Most code an AI agent writes looks right. It compiles, it reads well, it probably works. "Probably" is the problem. The agent stops when the work looks done, and "looks done" is rarely "proven done." You find the gap later, in production, in the case nobody tested.
proven-python is a Claude Code skill that holds an AI coding agent to the way a disciplined engineer actually works on Python:
It is model-invoked: the agent loads it when it is about to write, change, debug, test, or review Python, and follows the procedure instead of producing plausible-looking code nobody ran.
A discipline skill that fights your agent is worse than no skill. proven-python has an "apply with judgment" rule: scale the rigor to the code (a throwaway script earns the spirit, not the full gate), defer to your instructions and the codebase's conventions, and skip the ceremony on a one-line diff. It is built to align with Anthropic's own Claude Code best practices, not override them.
/plugin marketplace add shanwije/proven-python
/plugin install proven-python
Or copy skills/proven-python into your project's .claude/skills.
MIT licensed, principles distilled from the canonical Python and engineering literature with every source and its license listed in the repo. Feedback and PRs welcome, especially on where it is too strict or gets in the way.