 00:22:22.500 in order to create these things. So I did that. So there's skills and then there's the settings dot JSON which is where hooks stay and what gives permission for what Claude, uh, Claude to do. So these are the permissions that Claude has that could work in this fold, which is read, glob, grep, bash, 00:22:22.820 --> 00:22:28.020 and so on. And then there's hooks. So every time a new skill is created or renamed, 00:22:28.100 --> 00:22:34.225 Claude will update the claude.md file. So the Claude dot m d file again is the explanation 00:22:34.225 --> 00:22:40.945 on how, uh, this repository works and this Claude dot m d file is given to every single Claude agent. 00:22:41.345 --> 00:23:00.445 So then the Claude agent will know exactly how to work in this code space or knowledge base if you like to call that. Again, markdown files are just text files. If you're not a developer and you're confused what dot m d stands for, just think of it as like a text file that is written in a specific format. So these hashtags are kind of like headers, 00:23:00.605 --> 00:23:02.765 two hashtags like a header, 00:23:02.925 --> 00:23:05.085 uh, asterisk or kind of like bold. 00:23:05.325 --> 00:23:14.610 So it just looks like that. So if I want to read this in a more in a way that is more friendly I can click on control shift v, click on it altogether 00:23:14.690 --> 00:23:16.690 and it's gonna open up a preview. 00:23:16.930 --> 00:23:17.570 Okay? 00:23:18.050 --> 00:23:22.610 So technically this is just a prompt. If you just look at just like a nicely formatted prompt. 00:23:23.165 --> 00:23:27.645 So you we work with AI via prompts. So if I look through this CloudNMD, 00:23:27.805 --> 00:23:31.245 uh, this is what every Cloud agent will get so to understand, 00:23:31.565 --> 00:23:34.525 um, how the directory works, the routing. 00:23:34.765 --> 00:23:40.440 Um, the number one important thing here with working with AI is routing because when AI agents 00:23:40.440 --> 00:23:54.885 start to work, we want them to understand where to find the things to do the work properly. The last thing you want is AI to not have the context to do the work correctly, and then it does a completely wrong output. This is why it's important to have the routing because 00:23:55.045 --> 00:24:01.285 when AI knows the routing, it can look for the right tools to do the right amount of work or the right context to do the right work. 00:24:01.525 --> 00:24:09.230 K. So routing over here, if you wanna create a skill, we'll use these skills and so on. And then it'll go over the folder structure which looks like so 00:24:09.470 --> 00:24:20.990 and MCP servers will be configured on MCP JSON. I wanna do a quick recap before I go through the walkthrough so you can understand what's going on and you don't get lost. So the dot clot folder is pretty much a 00:24:21.405 --> 00:24:32.125 so the dot Cloud folder is where we're gonna store our agents. So custom agents or sub agents, commands, which is the buttons on or, like, packaged prompts that we want Cloud to do work predictably. 00:24:32.285 --> 00:24:35.900 Skills, which is agent skills, which is packaged SOPs, 00:24:35.900 --> 00:24:38.140 knowledge playbooks, and then the hooks configuration. 00:24:38.460 --> 00:24:46.860 Now it's in a dot Cloud folder because this is what allows Cloud to automatically discover these tools and to have Cloud work on it immediately. 00:24:46.860 --> 00:24:52.675 Instead of us having it stored in a different file format, this is automatically stored when Cloud is launched. 00:24:52.755 --> 00:25:01.475 So whenever I open a new, uh, Cloud or a new agent or I have multiple agents, all these agents will be reading the dot Cloud folder immediately. 00:25:01.555 --> 00:25:12.630 K? Similar with the cloud dot m d folder as well. And then at the bottom, there's a a option to work with plan mode or normal mode. So I can switch between those those modes 00:25:12.790 --> 00:25:27.765 by clicking shift tab on my keyboard. Now you wanna use plan mode when you are building stuff because you want Claude to sort of interview you and pretty much go through Claude and all of the requirements and what you're gonna build step by step. That way you do not make mistakes 00:25:28.180 --> 00:25:43.380 because if you just ask Claude to do something it's not gonna have the context to do it the way that you want it to do to do it. So plan mode is very helpful. Now once once you're done with plan mode you can click shift tab again and you can toggle between asking Claude to plan or asking Claude to just do the work 00:25:44.085 --> 00:25:59.700 already. So if I want to get started what I'm gonna do is I'm gonna click on Claude open over here because I already have Claude configured on my Versus Code I can just get started. If this is the first time you're sending this up you're gonna prompted over here to log in. All you need to do is use your cloud.ai 00:25:59.700 --> 00:26:00.420 subscription. 00:26:00.660 --> 00:26:11.785 You do not need to use the API. So if you already have a subscription with Cloud, just log in, uh, following the instructions over here. So our goal, again, like I mentioned, is we're gonna build a YouTube breakout researcher, 00:26:12.105 --> 00:26:12.665 and 00:26:12.905 --> 00:26:19.385 let's see how what we built. So let's start. Hey, Claude. Um, my goal is to build a YouTube breakout researcher. 00:26:19.705 --> 00:26:24.665 Um, in order for us to get started, we're gonna need to get access to YouTube 00:26:24.030 --> 00:26:24.910 MCP. 00:26:25.310 --> 00:26:27.550 So the goal is this. I wanna 00:26:28.110 --> 00:26:32.910 I want you to, uh, find an MCP server that is capable of 00:26:33.310 --> 00:26:37.710 finding a YouTube channel, their views, their subscriber count, 00:26:37.950 --> 00:26:46.305 and when the video was published, and potentially any other, uh, requirements for us in order to find their breakout score. So a breakout score is basically 00:26:46.465 --> 00:26:51.425 the, um, number of views divided over the number of subscribers on the channel. 00:26:51.870 --> 00:27:01.950 So ideally, wanna identify videos that have the highest breakout scores so we could model our videos on those videos so we can get the most amount of traffic to our channel. 00:27:02.270 --> 00:27:05.150 Does this make sense? K. So I'm gonna get started by 00:27:05.575 --> 00:27:11.895 asking Claude to find an MCP server because if we don't have the MCP server, then we can't really find the breakout videos. 