Transcription
Have you ever spent three hours building an automation only to have it break the next day? Ain't nobody got time for that. I have more times than I want to admit. I've built probably 3 to 600 automations by this point with Zapier, Make, Nan, you name it. So, I'm going to tell you something that might sound controversial. You need to stop building AI automations and start building agentic workflows. Let me explain.
Enthropic recently supercharged Cloud Code, and it's really completely changing how we build AI systems. No-code tools used to give us a visual way to build. But you still have to manually connect all the dots and troubleshoot when things break. But now, if you set up Cloud Code the right way, you just describe what you want in plain English, and Cloud Code figures out the how, writes all the code, and even corrects itself when it runs into errors. So, in this video, I'm going to show you the framework I use for building these agentic workflows. Then, we'll use Cloud Code to build an AI agent that scrapes multiple sources, analyzes the content, and generates even new content for me to publish. So, by the end of this video, you'll understand why this is the future of automation, and you'll be able to build your own automations just by speaking. Let's get into it.
So, what's actually wrong with the way most people are building automations right now? Here's the thing. No-code tools like Nan, Zapier, Make, they were a massive step forward. For the first time, you didn't need to be a developer to connect these apps together. You could just drag, drop, connect, done. But here's the problem. Look at this. This is an automation I built a while back. And when I built it, I thought I was being clever. Look at all these nodes, all these connections. It's doing a lot of things. But now, I have absolutely no idea what half of this stuff does. And that's the first problem. These workflows turn into visual spaghetti. Think of it like this. Imagine you're building a Jenga tower. Every block you add is a step in your automation. The first block, let's say, triggers when a new lead comes in. The second sends it to a spreadsheet. The third, a Slack follow-up. Stuff like that. Now, you've got this tower, and it works. It's standing. But what happens when one of those blocks change? Maybe there's an API update. Maybe the spreadsheet column gets renamed. Maybe you change the name of the Slack channel. One block shift, and the whole tower comes crashing down. And now you're spending three hours digging through nodes trying to figure out which one broke.
But here's the thing, that's not even the real problem. The real problem is you, the orchestrator. Every step, every connection, every piece of logic, every "if this, then that," you have to define it manually, one by one. You're not building an intelligent system, really. You're building a very complicated recipe, and if anything goes off script, it breaks. You're basically acting as the brain for the system that has no brain of its own. And I'll be honest with you, I spent years doing this, building elaborate automations, feeling proud of how complex they were. But complex doesn't mean smart.
So, what if there was a different way? What if, instead of manually connecting every single step, you just described what you wanted, and the system figured out how? And that's exactly what we're going to be building later in this video. But first, let me show you the framework that makes this possible. And this is where agentic workflows come in. It's a completely different way of thinking about automation. So, let me explain what I mean by that.
With traditional automation, you define every single step. You're the architect. You're the builder. You debug, too. You're everything. But with agentic workflows, you define a goal, and the agent gets to figure out all the steps. That's it. That's the shift. You stop being the orchestrator and you become a director. Think of it like this: an orchestrator is in the pit, waving their arms, telling every single musician when to play, how loud, how fast. If this conductor stops conducting, then the music stops. But a director says, "I want this scene to feel tense, and the audience should be on the edge of their seats. Go." And then the actors, the camera crew, the lighting team, they figure out how to make all this stuff happen. They curate this taste. That's the difference. One is way more exhausting, and one is very leveraged. And once you know how to harness all this, it's a crazy difference. And that's why I'm so excited to talk about this.
So, how do you actually do this? How do you become the director instead of the orchestrator? And this is where Cloud Code comes into play. And the first part is making sure we set it up the right way. Now, think of Cloud Code like a conversational coding environment. Here's the thing most people miss. This isn't just an AI that writes code. That's ChatGPT. That's Copilot. That's been around for a while. This is different. A code generator just gives you the code. You still have to run it, test it, debug it, all that. An agentic system does the whole loop. It writes, it runs, it tests, it fixes, and repeats until it works. The complexity doesn't disappear. It's just no longer your problem.
So, just to bring this back where we started, remember the visual spaghetti? All those nodes, all those connections, all that manual configuration. I'll show you a way we can get the same results, but the complexity is all going to be managed by the AI, not you. What we'll do together is we'll describe what we want and get the agent to figure out how to implement it. That is an agentic workflow. Now, you might be thinking, "Okay, this sounds great, but how do you actually structure this? How do you give the agent the right instructions so it doesn't go off the rails?" And that's exactly what we're going to cover next. I'm going to show you the three-layer framework I use for every agentic workflow I build. It's the difference between an agent that flails around and an agent that executes like an actual employee. So, let's get into it.
