Garry Tan runs Y Combinator — the startup program that funded Airbnb, Dropbox, and Stripe. This week he published the exact AI setup he uses to build software, and it's worth paying attention to.
The idea is simple but clever. Instead of asking one AI to do everything in a generic way, you give it different modes — like handing someone a different hat depending on the job. There's a mode that thinks like a founder (opinionated, fast, focused on what matters), one that thinks like a senior engineer (cautious, thorough), one that acts as a paranoid reviewer looking for problems, and one that just ships the thing.
The result feels less like talking to a chatbot and more like having a small team where each person has a clear role and knows when to stay in their lane.
There's also a quiet technical detail that matters even if you'll never touch it yourself: the setup keeps a browser running in the background permanently, so any task that involves checking a website happens almost instantly instead of waiting several seconds each time. Small thing. Adds up fast when you're doing a lot.
For a solo founder or a small team trying to move quickly, the real insight here is that how you instruct an AI matters as much as which AI you choose.
Claude Code — a version of Anthropic's Claude AI that can write, edit, and run software on your computer directly, not just through a chat window.
Slash command — a shortcut you type (like /review) to tell the AI which mode or role to step into.
Headless browser — a web browser running invisibly in the background, with no screen, used by software to visit websites and interact with them automatically.
Y Combinator — the world's most well-known startup accelerator, based in San Francisco.
If you work with a developer or an AI consultant, ask them: are we giving our AI clear roles, or are we just asking it to figure it out each time? The answer will tell you a lot.