00:27:12.615 --> 00:27:21.120 So it went ahead. It understood, and it's calling the create MCP skill. So I'm gonna do that. In the meantime, I'm gonna open up a new chat over here, 00:27:22.880 --> 00:27:25.760 and we're gonna do two things at once. So I'm gonna ask Claude. 00:27:26.160 --> 00:27:31.225 Hey, Claude. I'd like to work on building a skill where specifically 00:27:31.465 --> 00:27:41.065 we, uh, scrape through YouTube to find breakout videos. Now a breakout video is a video that has substantially more views than a number of subscribers. 00:27:41.225 --> 00:27:46.460 Ideally, we wanna find find out breakout videos that have a two x score or more. 00:27:46.940 --> 00:27:55.820 Now here's how the skill is gonna work. Um, I'm gonna give you a video idea and from that video idea, I want you to find five core angles 00:27:55.980 --> 00:28:05.465 or five core perspective shifts. And then from those core perspective shifts, I want us to identify five keywords for each core angle or perspective shifts. 00:28:05.785 --> 00:28:08.665 Then once we get through all those keywords, 00:28:08.905 --> 00:28:12.585 the next thing I wanna do is I want us I want us to spawn a sub agent 00:28:12.900 --> 00:28:17.540 for each one of those five keywords that will be searched, uh, in parallel. 00:28:17.700 --> 00:28:22.820 And then later on, I want you to consolidate all of that information and create a report for us. 00:28:23.060 --> 00:28:24.500 Now specifically, 00:28:24.500 --> 00:28:41.175 I want you to make sure that the videos that we look for are at least five minutes or longer and are not short. We do not want YouTube shorts because we post long form video. So five minutes or longer only. And I want you to create this where it's a phase based process where you ask me 00:28:41.690 --> 00:28:50.730 where we approve the keywords and so on so that we can work on in case, um, we wanna make any adjustments. Okay. So I'm gonna have that work and I'm also gonna tell it, 00:28:51.450 --> 00:28:54.010 uh, by the way, we're currently setting up the MCP 00:28:54.010 --> 00:29:07.225 so, uh, don't worry about that for now. We're working on that on parallel. Okay. So now I'm gonna go back over here, and it's gonna look for okay. So it's spawning an MCP finder agent to search the best YouTube MCP server options. 00:29:07.385 --> 00:29:11.305 So I have this agent already set up over here called MCP dash finder. 00:29:11.770 --> 00:29:16.090 So Claude already knew that this existed, so it's running this custom agent. 00:29:18.570 --> 00:29:23.290 K. So while it's running, it's going through the skills, and it's actually creating 00:29:24.145 --> 00:29:27.665 the skills. So I'm gonna go ahead and stop this very quickly. 00:29:29.025 --> 00:29:31.905 And what what I'm gonna say is I wanna do this plan. 00:29:32.545 --> 00:29:33.825 So can we do 00:29:34.705 --> 00:29:37.665 can we do what I shared earlier? 00:29:40.170 --> 00:29:42.010 What plan it together? 00:29:42.410 --> 00:29:43.930 K. The 00:29:43.930 --> 00:29:51.130 reason why I'm gonna do plan here, which I forgot about it. So Cloud Code has three different options. It has the normal option, 00:29:51.825 --> 00:29:54.225 which you just ask it, it'll start building things, 00:29:54.465 --> 00:30:09.760 then there's a, uh, plan mode. So in this case, creating a new skill, a plan mode would be best because then, um, we do not want Claude to make assumptions, we want Claude to create a plan. We review it, and then we work. So it's asking me right now. For the parallel search, 00:30:10.480 --> 00:30:15.760 should each of these five sub agents search all five keywords for the angles or should respond 25 agents, one for keyword? 00:30:16.625 --> 00:30:19.265 25 agents would be way too much, some say five agents. 00:30:19.505 --> 00:30:22.785 How many breakout videos should we aim find per angle or keyword? 00:30:23.185 --> 00:30:24.065 I would say, 00:30:25.665 --> 00:30:29.265 top three to five per angle is good. Sometimes it's really hard to find 10. 00:30:29.780 --> 00:30:33.540 What should the final report include? Video details, title views, subprocesses. 00:30:33.540 --> 00:30:35.300 I'm gonna go ahead and say 00:30:35.860 --> 00:30:38.740 yeah. We can do all this. I would say the most important important 00:30:38.740 --> 00:30:41.220 is the title templates 00:30:41.860 --> 00:30:43.300 so that we can 00:30:43.925 --> 00:30:44.965 brainstorm 00:30:45.045 --> 00:30:47.845 different ways we can implement 00:30:48.005 --> 00:30:50.645 our core ideas into those 00:30:52.325 --> 00:30:53.605 title templates. 00:30:54.165 --> 00:30:55.365 And then future 00:30:55.365 --> 00:30:57.845 have the option to download 00:30:58.085 --> 00:30:59.285 the thumbnails 00:30:59.600 --> 00:31:03.040 so we can use those as inspiration. 00:31:04.400 --> 00:31:05.760 Let me click enter. 00:31:05.920 --> 00:31:08.240 Where should the research agent output itself? 00:31:08.560 --> 00:31:10.800 Research YouTube breakout video idea. 00:31:10.960 --> 00:31:13.315 I think that's fine. Submit answer. 00:31:13.715 --> 00:31:16.115 Okay. So let's see here. It actually 00:31:16.355 --> 00:31:20.515 is, um, web searching and looking at these GitHub repositories 00:31:20.515 --> 00:31:23.875 over here, um, to find the MCP servers. 00:31:24.195 --> 00:31:32.130 So let's see here. For title template, should we extract exact pattern structure from breakout video titles, show both templates, and actual give templates plus real titles for context? 00:31:32.610 --> 00:31:33.650 Yes. Both. 00:31:33.970 --> 00:31:45.405 Uh, for the breakout zero calculation, we restrict about two x minimum or short range. Uh, two x only show tiered results categorized by breakout level. I like that. Should the skill save thumbnail URLs only or actually download the image files? 00:31:46.365 --> 00:32:03.640 Download yeah. Let user decide. K. What should happen if you wanna modify the angles or keywords during approval? Interactive edit, you suggest changes or regenerate. You directly edit the person's list and continue. Right? Okay. Interactive edit. After five agent after the five agents finished searching parallel, should I show you raw results before consolidating? 