Transcription
Cloud Code is one of the most powerful AI tools ever released. But most people are using maybe 10% of what it can actually do. I've spent thousands of hours inside Cloud Code building apps, automations, internal tools, and the gap between how beginners use it versus how it's meant to be used is massive.
In this video, I'll show you the system I use to get 10 times more out of Cloud Code, whether you're building apps, writing code, automations, or just trying to work faster. By the end, you'll have a framework you can apply to just about everything.
But before we start prompting, here's what separates the 99% from the 1%. Most people actually just open up Cloud Code and start typing. That's like trying to build a house without any blueprints. So, I want to show you a proven system that you can follow, even if you've never touched this stuff before. And it's called the SCOPE framework.
And the SCOPE framework starts off with setting up Cloud Code properly, making sure you have everything configured. So, let me show you how to do that. Okay.
So, to set up Cloud Code, what we want to do is go to Cloud Code setup. So, I'm just going to Google that. Okay. Then we come to this page and you see right here, depending on your operating system, all you have to do is copy this, open up your terminal, paste it here, and then hit enter. Okay. And then it's going to take a while, but that's all you have to do. So, that's step number one. Okay.
Step number two is you want to download something like Cursor or VS Code. VS Code is the lightest, so I'm going to download that. So, you just go to VS Code and everything is free. You just download that and then you want to open up VS Code after you have it downloaded. After you have VS Code downloaded, you want to go here and then hit this extensions button and then type in Cloud Code. After you hit Cloud Code, you hit install. Okay? And then right after you install, you'll see this right here. So, you want to hit this button and then you'll see Cloud Code is going to take you through some steps where it just wants you to open up and verify you've paid for it. After you have that, you see you have Cloud Code installed and it's really easy to work with. So, we're going to be mainly be working with this chatbot right here and I'll show you how to get everything set up.
All right. Now that we have set up Cloud Code, what we need to do is control what it does to make sure it's predictable in its output. And the reason we have to do this is because LLMs, their brains are inherently crazy. They're all over the place. What that means is let's say you ask a question to it one day, it's going to give you one answer and then you ask the same question to it 5 minutes later and it gives you a vastly different answer. This does not work in business. What we need is predictability and reliability. Okay, so if we want Cloud Code to be predictable and reliable, how do we do that? And the way we do that is by putting Cloud Code's brain. So let's say this is Cloud Code brain inside a box and we give it some directives. Okay, that is done through this thing called Claude.md. And what Claude.md is is just preloading Claude with some information to make sure you create this box before it starts doing any random stuff. Okay, you can think of the sides of the box with this framework and it's called the DOE framework.
And what the DOE framework really is, and I got to give some credit to Nick Rev. He has been making some really good videos about this and I've been pushing this to its limits. So what the DOE framework is is a simple way to control Cloud Code so that it's reliable and it doesn't have that LLM spaghetti, right? You can think of the directives as basically just the goals and the rules of what you want to do. So for example, like find me 50 plumbers in Miami and a personal one opener. This is like a high-level goal, right? What the LLM will do is take this goal and then create the subtasks it needs to do to achieve that goal. So that's the orchestration layer and that's the agent's brain. So it decides that first it needs to search, then it needs to scrape, then it needs to enrich whatnot. Okay. After that it goes to execute on what it deems to be the correct plan. And how that works is that it writes Python files. And Python files are purely deterministic. So if you guys have used like Naden or something like that before and you've seen a workflow where it's like node next node next node, a Python file will work like that. And you'll basically write a script that does the same thing over and over which means whatever Cloud Code does is going to be predictable and reliable. Okay.
So now that you understand what the DOE framework sort of is, it's actually really easy to install. So let me show you how to do that. Okay. So let's set up Cloud Code with the DOE framework. And just to make it easy for you guys in my community in YouTube video resources, by the way, all of this is free in the link below. You're going to come across this agents.mdcloud.md file. All you need to do is copy this. Okay. You copy this and then in your folder where you've opened up Cloud Code, you make a new file and then you do Claude.md. Then you paste that in. Then all you say is initialize my environment please. Cool. And then you wait. So what Claude will do now is essentially it will read the doc we just pasted in and based on that document it's going to create your environment so everything stays contained and you'll see it'll ask for some stuff like um some permissions and I just like saying yes and don't ask again. All right.
