Major projects are embedding rules for AI coding agents directly in their repos. Here's what $MSFT, OpenAI, Vercel, PyTorch & others are telling AI to never do:

$MSFT TypeScript: Don't delete checks to silence errors. Existing checks aren't too strict.

Astral uv: Never assume a failing test was broken before your change. It almost never was.

OpenAI Python SDK: Don't shrink tests or raise limits just to pass. Fix the real issue.

$MSFT Playwright: Never push without permission. Never rewrite old commits—add new ones.

Vercel Next.js: Never rerun tests with different filters. Save output once, read from file.

Bun: Never write a test that only checks "it didn't crash." That test can never fail.

PyTorch: Never post on GitHub autonomously. Never pass AI output as human-written.

Apache Airflow: If a required tool fails, stop and report exactly which call failed. No workarounds.

Cloudflare: Comments explain why, not what. Don't repeat what the code already says.

OpenAI Agents SDK: Only stop for missing decisions, permissions, or real blockers. Say exactly what's missing.

These are the rules worth stealing for your own AI agent workflows.