00:32:03.640 --> 00:32:04.840 Okay. I'm gonna say 00:32:05.160 --> 00:32:08.200 share with me the consolidated 00:32:08.360 --> 00:32:08.920 report, 00:32:09.160 --> 00:32:09.960 but also 00:32:10.395 --> 00:32:11.355 compile 00:32:11.355 --> 00:32:13.115 all of the agent 00:32:13.515 --> 00:32:15.355 results into a, 00:32:16.235 --> 00:32:17.755 uh, consolidated 00:32:18.555 --> 00:32:22.875 folder consolidated files. So I'm click submit answer. K. So the MCP 00:32:23.190 --> 00:32:30.230 finder agent is still looking. K. Perfect. Now I have a complete picture. Let me summarize the complete workflow. We'll build phase one, phase two, phase three, 00:32:30.550 --> 00:32:31.990 phase four, phase five. 00:32:32.470 --> 00:32:43.625 Okay. So output structure. So it's gonna create a new folder where it's gonna create these outputs for us so it organizes it, uh, for us. Does this capture everything correctly? Any adjustments before I start building? Okay. So it found 00:32:43.705 --> 00:32:51.670 the, uh, best MCP server for our use case. So I'm a say yes. Let's go ahead. Yes. Let's go ahead and install 00:32:51.670 --> 00:32:53.030 this MCP. 00:32:53.510 --> 00:32:57.350 So, uh, because this MCP server is on GitHub, I believe 00:32:57.590 --> 00:33:02.470 it'll be able to set up. So now Cloud is telling me that I need to get an API key. 00:33:02.710 --> 00:33:18.545 So I can follow these instructions. Go to Google Console. Go to Google Cloud console. So if I click on that, it'll pull up my Google Cloud. So I'm now on my Google Cloud. If I go back, I need to go to click a new project, enable YouTube data API, create API credentials, 00:33:18.980 --> 00:33:31.380 and so on. So I'm gonna go ahead and skip this. So I'm just gonna copy over the API, and we'll get started. Alright. So I got the API key. I'm gonna share it, and then Claude is going to configure the MCP for us. K. So it went ahead, 00:33:31.540 --> 00:33:39.285 and it created an MCP folder over here, and it added the API key. So I'm a click on yes. So if I go to mcp.json, 00:33:39.285 --> 00:33:47.525 we could see the API key over here, and then we could see the MCP that we're using. So this is how Claude gets access to the external world via the dot MCP dot JSON 00:33:48.350 --> 00:33:59.470 file over here. Go ahead and close that out. K. So it seems like this agent is still planning. K. So apparently, it says it recognize the MCP server is connected. So the way that I could check if an MCP is set up correctly, 00:33:59.790 --> 00:34:02.750 I could type slash MCP, manage MCP servers, 00:34:03.175 --> 00:34:04.615 continue in terminal, 00:34:05.095 --> 00:34:06.935 and it's going to open up 00:34:07.575 --> 00:34:17.575 a new terminal. So click on enter. So it is actually connected. So, yes, proceed. So I can type m c p. So we have two m c p servers, 00:34:17.655 --> 00:34:18.535 it seems like 00:34:19.140 --> 00:34:20.500 project MCPs 00:34:20.580 --> 00:34:25.300 and YouTube. So and Notion, which is I connected globally on my computer. 00:34:25.460 --> 00:34:29.620 So that's good. So we're all set. So I'm ask, are we done configuring 00:34:31.380 --> 00:34:39.085 MCP? Okay. So now it's asking me, do you want Cloud to use YouTube MCP tools without asking permission each time? I'm gonna say yes, 00:34:39.325 --> 00:34:46.285 uh, because we're gonna use sub agents in the future. If we do not add this, then the sub agents will fail. So I'm gonna say yes. 00:34:47.645 --> 00:34:51.210 So what's gonna happen here in the settings dot JSON, 00:34:51.610 --> 00:34:58.570 right now we have read, glob, grep, bash, bash. It's gonna add those tools within that MCP server that Claude can edit. 00:35:00.410 --> 00:35:01.210 Okay. 00:35:01.290 --> 00:35:01.930 So 00:35:02.090 --> 00:35:04.490 it created over your MCP underscore YouTube 00:35:05.255 --> 00:35:06.135 underscore 00:35:06.455 --> 00:35:08.215 excuse me underscore asterisk. 00:35:08.215 --> 00:35:15.415 Hopefully this will work. If if it doesn't work then we'll likely have Claude to actually put in the real, uh, tool names over here. 00:35:16.055 --> 00:35:17.895 K. So let's go back here 00:35:19.270 --> 00:35:25.990 and let's click on yes. So it's gonna so it's checking what tools are available from the MCP so we can build the agent properly. 00:35:26.550 --> 00:35:43.965 Um, let's see what happens. So you'll notice that Claude will ask you. You can just say yes allow if you wanted to stop asking you for permissions. Yes. So I'm gonna ask if we're done setting up the MCP here. K. So apparently, we'll need to reload and test, but I think the MCP servers are already working. So I'm gonna ask, can you test 00:35:44.285 --> 00:35:47.165 searching for a breakout video 00:35:47.760 --> 00:35:48.400 for 00:35:49.200 --> 00:35:51.280 business for productivity. 00:35:52.880 --> 00:35:54.320 Sorry. Uh, productivity 00:35:56.000 --> 00:36:11.885 for business owners. Click enter, and let's just see if it works. If not, then we'll have to restart to get the MCP to work. Alright. So this agent just finished searching how to use all the tools. So apparently, it's gonna update the tools needed. So it identified all the tools we needed. So 00:36:12.285 --> 00:36:15.720 and it's, uh, in the plan folder. So meanwhile, 00:36:15.720 --> 00:36:28.855 while that's working, I'm gonna go ahead and, um, I'm gonna restart Cloud Code because the MCP servers aren't quite, uh, the u two MCP servers aren't quite connecting. So in order for me to do that, I'm gonna click control shift p, and I'm gonna type reload 00:36:29.655 --> 00:36:31.175 reload window. 00:36:31.495 --> 00:36:34.055 And what's this gonna do is gonna pretty much refresh, 00:36:34.855 --> 00:36:38.215 Visual Studio Code. So I'm gonna ask it to test again. 00:36:39.655 --> 00:36:41.095 Sometimes MCP, 00:36:41.095 --> 00:36:42.775 uh, servers are quite, 00:36:43.950 --> 00:36:44.750 like, 00:36:44.750 --> 00:36:49.230 miss or match. Sometimes it works. Sometimes it doesn't. So this is why we need to test it. 00:36:50.030 --> 00:36:54.670 Okay. Let's continue. So this is why using the plan mode is very helpful 00:36:55.335 --> 00:36:56.135 because 00:36:56.215 --> 00:37:02.775 Claude can get all of the context and put it into a a plan folder, which is over here, Claude plans. 00:37:03.415 --> 00:37:03.975 And 00:37:04.535 --> 00:37:09.815 when you work in a different, uh, when you compact when the window compacts or you work with a new agent, 00:37:10.340 --> 00:37:11.460 we can continue. 