My AI Coding Workflow — Everything I Use Now (for peak performance)

So, I'm currently building a tool that lets AI agents take useful content that I find online and turn them into clean markdown they can actually use for their work. This could be a YouTube video, an Instagram Reel, an X article, or just an interesting blog post. I want to save everything into one place and make it available as context for my AI agents later. But, I'm currently facing a problem. The more I use my own tool, the more cluttered my sources have become. So, in this video, I'll be walking you through the new features I'm building to solve this problem. I'll also share a free tool I've discovered that allows AI agents to work with your databases in a more safe and controlled way. And I'll also update you on my current AI coding stack because the tools I'm using have changed quite a lot in the past couple months. So, to give some background, the app I'm currently building is called Create Skills. It's a browser extension that lets me save pretty much anything I find online into a clean markdown file. And then, I can feed that into my AI agents as context for its work. And recently, I started building out a custom MCP server that lets AI agents like Claude Code and Code X have direct access to the sources I've saved, which saves me time because it means I no longer have to context switch and copy and paste anything into my prompts. But, the more sources I've saved, the messier my library became. I had a bunch of genuinely useful resources in my library, but they were all just sitting in one unorganized list. So, that leads me to the two new features I'll be building for my MCP server. First is source organization. I want the agent to look at all of my saved sources and organize them by creating folders and grouping related sources together. And second is source retrieval. So, instead of manually scrolling through my feed to find something, I can just ask the agent something like, "Find the best sources for growing on social media and use it to create a new agent skill. " And then, I [clears throat] should get a small set of the most relevant sources back ranked by usefulness. And if I can get this working, my app stops being just a standard kept tool. It becomes a fully customized context system for all of my AI agents to use for their work. But, to build this out properly, I would need to test it out with my current data, which sits in my production database. And obviously, experimenting with production data for a new feature is not the

smartest idea. So, for this build, I'm going to use a free tool called Ghost, which essentially lets my AI agents create, fork, and discard their own databases when they're working. And I know I was just as skeptical as you are right now after hearing that. As the database is one of the most important parts of any app, handing that off completely to AI might not feel right. But, after trying it out and learning more about its capabilities, I was pleasantly surprised in how effective it was and how it has allowed me to build and test new features safely without running the risk of nuking or corrupting my data. It works by letting my agents spin up Postgres databases all by themselves that it can query and make changes to. Then, it will automatically perform any cleanup after it's finished, all without needing me to manually configure or manage any databases or having to worry about it modifying any real data in production. If you're curious about trying it out, here's how to get set up. So, the first thing you want to do is head over to ghost. build and then head over to the docs and the installation tab. As you can see, these are all the commands you can run to install the Ghost CLI. But, in this example, I'm going to use the Homebrew command and copy and paste that into my terminal. So, open up your terminal and then paste in the command and press enter. And once you let it run, it should successfully install everything that it needs. And once that's done, you just need to run the login command, which is ghost login. Press enter. And since I'm already logged in, it shows me this success screen. But, for you, you will just need to log in with your GitHub account and then authorize the app. And then once you've logged in, you just need to run ghost MCP install. And running this will install Ghost's MCP server locally to all of your coding agents. So, here, just select all of the clients you want to install the MCP server for and then press enter. And since I'm using Code X, I just want to check that the MCP server has been added. And as you can see, it's right there available and turned on. So, that's all there is to it. And now, you can just start using the Ghost MCP server in all of your chats. So, after setting things up, your coding agent can now directly use Ghost through their MCP server. So, instead of me having to learn a whole new app and UI, the agent can do all the tedious work of managing the databases for me as it reasons about the current task and what it needs to do. And I have to say, coding agents are now becoming way more capable than just generating code. If you give them the right tools and data, they can actually handle the product building and testing loop all by themselves. Which just means less babysitting for you as the builder and more time to focus on making important product decisions and marketing. So, I ended up shipping two core features for my custom MCP server. The first one being source organization where the agent can look at my safe sources and automatically sort them into folders based on their content and similarity. And the second feature is source retrieval where the agent can simply find the best sources for the current job. And behind the scenes, it gives each source a score and a rank for how relevant it is to the current task. And then for testing in this case, I made sure to call the Ghost MCP server to create and clone temporary databases so that I can test the features with real data and get feedback on what's working and what's not. And the nice thing about using Ghost in this case is that I don't have to be hands-on in the testing loop at all. I can just let the agent do its thing and it will iterate on itself until it fixes all of the issues that come up since it can manage the database life cycle all by itself. If I wanted to, I could spin up 10, 20, or even 50 databases running all at once and then test different approaches and keep the one that works best. So, if you're currently building agent-first applications like custom MCP servers or CLIs, I highly recommend integrating Ghost into your database layer to test your features with real data in a safe and controlled way. It's free to get started and they offer a very generous free tier that includes 100 hours of compute and 1 terabyte of storage. And with hard spending caps on by default, you can be sure you'll never get a surprise bill for overspending. There's currently no waitlist and you can get started by running just one command in your terminal. Check the first link in the description if you want to start building with Ghost. So, a couple months