So, what you'll notice right now, it's asked to do a couple of different things. And right here, it's made two different folders, directives and execution. So, inside the directives folder, all the high-level goals of what we want to achieve is going to live. So, it's going to be more .md files that you can just read in plain text. And in the execution folder, all the Python scripts that Claude will execute will live in there. And this Claude.md file makes it so that you have set up Claude where it will behave the same way all the time. All right. So you see it says the environment is now ready for your three-layer architecture. You can start adding directives and executions as needed. Cool. So we're good to go there.
All right. So now that we have Cloud Code set up in a controlled environment, now we're going to be setting up the plan of something we're going to be building live together and that is the orchestration part. So you can think of the orchestration piece as you being a composer and using Cloud Code which is the people to create whatever symphony or automation you would like. And to do that you would need to equip Cloud Code with some tools. So for example saying like use Airtable or use Google Sheets and this will be more evident as we dive into the build. So let me show you what we can build so you can learn all the concepts and whatever idea you have you can turn that into reality with Cloud Code.
And now we want to kind of design the system we want to build. So I really encourage you whenever you want to build an automation or a system that you kind of draw it out like I am right now. Okay. So I'm going to be explaining how I develop systems and then we're going to be taking that exact three-part formula and develop the automations with Cloud Code. All right.
So what is a system? Well, a system is basically consists of three things which is inputs, processes and outputs. And these processes can activate based on a trigger. Let's say it's 8 a.m. every day, it triggers or it's triggered by me manually. Cool. So if a system consists of the inputs, the processes and outputs. So the whole idea is we're going to use this framework to solve a problem, right? Okay. So what is the problem? All right. So let's say the problem we want to solve with this system is we want AI to come up with new ad creatives so we can grow our brand better. All right. So if the problem that we're solving is that we don't have enough ad variety, what are the inputs? Right? So the inputs that means the system will already have to know about the context of our product. So the inputs, a product context is an input. And then what would we do with this product context, right? We're going to do something with it. So we're probably going to come up with a new ad angle. And of course, because it's AI, not everything it's going to come up with is going to be good. So we should have a human in the loop or an approval process. So if good, then let's say we want an image with Nano Banana Pro to be generated. Okay. So the output right here would be would be the created image. Okay. Maybe we want this happening every day at 8 a.m. so we can review a bunch or we have an idea and then we manually trigger it. So let's try to develop this system.
All right. Now that we know what we will be developing, we need to think about how we will actually use it, right? And the way I'm envisioning it is kind of like a Kanban board where let's say there's an Airtable and I'm going to split it into two tables. Table one, table two. So table one, let's say it holds the context. So I can always edit the context whenever I'd like. Right? So what would the context be? So the context in this table it would hold the product capabilities. It could have constraints or ICP use cases. That should be enough. Okay. Now our AI will use all this stuff to develop the stuff in table 2. And how would I use something like this? So I'm imagining it would be kind of like a board where I have a card here. Let's say this is insight and let's say this is good and let's say this is approved. Oh, this isn't working. And this is published. Okay. All right. So, what we're going to do is if something is good right here, we want to generate an image. And then this will probably post with something. But I think we have a basic understanding of what we want where we'll have an Airtable with some context, a context table, and then we'll also have another table where these insights will be generated and then if I like the insight, then it's going to generate an image. All right, cool.
Now that we have that done, let's go build it. All right, so now that we have our plan and what we want, let's go back into Cloud Code. Let me close this out because all we need is this chat box. So, the next thing we want to do is plan out what we will do. So, I want to hit this again. And now you see how we're in plan mode. And I'm going to be right here and using my voice to talk to Claude and say what we want to do. Hey, so I want to build an automation where for my product Booked in AI, I want to have an Airtable and in that Airtable I want two tables. I want one table to contain the context. So stuff like what the product can do, its capabilities, etc. And then the other table I want it to be a Kanban board where it will generate new insights and prompts for images to be created. Okay. And after that is created, what I want is when I drag it over and approve it, it generates an image with Nano Banana Pro. I think what I need to do for my plan of action is first give you access to Airtable's MCP so you can build the table for me and then I want you to go do research and find an API that I can use to create these images for Nano Banana Pro. Let's make the plan.