00:37:11.940 --> 00:37:12.500 So 00:37:12.900 --> 00:37:18.980 looks like the MCP server is connecting, so it actually works when you restarted it. And over here, says, 00:37:19.220 --> 00:37:24.900 um, that we already have permission set up. So it's gonna verify the settings and then continue building. 00:37:25.935 --> 00:37:32.095 K. So it's gonna create the YouTube breakout search agents now. So I'm gonna do this. I'm gonna ask, can you 00:37:32.655 --> 00:37:34.415 make sure to read, 00:37:35.295 --> 00:37:37.535 uh, create agent skill 00:37:38.880 --> 00:37:39.680 before 00:37:39.680 --> 00:37:56.015 creating the agent so that I want I want them to I want Cloud to build it right the first time instead of us going back and forth. So I'm gonna make sure ask it to read the create agent skill before creating the agent. Yeah. So now this is the thing with Cloud is that sometimes you need to guide it. 00:37:56.495 --> 00:38:00.175 So it's creating this directory when we don't quite need it. 00:38:00.815 --> 00:38:03.855 Okay. Because we already have one that exists over here. 00:38:04.175 --> 00:38:05.615 So we already 00:38:06.015 --> 00:38:07.135 have an 00:38:07.950 --> 00:38:08.990 existing 00:38:09.390 --> 00:38:10.110 agents 00:38:10.510 --> 00:38:11.470 folder. 00:38:13.950 --> 00:38:21.795 K. I'm gonna click on yes. Okay. So you can see here that, um, we use the YouTube MCP. It works great. So it's found, 00:38:22.275 --> 00:38:24.195 um, some good performing videos, 00:38:24.675 --> 00:38:35.315 but as you can see here, the breakout score aren't that high. So 0.6, 0.3, zero point, uh, two, and so on. So this is good. So we're we're already set. So 00:38:35.950 --> 00:38:44.350 let's continue on this chat. So, yes, allow. So Claude made a mistake here, so it fixed it, and it went ahead and created this markdown file over here. 00:38:46.590 --> 00:38:50.110 Um, if you notice, it's not the same as this file, 00:38:50.885 --> 00:38:54.725 whoops, as this file over here. This is how an agent should be structured. 00:38:55.365 --> 00:39:01.445 K. So now Claude will need to update that. This is why I've created these skills over here, 00:39:01.765 --> 00:39:10.600 which teaches Claude the standard formatting to do things. So this is very important because Claude right off the box doesn't have context, 00:39:11.160 --> 00:39:11.640 so 00:39:11.960 --> 00:39:15.160 this is why it's very important. So you see it just made these updates over here. 00:39:16.920 --> 00:39:20.175 I made the other. So if I open up the folder, can see now it's, 00:39:20.255 --> 00:39:23.695 uh, you can see now it is it is using that the correct way. 00:39:24.015 --> 00:39:26.255 Um, also, with using agents, 00:39:26.255 --> 00:39:29.935 you can choose which model to use. So here we're using Sonnet because 00:39:30.430 --> 00:39:50.095 we could potentially even use Haiku over here. So if you want to, uh, conserve tokens, you can use Haiku. I'm gonna stick to Sonnet. It seems fine for this. This is why having custom agents is helpful because you can decide what tools to use. Like, for example, MCP Finder doesn't need YouTube search tools, so it's it's custom. 00:39:50.415 --> 00:39:53.055 And I can use this prompt that is predictable. 00:39:53.375 --> 00:40:09.780 Right? It uses a predictable prompt instead of having Claude need to create it all the time. So I could ask Claude to spawn sub agents, but now I'm gonna rely on Claude to create the prompt to send that sub agents to do that work. But if I create a custom agent, I'm more in control of the prompt 00:40:10.255 --> 00:40:15.135 of what it does and so on. So this is why, uh, agents are really helpful. 00:40:15.935 --> 00:40:23.935 Uh, skills are really helpful because skills is what gives pretty much context to agents on how to do the work. Okay. Let's continue here. Seems like 00:40:24.335 --> 00:40:32.230 it got stuck here. So I think we're done setting up the MCP, so we likely will not need this tab anymore. So I'm gonna close this. 00:40:32.550 --> 00:40:34.870 Um, if you close a tab and 00:40:35.030 --> 00:40:46.645 you're wondering where it went, you can always, uh, click the stop down menu over here and you can always find your past conversations. So don't worry about losing work. K. It's created this. So quality criteria, 00:40:46.885 --> 00:40:49.925 search all five keywords, filtered all results. 00:40:50.565 --> 00:41:02.010 So now it's creating the it's identifying the agent to make sure the agent does the work correctly. So I'm gonna click on yes. Also, noticed it has three to five diverse breakout views. I don't think that's necessary. 00:41:16.685 --> 00:41:21.805 Alright. So now it's gonna start working on the skill. So it's gonna create the skill directory. 00:41:22.685 --> 00:41:23.085 Yes. 00:41:23.780 --> 00:41:32.900 So if we pay attention to skills over here, we should see a new skill. So YouTube breakout finder, we're gonna find we're gonna so create a new folder. So this is how Claude 00:41:32.900 --> 00:41:34.740 scans for, um, 00:41:34.900 --> 00:41:36.740 YouTube. So I'm gonna make sure 00:41:37.245 --> 00:41:38.365 make sure 00:41:40.605 --> 00:41:41.565 to read 00:41:41.645 --> 00:41:43.005 create skill 00:41:43.965 --> 00:41:46.925 to know how to create the skill, 00:41:47.885 --> 00:41:59.590 please. Now this thing about Claude, sometimes it takes a while, so this is where you want to think about splitting up your work into multiple parts and have multiple agents work at the same time. 00:41:59.830 --> 00:42:00.470 Um, 00:42:00.870 --> 00:42:06.495 I'm working on this on a video, so I wanna show you guys everything. But if I was doing this myself or if this again, 00:42:06.655 --> 00:42:10.015 I'd probably split this up into number one, creating the agent, 00:42:10.255 --> 00:42:13.615 number two, creating the skill, and number three, creating the MCP, 00:42:13.615 --> 00:42:31.270 trying to do all those three things at the same time. So instead of us taking, let's say, thirty minutes, we could potentially get it done in, like, twenty minutes. So I went ahead and created the skill over here, and what I'm gonna do is I'm gonna click on control shift v so you guys can read this clearly. So 00:42:32.845 --> 00:42:34.685 YouTube breakout video finder, 00:42:34.765 --> 00:42:35.725 that's the 00:42:36.125 --> 00:42:37.645 that's the name, the title. 00:42:37.885 --> 00:42:42.205 So this is YAML front matter, which is literally like a title and a description. 