ago, I probably would have built these new features in a very different way with a very set of different tools. So, now I want to share with you my updated AI and coding tech stack and show you the upgrades I've made to help me build and ship my apps faster. So, let's start with the coding agent. I recently made the full switch from using Claude Code exclusively in the terminal to now purely using the Code X desktop app. There were a few reasons for this, but the main one was that in my opinion, the Opus 4. 7 model is just dumb compared to even 4. 6 or even 4. 5. It was really noticeable in the way it would over-engineer simple features or poorly manage its own context window, which then is a vicious cycle because that tends to lead to worse performance. And not to mention the whole usage limits restrictions they tried to pull on users behind the scenes recently. But anyway, these days I'm doing most of my work directly in the standard Code X desktop app, which is different from what I'm used to, but surprisingly, it has everything I need to debug and test any new features or apps I want to build. You can open up a terminal, browse your project files, review diffs, or even watch the agent run and test your app inside the built-in browser. And generally, I found that GPT 5. 5 is much more thorough and reliable than the current Opus models. For example, when I'm in plan mode, it's not afraid to go through multiple rounds of asking clarifying questions and then use a variety of tools without me asking and spin up multiple sub-agents for research before it generates a plan. And usually, the plan it gives me is solid and I'm confident that it will implement what I asked for. Outside of coding agents, some of my work still happens in Ghost E, which is my terminal workspace. Here, I usually run my dev servers when testing my apps locally. And I've been using this TUI app called Lazy Git to manage all of my current changes in Git and another one called Yatzi to navigate and manage my project files. And sometimes, if I need to do some simple tasks or research, I run some additional sessions in Open Code, which is an open source coding agent that allows me to hook up my Code X subscription and is way less buggy than Claude Code in the terminal. And for my IDE, I'm still using Cursor for some light edits. And I've tried out their new Composer 2. 5 model, which is actually really good for coding and is super fast and accurate. And some agent skills I've been using a lot these days are copywriting to help me draft high conversion copy for my landing pages, Grow Me to help me turn fuzzy ideas into clear actionable plans. SEO audit to make sure my technical SEO is all configured correctly so that web pages can be crawled, indexed, and mentioned in AI suggestions, which will just bring me more organic traffic and users. And then finally, UI/UX Pro Max and front-end design to make my UI more intuitive to use and to design it in a way that doesn't look AI generated. So the stack right now is basically Codex for building and testing new major features, Ghost theme for my terminal, Cursor for hands-on and light edits, and custom skills to improve the agent's quality and performance. But I'm pretty sure it's going to change again soon when the next big thing comes around.

And I also recently released a new batch of resources for my members of my private community, Startup Club. It's a new collection of in-depth guides for beginners that shows you how to validate your app idea, build your app and landing pages with the latest AI tools, and develop a distribution and marketing strategy to get your first paying customers. Check the link in the description if you're interested, and I'll see you in the next one. Bye-bye.