Notes

Claude Design makes prettier screens than Claude Code does. It isn't a smarter model, it's a different set of conditions.

Every part can be correct and the screen still be wrong. Nothing automatic catches that, so I write the rules down.

An afternoon with an LLM, one wrong theory, and a resource leak in an Apple daemon.

Write it down at the beginning, before anything can drift.

Building a skeuomorphic interface in pure CSS, no reference image, just me telling an AI what I wanted until it was right.

Why I built a curated panel of experts to answer the health questions I wouldn't hand to a generic model.

Keeping the context low and the cache high.

Making the file layout the flow, so nobody asks how the screens connect.

What Fincher's Mindhunter gets right about user research.