00:42:42.525 --> 00:42:49.510 K? So what Claude does when it looks for skills, it looks for the title and description so it knows to run the skill. So when to use, 00:42:50.230 --> 00:42:56.950 user says find breakout videos, user wants to discover, blah blah blah. Output where to output it, output it over here, 00:42:57.670 --> 00:43:01.765 and so on. Files produced, report, agent results, thumbnails, 00:43:01.765 --> 00:43:02.565 if you wanted 00:43:02.965 --> 00:43:04.165 to get the thumbnails. 00:43:04.245 --> 00:43:09.845 And then goes through the five phases over here. So okay. So now we're gonna test the workflow. 00:43:09.925 --> 00:43:16.405 Okay. So next thing, I wanna double check is review the the agent if it actually understands 00:43:16.405 --> 00:43:22.700 what what a breakout score is. K. It does view breakout score. Okay. So I'm gonna do this. I'm gonna ask, 00:43:23.180 --> 00:43:25.180 can you review the workflow 00:43:25.180 --> 00:43:26.060 from 00:43:26.060 --> 00:43:27.020 the skill 00:43:27.580 --> 00:43:28.860 to running 00:43:29.675 --> 00:43:30.395 YouTube 00:43:31.275 --> 00:43:32.555 breakout search 00:43:32.875 --> 00:43:34.555 as sub agents 00:43:34.715 --> 00:43:35.755 for those 00:43:36.795 --> 00:43:38.395 for each batch 00:43:39.355 --> 00:43:40.635 of five keywords 00:43:42.570 --> 00:43:43.530 before we 00:43:43.690 --> 00:43:53.370 run it. So I'm gonna ask it to review, and then once it's reviewed, then we can get started by testing it out. Um, this is very important because this technically is a workflow, 00:43:53.530 --> 00:44:00.885 um, because we're starting with a skill. There's, like, three phases, And then from the three phases, it's going to spawn, uh, sub agents. 00:44:01.045 --> 00:44:24.270 It's gonna go through keywords, and then the sub agents need to extract an output. And then from that output, we're gonna ask Claude to, um, analyze that output, give us a compiled analysis, and then we could use that analysis to download thumbnails. So this is a complete workflow. So I wanna make sure everything is working and make sure Claude okay. So, um, here's what Claude missed. So spawn parallel agent. So we need to do this. So looking at skill instructions, 00:44:25.625 --> 00:44:28.185 format might not be clear enough for the agent. Agent expects 00:44:28.585 --> 00:44:41.440 angle names, a string keyword as an array. So just the prompt should be more structured. Yes. Fix it, please. Also, as you as you work, um, with building skills, since I'm building this live, when you build it on your own time, like, work close like so, 00:44:41.680 --> 00:44:43.280 you'll have to test 00:44:43.600 --> 00:44:47.760 and see what works best for you and your your own business. 00:44:48.240 --> 00:44:53.200 So maybe there's a specific style of output you want. Maybe you want the agents to 00:44:53.755 --> 00:45:00.075 do something else, you can go ahead and customize it based on how you want it. K. So I think it fixed it. 00:45:00.395 --> 00:45:01.595 Did you fix it? 00:45:02.395 --> 00:45:03.995 Are you ready to go? 00:45:04.795 --> 00:45:06.315 And what I'm gonna do is 00:45:07.050 --> 00:45:11.930 I'm gonna close these tabs. I'm gonna open up a new Claude chat over here, 00:45:12.970 --> 00:45:16.330 and I'm gonna ask also, I forgot to mention beginning of the video. 00:45:16.970 --> 00:45:20.265 Ideally, we should probably be using Opus. I've been using Sonnet, 00:45:20.825 --> 00:45:22.745 so I'm gonna switch to Opus, 00:45:22.825 --> 00:45:29.465 which we should been should have been using from the start. So this is why it might have not have been making the best decisions possible. 00:45:30.105 --> 00:45:33.890 So it's important that you use the right model. Not that Sonnet is bad, 00:45:34.210 --> 00:45:49.345 but Opus is available, so might as well use the best model possible. Okay. So it's ready. So I'm gonna go ahead and open up in a new tab. Alright. So before I run the workflow, I'm gonna double check to make sure all the MCP connections are connected. So I'm gonna click control shift p 00:45:49.665 --> 00:45:51.825 and I'm gonna click reload 00:45:51.985 --> 00:45:54.225 and I'm gonna put in reload WinSrad 00:45:54.225 --> 00:46:01.265 to make sure everything is, uh, restarted just in case because sometimes MCP connection don't quite work. So I'm gonna ask Claude code. 00:46:01.880 --> 00:46:04.680 Hey, Cloud. Can we do a research 00:46:04.680 --> 00:46:14.120 to find breakout videos specifically for Cloud Code for business and productivity? Click enter, and let's see what happens. So ideally, it should open up the skill immediately. 00:46:14.605 --> 00:46:24.605 It should be able to pick up the skill, so it did. YouTube breakout finder skill to search for breakout videos, so it went through the skill and read through it. So it's gonna extract the core angles. 00:46:24.765 --> 00:46:29.660 Here are the five distinct angles. Okay. I don't don't like the comparison. Can we do 00:46:30.060 --> 00:46:37.020 AI? K. So extract the five core angles. I didn't like the fifth one, so I'd like Adjunctika AI, Adjunctika employees. 00:46:37.260 --> 00:46:43.425 So as you use a skill, you wanna make sure you guide AI because sometimes AI doesn't do the perfect thing or 00:46:43.585 --> 00:46:49.985 you would basically explain your thought process on why this is better so you can update the skill. K. So yes. Looks good. 00:46:50.305 --> 00:46:52.225 Let's con so now it's gonna identify 00:46:52.305 --> 00:47:00.980 identify 25 keywords. No. Looks good. Let's continue. So from these keywords, it's gonna look online to find, uh, videos that 00:47:01.460 --> 00:47:07.860 might match our style. So here it went. It's spawning five agents. So agent one, two, 345, 00:47:08.315 --> 00:47:20.315 and it's loading the instructions from that agent dot m d folder, which is, um, this one over here. So it's actually pulling this out. And by the way, I changed the model to Opus just to see how it would work better. 00:47:21.140 --> 00:47:24.260 I do like Opus better than Sun because sometimes Opus 00:47:24.340 --> 00:47:41.675 does do a better job in terms of tool calling. But that doesn't mean you can't use Sonnet. You could use Sonnet or Haiku. Really depends on on what you want. You wanna really customize things to what matches your business best. So as you can see here, the MCP is working. The sub agents are actually, 00:47:41.995 --> 00:47:48.395 uh, running the MCP and are getting, uh, videos. So it's going to automate business tasks, get 20, 00:47:48.635 --> 00:47:52.475 look for long long videos, and order it based on view count. 