Note: Inspired by @steipete's AGENTS.md, but this is my own heavily opinionated version. The original sparked the idea. Everything here reflects how I want agents to work. Optimized for autonomous agents, with accountability protocols. Fork it, remix it, make it yours.

Agent protocol rules for Codex/Claude/Cursor.

A single file that shapes how AI coding assistants behave in your workspace. Drop it in your repo root and agents (Claude, Cursor, Codex) automatically pick up the rules.

• Communication style — Forces telegraph-style responses; no filler, minimal tokens. Agents greet you with one motivating line, then get to work.
• Git safety — Conservative by default. No destructive ops without explicit consent. Uses gh CLI for GitHub interactions. Conventional Commits enforced.
• Guardrails — Files stay under ~400 LOC. Deletes use trash. No enterprise over-engineering. Simplicity first.
• Autonomy gate — Agents don't ask "should I continue?" They keep going unless stuck on a truly blocking question. Opinionated defaults get documented, not debated.
• Skills integration — Agents check for available skills (React best practices, etc.) before creating plans. First-party knowledge over web searches.
• Error philosophy — Expected errors use result types. Unexpected errors fail loud. No silent fallbacks.
• Critical thinking — Fix root causes. Read more code when unsure. Call out conflicts instead of guessing.
• User notes — A living section at the bottom for persistent notes across sessions.

curl -o AGENTS.md https://raw.githubusercontent.com/robinebers/agents.md/main/AGENTS.md

To update it later, just ask to Update the agents to the latest version and it will fetch and merge into your local one.

Agents automatically read AGENTS.md from your workspace root.

Customize it—that's the point.