Coding Agent Rules — Development skill for Claude Code

Ten Commandments for Coding Agents

**English** | [한국어](README.ko.md)


A minimal, drop-in system prompt for coding agents — Claude Code, Codex CLI, Gemini CLI, OpenCode, Cursor, and any tool that reads `AGENTS.md` or `CLAUDE.md`.

Two files. Ten rules. No fluff.

Why

Most agent prompts are bloated with role-play, persona scaffolding, and tool-specific tricks. This one is the opposite: a tight contract about how an agent should behave when it touches your code.

Every rule comes from a real failure mode:

• Agents that assume instead of asking.
• Agents that "improve" code you did not ask them to touch.
• Agents that pile every concern into one 2,000-line file.
• Agents that hide errors behind try/catch and call it a fix.
• Agents that claim success without running anything.
• Agents that hardcode secrets or run rm -rf to "clean up".
• Agents that bury the decision in narrative and never say why.
• Agents that read the whole codebase into the main context, then ask you to re-explain it.
• Agents that ship a patch that passes tests but fights the architecture.

The Ten Commandments — plus the short *Response & Documentation Style* section — block all of those.

Install

Pick the file your agent reads. Drop it in your home directory or project root. Done.

Tool	File
Claude Code	~/.claude/CLAUDE.md or /CLAUDE.md
OpenAI Codex CLI	~/.codex/AGENTS.md
Google Gemini CLI	~/.gemini/AGENTS.md
OpenCode	~/.config/opencode/AGENTS.md
Cursor / Windsurf / others	/AGENTS.md

One-liner

curl -fsSL https://raw.githubusercontent.com/cskwork/coding-agent-rules/main/AGENTS.md -o ~/.codex/AGENTS.md
curl -fsSL https://raw.githubusercontent.com/cskwork/coding-agent-rules/main/CLAUDE.md -o ~/.claude/CLAUDE.md

Single source of truth across CLIs

If you use multiple CLIs, keep one copy and symlink the rest:

git cl