The AI Stack I Actually Use to Build

Not the tools I’ve heard about — the ones that survived on my machine. A working founder’s honest map of which AI does which job, and the input method almost nobody talks about.

There are a thousand “best AI tools” listicles, and most are written by people who tried each tool for an afternoon. This isn’t that. This is the short list of what actually survived on my machine after a year of real building — and, more usefully, which tool I reach for which job.

Because that’s the real skill now. It isn’t finding the one magic model; it’s knowing that different tools have different temperaments, and matching each to the work it’s best at.

Claude Code for the whole picture

For anything that has to hold a lot in its head at once — architecting a feature, reasoning through a messy codebase, or writing prose — I use Claude Code, currently on Opus 4.8, and Fable 5 when I want the frontier model. It’s what I trust for the robust, end-to-end approach rather than a single throwaway snippet.

The reason is harder to benchmark than it is to feel: it simply syncs with how we think. It holds context across a whole project instead of losing the thread, and the collaboration just vibes right. That fit matters more than any leaderboard, because a tool you argue with less is a tool you’ll actually keep using.

ChatGPT for the granular work

For the smaller, well-defined jobs, I reach for ChatGPT — writing individual unit tests, knocking out a specific function, quick code chores with clear inputs and outputs. On the bounded stuff, it’s fast and reliable.

This is the whole philosophy in miniature: horses for courses. One model for the sprawling, judgment-heavy work; another for the tight, defined tasks. You don’t need the “best” AI. You need the right one for the shape of the problem in front of you.

Voice is the most underrated input there is

The biggest productivity jump of my year wasn’t a smarter model. It was changing how I talk to them. We use Whisper Flow to drive nearly everything by voice — and, quite literally, this piece is being spoken, not typed. Everyone obsesses over multimodal output; the genuinely underrated frontier is multimodal input.

Speaking is several times faster than typing, and more importantly it lowers the friction of thinking out loud until ideas flow the way they do in real conversation. If you take one thing from my stack, don’t copy the models — copy this. Add voice, and you change not just how fast you work, but how readily you’re willing to begin at all.