All right. So I just spoke to Claude on what the plan I want is and now it's just going to load. It's going to keep thinking and I'll get back to you when it's done. All right. So, this is really cool. You can see how Airtable setup. How would you like to set up the Airtables? Use an MCP to create it. And it says recommend. So, I'm just going to click that. And then it says what am I going to use for our image API. So, I'm just going to select maybe Wavespeed. And in terms of triggering, let's just do it manual for now. Yeah, I like the first one. All right. So it got done with the plan. I'm just skimming it over where it literally got down what we want to do. It's going to be using Wavespeed. Okay, the Airtables. It knows what we want. Honestly, we don't really need to know much of what's going on. It's more so that hey, this made a plan. Cool. Now you can follow this. All right. Yes. And auto-accept.
All right. So this part is really important guys. So right now it's saying API keys needed and it wants the Airtable key. It wants the Wavespeed API key and the Entropic API key. So let me go and find those and I'm just going to put it all in here. This part you can always just ask Claude on what to do. But for Airtable you can just create a token, take that, paste it in. And Wavespeed, uh, right here they have an API key which you can just copy and then paste it. So let me just paste those. And just so because this is sensitive information, I'm just going to skip through to the next stage. Okay.
So at this point, what I did is I took my keys, I put it into Cloud Code, and Cloud Code updated my .env file. So all those keys Cloud Code has access to use them and I don't have to give it again. And also you can check that it has access to the MCP and you can do that by doing /mcp and if you do mcp status and you see Airtable MCP connected. Okay, that's awesome. So now we have given Cloud Code the power to do external actions as well. So now where are we? So we have a plan that Claude will execute. We have given it access to the MCPs and all the resources it needs. So now it's just time to start executing on the plan. So what we want to do is let's build the MVP of the system we want and then iterate over that. What we can do is let's start making an Airtable base. Okay. So let me make a new Airtable base. Let's call it Ad Generator. Great. Now let's go back to VS Code and let's ask Claude to do the rest. So this is my plan. Okay. So, uh, I made an Airtable base called Ad Generator. I want you to follow the parsed stargazing.md plan and build out the tables based on the plan.
Okay, great. So it's saying your Airtable structure is ready based on what we asked it. Let's go look at it. Product context, content pipeline. Okay, nice. So, we have a bunch of stuff here. Now, let's start actioning it more. I want you to populate the Airtable book context based on any information you find in Bookton.ai and uh booked context.md. Okay, so what I did right now, I just told it go populate this Airtable base. Go do your research from my website and all that stuff and make sure this product context is populated. Cool. So this thing got done. So it should have populated my table. All right. So looking at it now, it has all my context. So it went to my research, my website, my YouTube videos and built all this stuff out. You could have also just dumped it in. And here are the ad insights. Let me look at it. White label AI agency gold. Okay. It has a target audience and I guess this is where all the images would lie. Okay, so it populated some mock data. Let me open this up. Okay, so there's some images here. You can see it also showed that it used Nano Banana Pro. If I scroll down, cool. So here's an image it kind of created. You can tell it's Nano Banana Pro because all the text is correct. This one looks pretty bad. Okay. Why is that? Oh, okay. Image model is Flux. So it lets us use a couple different image models as well. Cool. So it's also adding these little things in. All right.
So now we have a working system and all we did to build this was just prompt it. All right. So why this is so cool is because you see how Cloud Code has literally created these directives which are the instructions the AI is going to follow to build these and it knows about if anything has gone wrong in the past and also it has created these Python files which I don't know what it does but the cool part is that if something breaks, it knows it based on the directives and it'll fix it and make it better. So all this means is because we set up Cloud Code in the right way from the start, we are now able to build reliable systems because it's not the AI that's generating random stuff. It is done through just triggers of these Python files. And all of that is just done through setting it up and by talking to Claude. Great.