00:47:53.250 --> 00:48:04.370 So it's finding the most viral videos or the videos that got the most views for all of these keywords. And the benefit of running sub agents is that these are all are all running in the background. 00:48:04.770 --> 00:48:08.545 So it doesn't take the context window of the main agent. 00:48:08.785 --> 00:48:12.865 So once once the sub agents are done, I can continue working 00:48:13.025 --> 00:48:34.425 with this agent. So let's say I was working on a a YouTube video. I shared the YouTube idea with with the agent and I wanna continue working with it so it doesn't lose context, I would use sub agents in this case. So this is why it keeps the main context window very clean. It doesn't pollute it. And if I were to ask the main agent to do all the search itself, it probably can, 00:48:34.745 --> 00:48:36.025 but it will eventually, 00:48:36.665 --> 00:48:42.265 for certain, reach the 100% capacity of the context window, and then it will compact again. 00:48:42.585 --> 00:48:45.590 So this is a game that you have to learn is 00:48:45.750 --> 00:49:05.965 understanding the context windows limits and knowing how you could split that work with sub agents and so on. Alright. So this might take a while. So just let Cloud Code cook while you work on something else. So this is very powerful because now you do not need to do this work yourself. So if you were to do this, you'd probably go on YouTube, search these keywords itself, 00:49:06.125 --> 00:49:13.965 and probably do the analysis yourself or use a tool or a software or probably hire a VA to do this one. Now we can just have AI scrape 00:49:14.125 --> 00:49:18.200 over a 125 keywords and then from those keywords, identify 00:49:18.280 --> 00:49:19.240 which video 00:49:19.400 --> 00:49:30.520 title would you like to use the best. So now this really gives you a better chance in terms of positioning your YouTube videos to get more views. Alright. So all five search agents have completed and, 00:49:30.760 --> 00:49:38.575 um, has found some excellent breakout videos. Here's a preview of what we found. The workflow automation, how to instantly build AI agents in NA 10 using Cloud. 00:49:38.815 --> 00:49:40.255 Well, that's great. So, 00:49:40.735 --> 00:49:47.200 um, would you like to down the thumbnail image for inspiration or just save the URLs in the report? For now, let's just save 00:49:47.520 --> 00:49:49.440 the URLs in the report, 00:49:49.680 --> 00:49:50.960 and then we can 00:49:51.360 --> 00:49:52.080 select 00:49:52.160 --> 00:49:52.880 which 00:49:53.120 --> 00:49:54.800 videos we want 00:49:55.760 --> 00:49:57.520 the thumbnails from 00:49:57.920 --> 00:49:59.440 to download. So 00:49:59.605 --> 00:50:02.885 instead of doubting or thinking or guessing 00:50:03.045 --> 00:50:05.925 whether a video will get views, you might as well just 00:50:06.165 --> 00:50:11.445 figure out what works and just emulate from it. K. So now it's generating the consolidated report. 00:50:11.445 --> 00:50:17.730 It went ahead and created a new folder over here. So research, YouTube breakout, Cloud Code, 00:50:18.210 --> 00:50:19.810 uh, business productivity, 00:50:19.810 --> 00:50:25.330 and it should create a a report that we could read, which is gonna be in a markdown 00:50:24.975 --> 00:50:27.135 Alright. So it created a report here, 00:50:27.775 --> 00:50:28.415 um, 00:50:28.655 --> 00:50:41.590 with the top title, how to action and time frame, how to instantly build AI agents, how I code Progible app app solo, how to get unlimited cloud code free. So it's giving us the title templates that we could use as a foundation to build the title. 00:50:42.150 --> 00:50:47.910 K. So don't sell NA 10 workflow, sell AI infrastructure. So this could be another thing. Don't build on NA 10, 00:50:47.990 --> 00:50:50.115 build on Cloud Code. 00:50:50.355 --> 00:50:59.555 So these are ideas now I can have to create the k AI agents explained like your five, AI agents explained, so and so on. And then there's breakout videos by tier. 00:50:59.795 --> 00:51:05.430 So it's sorted based on tier. So let me do this. Let me click control shift v so it's easier to read. 00:51:05.910 --> 00:51:06.550 So, 00:51:06.710 --> 00:51:11.430 yeah, this is much better. So I can go through this, and I can figure out how to take action with this. 00:51:12.150 --> 00:51:15.510 So all these videos, the links are here, I can click on watch 00:51:15.590 --> 00:51:18.470 if I wanna check this out, and it will open up the video. 00:51:18.710 --> 00:51:28.545 Same thing here. And all the thumbnail URLs are here as well. So if you wanna download this, you can check it out. So yeah. So what would have taken you, who knows, like hours 00:51:28.625 --> 00:51:35.950 searching YouTube or hiring a VA to do this, spending a whole day to find out, uh, analytics or what what video would work, 00:51:36.430 --> 00:51:56.175 you basically have a research report that you can make data driven decisions to predictably grow your YouTube channel. And this is what I've been using to grow my YouTube channel. And if you look at my recent videos, most of my recent videos are breakout videos because I take this I follow this, uh, scientific approach. So I don't just guess what video will do well. I use 00:51:56.335 --> 00:52:02.470 data to figure out what titles work, and then I plug and play those in. So I'm gonna say, let's continue. 00:52:02.790 --> 00:52:03.670 Let's 00:52:04.070 --> 00:52:05.030 download 00:52:05.750 --> 00:52:07.270 the, uh, thumbnails. 00:52:07.270 --> 00:52:18.385 K. So now it's gonna create the thumbnails folder and then it's gonna go ahead and download them and place them in here so that we could preview them and then we could think about how we could potentially create, 00:52:18.945 --> 00:52:25.425 um, these these, uh, thumbnails for us. So now you could then remix so you could pick a title 00:52:25.665 --> 00:52:27.505 that got the highest 00:52:28.145 --> 00:52:33.420 breakout score, and then you can pick the thumbnail that got the highest breakout score and then just combine them, 00:52:33.740 --> 00:52:43.500 and then you have a higher chance of getting more views on YouTube. Alright. So apparently, there was an error, so I asked Claude to fix it. So now it's actually downloading it. So if I click on one of these, 00:52:44.365 --> 00:52:45.885 pull this on the right side. 00:52:46.605 --> 00:52:57.565 So this is one of the thumbnails, so this could give me an idea. Alright. So CloudCoder is downloading all these thumbnails. So if you check them out so you can you you can potentially get inspiration from these like so, 00:52:58.460 --> 00:53:02.860 and then you can get some idea. Most of these has a face on the right side 00:53:03.580 --> 00:53:05.980 and then the and then something on the left side. 00:53:06.220 --> 00:53:08.780 So now you can start to see the patterns. 