I Hacked Claude Desktop to Build an AI YouTube Assistant (No Code)

I got the Claude Pro Max 20, the best one. And I want to talk about how I ran a bunch of different Cloud Code instances, but also just leveraging the absolute crap out of the desktop app. So, let's cover my first impressions, things that I've learned that I feel like no one knows about if you're on the fence about it. So, first of all, I found that the desktop app is in beta, so it's a little bit finicky. It does have this. So, it's like raycast where it's I bound this to option X and then I can pop that open. But the most interesting thing I've come across is that you can actually hack together these agentic workflows and pretend it has tools by just making really good context in the prompt like in the system instructions. So what I did was I wanted to have this YouTube assistant that can help with this. So when I make a video, I just made a video about the remote agents in Augment and I wanted to have a way to make it easier. So, I'm not a script guy, but I wanted an outline so I could at least read it ahead of time and see how it could be better. And it did a really good job there. So, what did I do? I wanted to have something I could talk to where it'd be conditional based on where I'm at in the video making process. So, you either have the idea, but you don't know how to package it. So, you either have the idea, but you don't know how to package it, meaning what's the title and what's the things that you're going to talk about. That's the earliest version. or you have the video you've already made, which is what I usually do. I just flip the camera on and I yap and then I want to pass in a transcript so I can figure out how to position it. And so I found that first of all, you want to actually do this ahead of time because then you can bake in stuff for hooks, right? The thing that's going to get the person to watch. But what I did was I ended up making this project and I made the project such that I could have project knowledge. And in the project knowledge, this is where I actually bake in a bunch of the tools. So it's funny because on one side of the coin, all the models are saying you want to have less system prompt and more context. It's context engineering is the thing that they've coined it as the people from Devon. And I think that's true, but with this tool thing, I don't know how it works. So what I did was I actually opened a new chat with Claude Opus and I explained what I was trying to do. I explained the user story from my perspective of I am a creator. This is my context. This is what I'm trying to do. And then I listed out the steps of what I want to do. So I either have a video idea that I need to be turned into some packaging. So I'll explain the idea using voice with whisper flow and then it'll turn it into 10 clickbaity titles that are not like overbearing with like urgent, you're going to die if you don't do this or something super cringe. And then the outline. So those two will come back from the idea or I have that script right like I said. So I provided that context in a different chat and then I said give me the system instructions on how I'd make an assistant. This is what it gave me. So it basically said here's your role. Here's the background of the creator. So it's specific to me and then here's the tools you have access to. So tool one is the video outline and title generator. Pretty cool. Here's what is the expected output. So the compelling video outline within this it has all the different parts of a good hook. So this is from Alex Ramosy's framework of proof, promise and plan. Proof, why are you the person to listen to? Promise. What is it that I'm going to learn in this video by spending my time with this person? And then plan, how are we going to get there? So that's nice that it's baked in. And then the other output is those titles. The second tool is the title generator only. So that's if I have already made the thing, I pass in the script and then the delivery as well as another tool which is the YouTube description. So then I can pass in some examples, how to determine which tool to use based on some logic and then some best practices around all the things. Then in that initial chat when I was like how am I going to make this? How can I improve this? It's yo, you should definitely have some examples. So, I just threw in a bunch of examples for titles, for descriptions. And then I said, "Is there any way I can make this even better?" And then it says, "Write a user manual." So, I thought this was mainly for me, so I could know. But it turns out that the assistant really likes that, too. So, I found that to be really awesome. What this lets me do is instead of going and I actually did try this in Open AI and it didn't work. I tried to use their assistant. Nope. And then I tried to make a custom GPT. Double nope. Even with the actions built in, still not working the way I wanted it to. I think that Claude is just superior at writing. It's also superior at coding. I still think that's why I like Augment is cuz they have deep partnership there. But if I go back into here, we can see this is the overall stuff that I was doing with it. I think probably covered most of this. Next, we want to cover some of the questions that came in. I think I answered all of them actually. Cut. Then when it comes to actually using Claude Code, I think it's really good if you're comfortable in a CLI. I'm still in my honeymoon phase with it. I started using it, I don't know, 11:00 a.m. or 2:00 p.m. yesterday and then I was up super late using it cuz it's just if you know your code base, like the back of your hand, you know exactly how to implement it, then you're doing a lot of work for the thing. Like I could literally oneshot a blog, right? That should be pretty simple. But there's a lot of stuff around it that people don't think about it where it's okay, we want to have for/blog/category. Inside of a dynamic route inside of that dynamic category route, we want to have a dynamic page.tsx, which is the dynamic slug. We need to have a blog.ts that has all the categories stubbed out such that if we add a new one, then we can surface that in the dynamic page. The category page needs its own gallery and these tags. And then on the page of the individual blog post, we need to have estimated reading time. We need to have MDX with front matter of all the typical blog things like who published it at what time and what's the topic. And then we also need to have MDX support such that we can have custom components that are passed in so that we can have rounded images or a nice YouTube embed or custom links where when you scroll over them, it shows you a preview of the page before you went there. It's pre-fetching. And then all of it needs to be SEO optimized both on page and off- page. Are you using JSON LD? Are you putting in the right metadata and then you need to make sure everything is server side rendered so that you're just not missing the boat on all those benefits. So you put that in and it just did it, which is crazy. So those are my first impressions on it. I hope you guys learned something. If you did learn something, definitely subscribe to this channel. You'll see new videos, you'll learn more stuff. If you're interested in joining VAI with engineers from fast growing startups and Microsoft and Google, go check that out. The link is in the description. And I'll see you in the next one.