Okay, so we established the problem, and now we see what an agentic workflow looks like. But here's the thing. Just knowing that Cloud Code exists doesn't mean you'll get good results out of it. You need a framework to use this tool properly. And I like calling this the agentic stack. This is basically a step-by-step set of instructions written in plain English. You're answering one question: What do you want to achieve? So, for example, let's say I want to monitor five AI creators on Instagram. I want to scrape their last posts, summarize key insights every day, and I want a daily digest so I can stay on top of what's happening in the space and remake the content, possibly. And that's a directive. That's the job we're going to get the AI to do. That's defining the goal. So that was layer one.
Layer two is the orchestration. That's the middle layer. This is the "how." This is not your job anymore. This is where Cloud Code lives. It takes what you just told it, that's the directive, breaks it down into steps. It figures out what tools it needs and what order to do things in, and how to even handle edge cases. It's doing the thinking you would usually have to do. Remember the orchestrator from earlier, the one in the pit waving their arms, telling the musicians what to do? This is that layer, except now AI is the orchestrator, not you.
Then we move on to layer three. That's the execution. This is actually playing the instruments, the doing of the work. And to do that, what we do is we get Cloud Code to write the actual code. Like, for example, Python scripts, calling the APIs, organizing the files. Cloud Code writes all the stuff. It runs it. It tests it and fixes it when it breaks. You're not supposed to touch any of this stuff. You don't even look at it unless you really want to.
Now, here's why this matters. Before, you were responsible for all three layers. You needed to define the goal. You needed to figure out all the substeps. You wrote the code. You fixed the bugs. Now, you're only responsible for one layer: the directive. The AI handles all the orchestration and all the execution. You went from doing three jobs into one. And this is where all the leverage comes from. Think about it. Most of your time wasn't spent on the goal. Most of your time was spent on, "How do we get this damn thing to work?" Figuring out the right sequence, debugging the code, handling the edge cases. But with the agentic stack, you're reclaiming all that time. You focus on what you're good at: knowing what needs to be done. And the AI focuses on what it's good at: figuring out how, and actually doing it. This is a completely different way of working. And I'm super excited about it.
Okay, so the DOE framework. So now you've got that framework, and you understand what agentic workflows are, and you've got the mental model to structure it. Now, let's actually build one. I'm going to show you exactly how I use Cloud Code to build a content creator scraper. That's something that monitors all the AI creators I like on Instagram, pulls their latest content, and summarizes it for me, just so I can recreate the content if I want and grow my audience, all by just describing what I want. Let's get into it.
Now, look, understanding this framework is the first step, but actually applying it to your own business problems, that's where the real power is. Actually, if you guys want, I've put together an agentic stack. It's in the link in the description below. It's free. It's in Simon's resource hub inside YouTube video resources. And we'll be using these to get started. Okay. It walks you through how to identify the best opportunities for automation in your business and gives you three ready-to-use prompts to build out your first agentic workflows with Cloud Code. Links in the description, and it's the perfect next step after the video.
Okay, so let's start actually building this with Cloud Code. We're going to be building the content creator scraper we talked about initially. And to do that, we're going to get started. And to get started with Cloud Code, you're going to need one of the pro plans. And then after you have the pro plan, you want to download the Cloud desktop from cloud.ai. Okay. And then after you have the plan, you'll see right here it says Code. So you want to hit that Code button. All right. Then you see it's going to ask you what's going to happen. And all you need to do is just interact with this chatbot. It's going to be super simple. Cloud actually Cloud Code made it way better. So now I'm on Mac, but what it first you need to do is make sure you select a folder. [snorts] So I'm just going to go to my Finder and I'm going to right-click and I'm going to make a new folder. Let's say "content scraper." Cool. Now that we have that, let's go here. Let's hit open, and then "content scraper." You see that it's empty. Let's open that. Okay, great. Now what we want to do is let's go back to that chat and then you see we have "Get Started." So I'm going to take this agents.md. All right. And you see this has the three-layer architecture we talked about: directives, orchestration, and execution. So we're going to be taking this and then using that to build all our automations. So we just take it, go here, we hit copy, and then let's go back to Cloud Desktop and right here I'm going to paste this here and I'm going to ask, "Hey, can you please create a claude.md for this?" Okay. And I'm going to say, "Trust this workspace." Cool. And I'm going to hit allow. And now I want to make sure we're only just chatting with this chatbot here to get our work done. And you see it's going to ask you if you always allow. You can hit allow. Okay, cool. So this is done. Now, just to verify, I want to show you this. If you go in