AI agents do not feel the cost of bad code. We do.

When a human maintains a growing codebase, the pain compounds. Every shortcut, every duplicate function, every defensive check that can never fire makes the next change a little harder.

At some point the pain gets big enough that we are forced to stop and refactor. That feedback loop is annoying, but it is also what keeps a codebase alive.

An agent never feels any of that. It solves the ticket in front of it and moves on. So it extends bad structure without noticing, adds code without ever deleting, and guards every edge case it can imagine because "comprehensive" looks safer than "minimal."

None of that breaks a test, so none of it gets flagged. The codebase quietly fills up with noise, and the person who inherits it is the one who pays.

So I wrote a small CLAUDE.md with one job: write as if a human will maintain this, because a human will. Five principles:

1. This is not your last task. Fix structure, do not stack patches.
2. Subtraction is a feature. Deleting code is real work.
3. Say no to abstraction. Rule of three. Nothing nobody asked for.
4. Completeness is not thoroughness. Guard real boundaries, not imagined ones.
5. Match, do not improve. Every changed line traces to the request.

The agent is good at producing code. Our job is making sure it respects the next person who reads it.

Repo is open. Built on Sonar's AC/DC cycle and Karpathy's CLAUDE.md.