Agentic development harness
The tooling and conventions I use to run Claude Code on real projects: generated context maps, session preservation, and measuring whether the agent actually followed the rules.
- Claude Code
- Python
- hooks
Most of my recent projects are built with an AI agent doing much of the typing. The interesting engineering problem turned out to be everything around the code: what the agent knows at the start of a session, and how you'd know if it ignored your standards.
The pieces:
- Context maps: a generator walks the code, pulls out annotated headers and public signatures, and writes a versioned map per module. The agent reads the map instead of rediscovering the repo every session.
- Session preservation: session logs get mirrored into the repo and turned into readable transcripts and summaries, so decisions made in conversation survive.
- Rule adherence and cost: scripts measure whether documented process rules were actually followed in each session, keep a ledger of deviations for a human to judge, and track cost and output per session.
- Conventions:
CLAUDE.mdinstruction files, ADRs, and a memory-versus-source rule: anything another session must inherit lives in the repo, not in the agent's memory.
This site's own CLAUDE.md is a lighter version of the same setup.