here, okay, so this is still happening. It's not done yet. And allow for session. I'll say, "Allow. Always allow for session." Cool. And I'm just going to let it run. Okay. So, you hear that little ping when it's done. [snorts] Cool. All right. Now, let's go back to our Finder. And you can see this claude.md file. Okay. I don't want you to touch it. I just need you to know that it's there. If you open it up. Okay. So, if you open it up, you see it's the exact same thing that we copied. All right. Okay, so that's good. Now let's say, "Hey, based on the Cloud MD, can you set up our environment?" Okay, great. So now it's set up our environment with all this stuff. And if I go back to my Finder, if you want to see it, now it has the directives folder and it set up our executions folder. Okay, cool. Now, what do we have to do? Okay, great. So, one thing I want to show you is the reason we're building this automation is because I actually use it all the time. And this is the Nan automation that we're going to be rebuilding. Normally, to figure this out, it took quite a few hours. But right now, I want to show you how all this can be done using Cloud Code without understanding all the little things that's been going on here. All right. So, we also have a test on the back with Nan if we want to verify if anything's going wrong. But we don't need any of that right now. We just need to tell it what we want. So, this automation, what it's doing is literally scraping creators that I like, and for each of those creators, it's, you know, generating their content. It's it's scraping their content and then using that to create a script for myself that I can record and publish that content. All right. I'll show you some stuff. I actually have a video of me building this automation, and this is something we're just going to be updating. Okay, so to do that, what am I going to do? Let me just start speaking to Claude again. Let me keep this simple. So, I want to scrape the latest Instagram posts from a couple creators in the AI space. For each post, I want to extract the transcript from what they're talking about, store that, and extract any key insights. Then, I want to summarize it and maintain that theme across all the creators. I want to make sure for each video that they have posted, I want to rewrite a script with my own prompt B and using theirs as an input so I can create content in my voice on autopilot. So, [snorts] it'll be a step-by-step framework where every day, let's say at 8:00 a.m., you're going to scrape their latest posts. And then after you scrape their latest posts, you will transcribe all of that. You'll store all of that. And if we haven't processed any of those new transcripts, then we are going to extract what they're talking about and then create a new reel, like a script for a new Instagram reel that I can read off and publish. And the creators I want to scrape are, let's say, Nick Serev, Matt Farmer, Nathan Hods Gason, Dr. Alvaro Centas, and I'll paste their Instagrams for you. And let me just go find their Instagram. I'll paste it here. Cool. And this can be anyone you like. And then I'm just going to hit enter. Okay, cool. I'm going to hit ask. And then, uh, let me go back. Okay, so now you see what's happening. It's asking for my scrape creators API key. Uh, and I'm going to go find it and paste it now. That's it. That's my contribution. I told it what I want. Now watch what happens. And you see here that it's doing all this crazy stuff that it's breaking down the problem. It's not jumping straight into the code before I approved it. Right here, it was figuring out what tools do I need? What order should I do things? What might go wrong? And it's the orchestration layer in action. This used to be my job, but now Claude's doing it. And did you see how it was deciding to use the scrape creator tool? I didn't have to mention that. It went to the internet and found it. And now it's writing the actual Python scripts. This is the execution layer. I'm somewhat of a Python expert, but right now I don't need to be. Like I'm no crazy expert, but I don't need to be. I just need to know what I want. You see how it's handling specific things that would have taken me hours to figure out. It's literally running it. You can see the to-do list. It's it's going to be running it. It's going to be testing it. And if something breaks, it'll read the error and then fix that, too. So, we can literally see the updated to-dos on what it needs to do. That's the really cool part. That's the error handling. Okay. You see how the database, it set up the entire database with all the creators, and it's testing the scrape creators API now without me having to ask it to do anything. And now the API is working and it's fixing any errors that it comes into. Normally I would have to do that, read the documentation, something doesn't work, change it around, all that nonsense. I'm just chilling. All right. So, it's saying it's done. Let's check it out. I just need to hit this. Okay. That's actually really cool. Let me show you guys. Okay. So, you see how there is literally like a Kanban board. It made that. This is so cool. Creators for the content script. You can refresh it and then make it, uh, let's see, next rev. This is off this source. If you're still paying for API keys, you're wasting your time and money. It's pretty cool. Generate script. Good news, you can access an AI model for free, etc. And then I guess I I'm able to change a lot of the stuff on how, let's see, let's review. Oh, we can improve. Move. Okay, this is really cool. Now, if, for example, let's say I didn't like this output, right? I can just be like, "Hey, what's um, what system prompt are you using to generate the generated script with?" Oops.