So now we've built the first version of the system we want. So the next step becomes our responsibility to polish the system and make it a lot better. And all that is done is just by going through and talking to Claude. And let me show you what I kind of came up with. So this is more or less the same system. And you see I changed up a bunch of the stuff just by prompting. And now it comes up with way more in-depth understanding. And I also added functionality and features. Now you can see it comes up with these kind of short style images where it tells a story. And then also I said, "Hey, can you make sure that we make scenes of this?" So there's a start frame and an end frame and then there's some motion uh if I like it. So for example here, this is a start frame and an end frame of one scene and I also with the same API got it to make a video like this. You see how now I can giving life to what I built. And all of this is done was literally just by prompting the same engine over and over. That's basically just me polishing the system over and over by talking to Cloud Code, judging the outputs. I'm reviewing if I like it, if I don't. and I'm using Cloud Code as my workers to build iterations and then I judge if I like it or not and that's based on my taste and then I tell it what to do and it becomes as simple as that.
Okay, now that we've gone back and forth with Cloud Code and built the system to a point that we like, we want to make sure we can export the system so it's sharable and it's always running in the cloud. This makes it so your computer doesn't always have to be on to be using these systems with Cloud Code. And let's say you want new ad sets to be created at 8:00 a.m. every day so you can review them. You can do that. So let's get into exporting these kind of Cloud Code systems and how you can take it, give it to other people or run it based on any triggers that you would want, which begs the question, how do we share these systems? And the way we're going to be doing that is with a service called Trigger.dev. And the reason we use Trigger.dev is because it has an MCP and it's code-based, which essentially means we don't have to know how this works. We can just tell Cloud Code, hey, here's the MCP. Go figure this out and use it. And it also really helps that it's really, really cheap. So this is Trigger.dev. You can just get started. And you see I already have a bunch of workflows here. What you want to do is if you just Google Trigger.dev MCP and then the first one is all you need to do. There's an automatic installation. So you just copy this and then let's go back to Cloud Code. Can you install the Trigger.dev MCP for me? Ask me for the API key when ready. Okay.
So now I'm telling Cloud Code to just install the Trigger.dev MCP. And now what I have to do is go here and find my API key by going down here to API keys and then you just hit copy here. Okay, after you hit copy here, let's go back to Cloud Code and set. I'll provide it in the text. All I said was, hey, can you install Trigger.dev MCP for me? And then it's asking for my API key. So I'm going to paste it here and then come back. All right. So it got done. Now I can verify by doing /mcp and then doing mcp status. And then you can see we have access to Airtable and Trigger. Okay. Great. Now I can say, hey, can you package all this up and then use the Trigger MCP to deploy this in the cloud and run this workflow every day at 8:00 a.m. So it generates me new ad creatives or insights that I can review every day. And then you hit go. So that's about it. So, uh, I'll let it work for a little bit and come back to you, but that's about it. So, now Cloud Code has context of your entire project and it's going to take the Python files it wrote. It's going to deploy it to Trigger.dev and then run it every day. Now, essentially, you would have a system like I showed you where every day it's going to come up with, let's say, strategies or stories like this, really cool ones, and if I like it, I can just drag it over. And then if I like that, it's going to generate these kind of images. And if I like that and I drag that over, it will generate the videos that I showed you before.
So that's an overview of how we kind of use Cloud Code like a crazy, crazy power user where now we're getting it to automate things that we don't want to do and we don't need any know-how. We just need to understand some little concepts, do the setup properly, and make sure that this agent is running in the background on its own and doing any tedious work for you. So that's SCOPE. That's the framework you can use to be better than 99% of people using Cloud Code. You have to make sure you do the setup correctly, then set the right controls in place, make sure you orchestrate what the agent needs. You polish it up by talking to the agent back and forth, and then you use services to export all these workflows. So, they're reliable systems that live in the cloud.
But here's the thing. The real power really comes when you start building AI agents, not just little apps like I showed you. And Google actually just released something that makes this ridiculously easy. So, click this video here because I'm going to show you how to build AI agents with Google's new anti-gravity. And we're talking about voice agents, automations, the kind of systems that actually replace manual work. Everything I showed you with Cloud Code, but taken to the next level.