00:53:09.340 --> 00:53:18.725 How can you create a thumbnail that will for sure get viewed. So, yeah, now you can see how Cloud Code can save you hours of your time instead of you downloading all of this manually, 00:53:18.885 --> 00:53:20.885 needing it to organize in the folder. 00:53:21.125 --> 00:53:38.520 Cloud can do that. If you want to search online, scrape stuff, Cloud can do a lot of work. So this is just one of the examples. Alright. So next step is let's say you have this report and you have all these thumbnails and the next thing you wanna do is you wanna store this somewhere where someone in your team or other people can easily view it because having it all stored, 00:53:38.760 --> 00:54:00.380 uh, within your local files might not be the best way for other people to view that. So let's say I wanna connect this to, uh, Notion. So I went ahead and I asked Notion to create this database, then it created a quick database for the breakout scores with all of the properties already. So what I'm gonna do is I already have Notion MCP connected to my cloud code. If So I go to your MCP 00:54:00.700 --> 00:54:03.260 and go to manage MCP servers 00:54:03.420 --> 00:54:19.605 and continue in terminal, so I can see that we have YouTube connected and we have, uh, user MCP's which is Notion Notion hosted. So for this one I'm gonna use Notion hosted because I found it to work better. Uh, Claude was able to use it. So I'm gonna use that. So I'm gonna close this over here. 00:54:19.845 --> 00:54:20.965 Close this tab. 00:54:22.005 --> 00:54:27.420 Close that. And we're gonna get started. So what I'm gonna ask is I'm gonna make sure 00:54:27.580 --> 00:54:28.220 that 00:54:29.100 --> 00:54:30.540 this one over here, 00:54:30.940 --> 00:54:32.460 I'm gonna take this link. 00:54:38.780 --> 00:54:50.865 I'm gonna paste that over here and k. So I'm gonna I have the copy I have the data source ready so I'm gonna ask Cloud Code. Hey, Cloud Code. I like to create the slash command that once I'm finished with creating a YouTube report, 00:54:51.105 --> 00:54:56.390 I'd like us to save all of the videos that we identified that are breakout videos 00:54:56.470 --> 00:54:59.670 and the thumbnails into a Notion database. 00:54:59.830 --> 00:55:04.230 Now I want to make sure that you fill out all the properties in this data source 00:55:04.470 --> 00:55:07.110 and we also paste the thumbnail 00:55:07.375 --> 00:55:13.775 within the page content so we can, uh, see it in the gallery view so can quickly, uh, scan through it. 00:55:14.175 --> 00:55:15.615 Um, does this make sense? 00:55:16.815 --> 00:55:21.135 So I'm gonna make sure I put it in plan mode again because again I'm gonna be building something. 00:55:21.870 --> 00:55:23.630 If I were to just ask it, 00:55:24.110 --> 00:55:30.910 uh, right away then it would probably make a lot of mistakes because it will forget some of the context. So so first it's 00:55:30.910 --> 00:55:35.470 checking if it can find the the, uh, database which it did over here. 00:55:36.350 --> 00:55:41.865 It found it and then it's gonna explore the skills in YouTube data path. So over here it's reading through the report 00:55:42.105 --> 00:55:48.345 and it's trying to learn. It read through, uh, the skills and so on. I think it actually spawned a sub agent over here 00:55:49.225 --> 00:56:06.785 while it's doing this. So Claude can spawn sub agents and still work on stuff. So if you wanna do that, just make sure you ask Claude whenever you're, like, working on a task. Just say, hey, Claude. Can you spawn a sub agents in the background? And can we continue working on this while those agents work? So this is why Cloud is super 00:56:07.105 --> 00:56:21.040 powerful, man. You can have as many sub agents working in the background for you. So what would take you eight hours of work before will take not two hours. It'll take, like, thirty minutes or ten minutes. So you can get a lot of work done. Technically, I could ask Cloud Code to 00:56:21.680 --> 00:56:29.760 to spawn, like, I can have five different tabs and I could do research for five different videos at the same time and get five different reports 00:56:30.005 --> 00:56:35.045 if I really wanted to. So, hypothetically, if you're working with, like, let's say, five YouTube clients, 00:56:35.285 --> 00:56:37.205 if you're an YouTube agency client, 00:56:37.365 --> 00:56:43.685 you could do that, pretty much do that, and then serve five clients at once instead of you needing to hire five VAs, 00:56:43.925 --> 00:56:56.170 just have Cloud Code and you spend, like, what, a $100 a month, $200 a month, and you get, like, a lot of value out of that. And this also extends to, like, other stuff in your business. This is in this video, I'm just going over YouTube. 00:56:56.250 --> 00:56:58.810 This could be something else. This could be, like, writing content. 00:56:59.435 --> 00:57:00.715 This could be like, 00:57:01.035 --> 00:57:05.995 um, breaking up content from long form into short form. This could be like any task 00:57:06.155 --> 00:57:15.910 that you would that AI could potentially do is completely scalable now with Cloud Code because just the ability to spawn sub agents is very high leverage. 00:57:17.190 --> 00:57:23.910 Okay. So, apparently, it found our existing breakout video database. It's gonna fetch the schema to understand the properties. 00:57:24.630 --> 00:57:26.310 So now it has a complete picture. 