[snorts] So, you notice how now I am the director. I'm like talking about it, and anything that it needed help with, it fixed it itself until it got done. And that's what I really like about this. It made the summaries for each of the creators, the key themes, all of them. This would have taken me, what, like an hour or two more, maybe more. And then now I can change it. "Can you write a much better script? Really think about it. And the model I want you to use, maybe use Sonnet 4.5 because I think that's the best. Um, put a lot of emphasis on the hook and then fix that." Cool. All right. Uh, let it do its work. And the real cool part is like, remember this like visual spaghetti that we had? This is the same task that it's doing. It's just a completely different experience in building it. One requires you to be the orchestrator. Like I need to know how all this stuff works. And now I have a completely different experience. Like this. Look at this. This is so cool. This took me a long time to build. Like all this automation with the Airtable where I can see the content pipeline like this. Now, we just did it in a couple minutes. Of course, you can polish it, but the polish is just me talking to the AI a bit more and improving the system prompt. So, that's the demo. That's what building with Cloud Code actually looks like. You might be wondering, when should I use this? And is this the right tool for everything? Let's talk about that.
So, I want to show you this. So you see here it's running into errors, and you see what happens when, like, Cloud Code hits an error in your automation, that's not a bug, that's a feature. It corrects itself. It's one of the superpowers of this approach. So let me break down three things that make agentic workflows a completely different ballgame. And the superpower one, I will say, is the self-correction it does. So when a traditional automation breaks, like this Nan one I was showing you, what happens? Maybe you get an error. Maybe you get an email notification if you set that up. And then you have to go in, you have to figure out what went wrong and then fix it. But with an agentic workflow, the agent reads that error message. It understands the problem and rewrites its own code to fix it. Like literally, like this, it's rewriting its own code and fixing it. Now, think about what that means. You're not a debugger anymore. You're not getting paged or tearing your hair out because some API format changes. The agent handles it. And if it can't handle it, then it tells you, but 80% of the time, it fixes itself and moves on. That's not automation. That's resilience.
And the second superpower I love about Cloud Code is its parallel processing. And here's something that would take you hours to set up in Nan that we just did right now in a few minutes. With Cloud Code, you can say, "Make it faster." But the cool part is that you're not limited to just building one automation at a time. You can just hit "New Session" and then do it over and over again. And Cloud Code lets you do that all sequentially in parallel. And that's it. That's the instruction. And Claude is going to rewrite the code. And it maintains its own context space, one at a time. In a no-code tool, you'd be watching YouTube tutorials for an hour trying to figure out how to set up parallel branches and all that stuff. Here, you just describe what you want, and the agent figures out the implementation for many of your workflows.
And the third superpower I would say is learning from documentation. So, Claude actually documents to itself if it's failed and tried something. So, it doesn't have to do it all over again. And the cool part is you can give it MCP access, like context 7. And rather than me having to go to the APIs and understand how this works, you can give it some API documentation, and Claude learns on its own and does all that trial and error. So you don't have to think about that for a second. You don't need to read through pages of docs anymore. You don't need to figure out authentication endpoints, rate limits, all this random garbage. You just say, "Hey, here are the API docs. Use this to get the data I need," and the agent reads the documentation, understands it, and writes the integrations. I literally used to spend afternoons figuring out how to connect new APIs. Now, it takes literally five minutes. And that's the 10x difference.
So, let's add all this up. There's self-correction, that means you're not the debugger anymore. There's parallel processing, that means you're not the performance optimizer, and you make many workflows at the same time. And then you get to learn from docs, which means you're not having to read all this stuff and waste all your time. You're just the person that knows what needs to be done. And that's not a 2x improvement. That's a 10x shift in how you work.
So now you've seen the problem with traditional automations, the framework we use with Cloud Code, the demo, all the superpowers. The question is though, what are you going to build? If you understand all of this stuff, you're technically can build whatever you want. And by the way, if you want to learn more on how to use Cloud Code way better, click the video right here. I'm going to break down exactly how you can learn to use Cloud Code better than 99% of people, no matter if you're using it to make your own programs like this or if you're coding real apps with the intention to sell them. And if you've gotten this far, leave a comment below on what kind of automation you'd like to see, so I can even build it.