Notes on building real software with AI coding agents — mostly Claude Code — and the workflows and tooling that make a human-in-the-loop setup actually ship working code instead of plausible-looking code.

What this is about:

  • Agent workflows — how to scope work, plan before coding, and keep an agent on rails over a multi-step task. Strong success criteria over vague prompts.
  • Tooling — the glue that makes agents useful on a real codebase: persistent memory across sessions, project context, search, and clean hand-offs.
  • The human-in-the-loop — where judgment still has to live, what to review, and how to catch the confident-but-wrong answer before it lands.

Most of the projects on this site were built this way — the birdfeeder vision pipeline, the self-hosted web apps, the firmware. The interesting part isn’t that an agent wrote the code; it’s the discipline around it that makes the output trustworthy. Writeups and patterns land here.