I build products with AI, and I write down how.

I've written software for years. AI didn't change what I know how to do. It changed how much of it I can get done, and it carried me through the parts I'm genuinely bad at, like design and marketing.

The interesting thing I found is that the model is rarely what decides whether a build works. What decides it is upstream: whether the idea survived a hard look, whether the model got enough context to do something specific, whether the work was cut into pieces small enough to review, and whether anyone checked the result before it went out.

So that's what I document. The Builder's Toolkit is the system I run: decide what to build, write the requirements down, set the agents up so they know the project, work the build and review loop in bounded tasks, then verify and ship.

I write it from builds that are in progress, which means the mistakes are in there too. A method that only shows the clean version isn't much use the first time something breaks.