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
Include only what applies to EVERY task. Agents may ignore content deemed irrelevant to current task.
4. Agents Are In-Context Learners
Don't document what the agent can infer from code. Style, patterns, and conventions are learned from existing files.
5. Progressive Disclosure
Point to files instead of embedding everything. Let agents discover details when needed.
Structure Template
# AGENTS.md
## What This Is
[1-2 sentences: project purpose and type]
## Tech Stack
- **Runtime:** [e.g., Bun, Node, Python]
- **Framework:** [e.g., Astro, Next.js, Django]
- **Language:** [e.g., TypeScript strict mode]
## Quick Commands[dev command] # [description] [build command] # [description] [test command] # [description] [lint command] # [description]
## Project Map
[directory tree with brief annotations]
## Key Conventions
- [Only non-obvious, universally-applicable conventions]
- [Things the agent cannot infer from code]
## Verification
[How to verify changes are correct - primary feedback mechanism]
What To Include
| Include | Why |
|---|---|
| Project purpose (WHY) | Agent needs to understand intent |
| Tech stack (WHAT) | Core tools and runtime |
| Essential commands (HOW) | Build, test, dev, lint |
| Project structure map | Navigation aid |
| Verification method | How to confirm success |
| Non-obvious conventions | Only what can't be inferred |
What To Exclude
| Exclude | Why |
|---|---|
| Code style guidelines | Agent learns from existing code |
| Import patterns | Inferred from codebase |
| Formatting rules | Use linters, not LLMs |
| Detailed error handling | Context-dependent |
| Dependency descriptions | Agent reads package.json |
| Example code snippets | Become stale quickly |
Workflow
Creating New AGENTS.md
- Read
package.json, config files, and 2-3 representative source files - Identify: runtime, framework, language, key commands
- Map project structure (top-level directories only)
- Write draft using template above
- Review: Is every line universally applicable?
- Trim to essentials (<60 lines ideal)
Improving Existing AGENTS.md
- Read current AGENTS.md
- Check line count - if >100 lines, needs trimming
- Identify:
- Style guidelines (REMOVE - agent learns from code) - Linter-like rules (REMOVE - use actual linters) - Task-specific instructions (MOVE to separate docs) - Stale/outdated info (REMOVE)
- Verify remaining content is universally applicable
- Ensure WHAT/WHY/HOW structure is clear
Progressive Disclosure Pattern
For larger projects, keep AGENTS.md lean and point to detailed docs:
## Documentation
For detailed information, read these files when relevant:
- `docs/architecture.md` - System design and data flow
- `docs/testing.md` - Test patterns and fixtures
- `docs/deployment.md` - CI/CD and release processVerification Checklist
Before finalizing, verify:
- Under 60 lines (strongly preferred) or under 100 lines (acceptable)
- Every line applies to ALL tasks
- No code style rules (agent learns from code)
- No formatting rules (use linters)
- Commands are correct and runnable
- Project map reflects current structure
- Clear verification method specified