AGENTS.md Guidelines
Core Principles
- •Keep root AGENTS.md as small as possible; it loads on every request.
- •Use progressive disclosure: link to deeper docs for details.
- •Prefer stable concepts over volatile paths or time-sensitive notes.
- •Avoid “ball of mud” growth; don’t add rules for every mishap.
- •Keep instructions concise; assume the agent already knows basics.
Essentials Only
Root AGENTS.md should usually include:
- •One-sentence project description
- •Package manager if non-default
- •Non-standard build/typecheck commands
- •Pointers to deeper docs
Progressive Disclosure
- •Move language, testing, and workflow rules into separate docs.
- •Link to those docs from AGENTS.md.
- •Keep guidance domain-based, not file-path based.
- •Avoid deep reference chains; link one level deep when possible.
Staleness Rules
- •Avoid exact file paths unless they are stable interfaces.
- •Avoid dates or time-based instructions.
- •Prefer capability descriptions over structure hints.
Monorepo Guidance
- •Use root AGENTS.md for shared rules.
- •Use package-level AGENTS.md for package-specific conventions.
- •Keep each level focused and minimal.
Instruction Design
- •Start with a TL;DR checklist for fast scanning.
- •Include clear triggers and skip cases to prevent overuse.
- •Provide ordered steps with explicit outputs.
- •Use strict format rules with correct/incorrect examples.
- •Add a decision tree for edge cases when needed.
- •Include validation and debugging tips to reduce failure loops.
- •Make naming conventions explicit.
- •Define archive or completion criteria.
- •Keep terminology consistent.
Quality Checklist
- •Description includes what + when, third-person.
- •Root AGENTS.md is short and link-heavy.
- •TL;DR + triggers + steps are present when needed.
- •No duplicated guidance across files.
- •Terminology is consistent.
- •Examples are concrete and brief.
- •Instructions fit an "instruction budget".