AGENTS.md Skill

Create effective AGENTS.md files that onboard AI agents to codebases.

Core Principles

1. AGENTS.md Is High-Leverage

AGENTS.md goes into every conversation with the agent. A bad line affects every task. Craft each line carefully.

2. Less Is More

• Frontier models follow ~150-200 instructions reliably
• Agent harnesses consume ~50 instructions in system prompts
• More instructions = uniform degradation across ALL instructions
• Target: <60 lines, <300 lines maximum

3. Universal Applicability Only