My AI tool stack, and the 0.1% I still do myself.
It's October 2026, and my AI tool stack looks nothing like it did in May. Here's what I use now, what I use it for, and what I've learned.
What I use now
Three tools: Claude, Codex, and ChatGPT. That's it.
Antigravity is gone. It was such a good tool. Then they moved it to be all Gemini, and I couldn't figure out the new update. It didn't fit what I wanted anymore, so I left. Basically, they updated it.
I stopped using Perplexity too. I don't have a use for it anymore.
NotebookLM is still a very solid study tool, but it's not my main one now. Nothing really replaced it. I'm still looking for a good one.
How I use Claude Code now
Everything I possibly can goes into Claude Code. Even stuff that probably shouldn't.
I used to run it inside Antigravity. Now I use it in the Claude app, and very rarely in the terminal. I also stopped working on my own computer and started working in a GitHub repo. That's just the type of work I'm doing now.
Here's a habit I really like. For a new project, I make a new GitHub repo. Every time, I have Claude summarize our chat and put it in the repo. That way I can always go back and look if something breaks.
What each one is for
Claude is for creating things. It's also where I brainstorm and really think.
ChatGPT is more for research and getting stuff done.
Codex I use in the app, and most of the time it runs on my computer. Honestly, if I had to drop one tool, it would be Codex. The main reason I use it, and ChatGPT too, is to save my Claude limits.
Bouncing them off each other
I don't always switch when I hit a limit. If Claude is in the middle of something, I don't want to mess it up. But most of the time, I'm switching back and forth.
I'll ask ChatGPT, "What do you think of what Claude did?" And it'll say something like, "Oh my God, that's terrible. Here's how I could do it better." Then I go to Claude and say, "This is what ChatGPT said about your work. What do you think?" And we go back and forth like that.
It gives me a very good result.
The 0.1%
People ask if I just let the AI do everything. I'd say it does about 99.9% of the work. That's more of a feeling than a number.
The 0.1% I do is making sure the AI gets it right.
It doesn't get things wrong the way you'd think. It just misses stuff. It'll miss a major SEO thing. It'll forget to add security. It won't think ahead, and it won't always follow the proper guidelines. That's where I come in. That's the human part.
My job has changed. It used to be reading and writing code. Now it's more like reading and writing summaries. I notice the important things the AI might miss, and I help it fix them.
You don't need to know a lot of code for this. But you need a little bit. You need to know what code is supposed to do, and the proper way to do things.
How I prompt now
I still agree you should give Claude goals. But I do more feedback loops now, with a human in the loop. That's me. I'm okay going a little slower if I get a better result. It's still millions of times faster than going with a developer.
I'll also tell Claude things like, "You've got to make the Sonnet family really proud. You've got to be good. This is representing you."
A real example
When I build a website, I use a collection of skills. Skills are saved instructions I give Claude for one job, like checking a site for security problems. When the design is done, I run my security skills over it, and then my SEO skills. I think those are really crucial. I might share some of them later.
The most surprising thing an AI has done for me was on a big redesign. I used Claude Design, and when I saw the result, I was like, "Holy smokes, this looks like a real AI app." It looked professional. That was a huge moment for me.
I'm not afraid to say I use AI
I built this website with Claude Code and ChatGPT. I'm not afraid to say that. And I think some people are starting to get really afraid to say it. They're worried: "My boss could just do this with ChatGPT."
But I know what kind of security needs to be in there. I know the best AEO and GEO practices. That's answer engine and generative engine optimization, the new SEO. I take everything I know and I use Claude's knowledge too. But Claude can't understand my goals. So I use my knowledge with Claude, and we work together as a team. It's like having another employee.
If you want to start
If a kid my age wanted to start building with AI tomorrow, here's what I'd do.
Pick a couple of projects. The first one: make your own personal website, on your own computer. Later, if you want, you can spend $10 or $20 on a very cheap server and put it online.
I use Wix, which is on the more expensive end. It gives me more flexibility and more features. But I wouldn't really recommend it these days.
Ask Claude how you should prompt it. It'll tell you. Watch videos from someone good. A person I like is Nick Saraev. He's very, very smart.
Then learn by testing. If I prompt like this, this is the result I get. And when you have your beautiful website, ask: what did I miss? What security did it miss?
Why I use two, not one
I used to say you should stop using one AI for everything. I still partly stand by that. But I use two AIs now, Claude and OpenAI's ChatGPT and Codex, and I think two is a good amount.
I still wouldn't use just one. You're relying on a single company, and if another one comes out with a better product, you're stuck in that ecosystem. I'm flexible. Because I switch around a lot, I can go and test the Gemini models whenever I want.
This is where my stack is as of October 2026.
Drafted and created with the help of AI.