Working with coding agents: notes from the passenger seat

(Sample post — placeholder content for the new posts collection.)

Most of the code on this site was written in collaboration with coding agents, and the workflow has settled into a rhythm worth writing down.

What the agent is good at: the first draft. Given a clear brief — “a CRT terminal that boots the site, with a working fake shell” — it will produce something runnable in one pass, including details I wouldn’t have bothered with (reduced-motion handling, mobile keyboard focus).

What still needs a human: taste and constraints. The agent will happily cover an art composition with a wall of cards unless you tell it the composition is the product. The most valuable prompt engineering turns out to be ordinary product judgment, written clearly.

The surprising part: documentation pays off immediately. Every gotcha recorded in the repo’s notes (like the scoped-CSS trap) is a mistake the next agent session doesn’t make. The repo’s memory becomes a shared brain between sessions — which is a strange and useful new reason to keep docs current.