00:57:27.355 --> 00:57:39.595 Let me check existing research output and understand how breakout videos are stored after research YouTube research session. So now it has everything. It's gonna create a plan for us. Okay. So here's the plan. We're gonna create a slash 00:57:39.130 --> 00:57:51.050 save to Notion and read through this. So create a slash command slash save to Notion that saves breakout videos from YouTube research report to existing Notion database with, uh, thumbnails embedded. Okay. 00:57:52.385 --> 00:57:53.025 Cool. 00:57:53.905 --> 00:57:54.625 Cool. 00:57:58.065 --> 00:58:00.865 Okay. Cool. This works. Okay. So I'm gonna ask it, 00:58:01.105 --> 00:58:01.985 change 00:58:01.985 --> 00:58:03.505 the slash 00:58:03.585 --> 00:58:04.225 command 00:58:05.505 --> 00:58:07.105 name to 00:58:08.440 --> 00:58:09.320 save 00:58:09.560 --> 00:58:11.160 YouTube to Notion. 00:58:11.720 --> 00:58:13.160 Because I wanna make sure 00:58:14.680 --> 00:58:36.205 it's because I put save to Notion, this might be something else. Let's say I wanna save something else to Notion. So wanna be careful where it says save YouTube to Notion. But other than that, I think it's good. So it updated the plan, and I'm gonna say yes and auto accept. If you wanna review in more detail, then you can say yes, manually accept. So it's gonna create the command file. I I hope that it creates a command file correctly. 00:58:36.970 --> 00:58:40.890 So I'm just gonna be really safe here, and I'm gonna say double check. 00:58:41.290 --> 00:58:44.730 Can you double check the format of the 00:58:45.050 --> 00:58:45.930 command 00:58:46.890 --> 00:58:47.850 of the 00:58:49.370 --> 00:58:50.490 slash command 00:58:50.570 --> 00:58:51.370 with 00:58:51.450 --> 00:58:53.530 create 00:58:52.535 --> 00:58:53.815 command skill. 00:58:54.135 --> 00:58:57.255 K. Because I did the work previously in creating this, 00:58:57.575 --> 00:58:59.975 um, repository or workspace 00:59:00.295 --> 00:59:09.080 by having all of these structures set up previously. So we're gonna make sure. So that's why I asked it to reference it because sometimes again, this is an example of context. 00:59:09.640 --> 00:59:10.200 Like, 00:59:10.600 --> 00:59:14.840 even though, like, Claude might have read it at the start of the chat over here, 00:59:15.320 --> 00:59:22.295 it went through my whole code base. It still forgets things. So this is why I make sure to remind it. So so it went ahead 00:59:23.415 --> 00:59:24.775 and updated it, 00:59:26.135 --> 00:59:28.375 so hopefully it will work. K. So 00:59:28.855 --> 00:59:37.730 we'll know if it works. If I type slash and I'll be able to see it. So if I put type slash YouTube to put type slash save to Notion, I still don't see that. 00:59:38.130 --> 00:59:44.130 So we're gonna we're gonna check it now once it creates it. So I'm gonna go ahead and refresh reload window 00:59:44.505 --> 00:59:54.185 so that if the command is working we'll be able to see it. Once Claude creates the slash command, I can open up a new chat and I can type slash save 00:59:54.505 --> 01:00:05.670 and I should be able to see it. So save dash YouTube to Notion. If you don't see it then you always again, you wanna make sure you reload the window or you make sure that Claude actually created the command correctly. So I'm gonna click on this 01:00:05.910 --> 01:00:07.750 save tab 01:00:07.910 --> 01:00:10.550 and then I'm gonna make sure that I select 01:00:11.430 --> 01:00:14.495 the actual report. So I'm gonna ask it to, 01:00:15.135 --> 01:00:16.175 uh, cloud 01:00:16.975 --> 01:00:17.695 code 01:00:17.775 --> 01:00:18.495 business 01:00:18.815 --> 01:00:19.775 productivity 01:00:20.095 --> 01:00:26.890 because let's say you have multiple reports and you want one specific report, you don't want another report. So you wanna specify. 01:00:27.050 --> 01:00:30.330 So I asked save YouTube to Notion, Cloud Code business productivity, 01:00:30.730 --> 01:00:44.775 and Cloud said, I'll save the breakout videos from YouTube research to Notion. Let me find let me first find the most recent research report. So I was gonna go ahead and save all 12 videos. So let's do this. Let me see if I could open this up so we could see it work at the same time. 01:00:47.415 --> 01:00:49.575 Put on Notion here and 01:01:04.560 --> 01:01:09.985 Alright. So as you see, it just added all of the videos in one shot. That's sick. That's so cool. 01:01:10.305 --> 01:01:13.185 And then I wanna make sure the thumbnails are shown. 01:01:14.065 --> 01:01:17.105 So great. Even the thumbnails are here. Awesome. 01:01:17.505 --> 01:01:25.760 So now it's gonna add the next three videos which is the five x tier. So if I work with, uh, Notion, I can always go here to property visibility 01:01:26.080 --> 01:01:31.520 and I can put out the breakout score and I could put in the, uh, channel 01:01:32.160 --> 01:01:34.000 number subscribers so I can adjust 01:01:34.875 --> 01:01:38.075 what I want. I can put the the URL as well 01:01:38.635 --> 01:01:41.195 and I can even sort it 01:01:41.595 --> 01:01:45.275 based on, uh, breakout score or views. 01:01:45.675 --> 01:01:50.010 So I'm going do by breakout score. So apparently this one got the highest 01:01:50.810 --> 01:01:52.010 breakout score 01:01:52.410 --> 01:01:54.490 for this title and this thumbnail. 01:01:54.570 --> 01:01:57.290 Right? So you can now use this as an analysis. 01:01:57.690 --> 01:02:57.830 So, yeah, this is sick. This is awesome. Now you could get work done and share it with your team on Notion by connecting YouTube and Notion MCP. So in less than an hour, we built a complete AI system that gets YouTube breakout videos for us. Now what would happen if you spent three weeks in building AI systems and how much more you could get done? That's exactly what I'm doing in a new challenge called your first AI employee, I'm running on February. Our goal is to first x-ray your complete business, so we're gonna find opportunities where you could use AI agents in Cloud Code, and then we're gonna delegate those activities to Cloud Code so Cloud Code can become an actual employee in your business. And, ideally, this employee will generate value in your business and generate business outcomes where, at minimum, our goal is to have Cloud Code generate you $10,000 a year extra in value generated or revenue generated. Now if you want that, I put that link in description. We're gonna get soon, and I'm only taking 10 spots. And if you found this video valuable and you want to apply what we just did, you can rewatch this video again, and you can grab the GitHub repository, 01:02:57.910 --> 01:03:03.830 uh, starter link in the description below. And a like and subscribe would be very helpful. I'll see you guys in the next