If you use more than one AI coding tool, your repo probably has this problem: the same instructions copy-pasted into .clinerules, .cursorrules, AGENTS.md, CLAUDE.md and .github/copilot-instructions.md. A month later they disagree, and each tool behaves differently on the same codebase.

A developer on Hacker News described exactly this recently: switching harnesses when one hits its usage limit, and maintaining 20+ config files just to keep a project portable. Here is a low-tech way to keep it under control without any new tooling.

1. Pick one source of truth

Write the rules once in AGENTS.md (plain markdown). Everything else is generated or a pointer.

2. Keep rules short and testable

Rules that work across tools are concrete:

  • "Run pytest -q before declaring a task done."
  • "Never edit files under migrations/ that are already committed."
  • "Do not add a dependency without saying why in the PR description."

Vague rules ("write clean code") get ignored by every tool.

3. Sync with a 15-line script 

from pathlib import Path src = Path("AGENTS.md").read_text(encoding="utf-8") for target in [".clinerules", ".cursorrules", "CLAUDE.md"]: Path(target).write_text( "<!-- generated from AGENTS.md, do not edit -->\n" + src, encoding="utf-8", ) print("wrote", target)

Run it in a pre-commit hook so drift is impossible.

4. Split by repo type

A Python API, a Next.js app and a CLI need different rules (test command, lint command, forbidden directories). Start from a template per repo type instead of one giant generic file.

I put together six repo-type templates (each as .clinerules and .cursorrules, plus a short guide) as a pay-what-you-want pack, minimum $1: https://renevibe76.gumroad.com/l/toahua

The sync script above is free and works with any rules you write yourself.

What tool combination are you juggling, and which rules file do you trust most?