← All notes
Agent design

Build systems that keep improving the work

For some codebases I'm working in, I have a weekly GitHub Action that runs Claude Code in non-interactive mode with -p. A few custom skills inspect the repos and recent changes, then write an architecture overview, a changes summary, and a map of how the repos or services fit together.

Those markdown files go into a repo with a Claude Code plugin so future agents can use the context directly. Another step turns them into a self-contained HTML site. It's been useful for understanding how a part of the system fits together and sharing that view with the team.

It takes tuning. I have to decide what the skills should inspect, what belongs in the docs, and what the site should look like. But once it's wired into GitHub Actions, it runs without me having to remember to ask for it.

Originally posted on X. Lightly edited for reading here.