AGENTS.md Management

When to Use

• "Create AGENTS.md for this repo."
• "Update agent documentation for this monorepo."
• "Set up hierarchical AGENTS.md files by package."

When Not to Use

• Pure code implementation work with no documentation update.
• One-off prompts where repository guidance files are unnecessary.

Principles

• Keep instructions concise, concrete, and path-specific.
• Prefer references and indices over duplicated prose.
• Optimize for nearest-file relevance in hierarchical layouts.

Workflow

1. Discover repository shape and technologies. Output: simple vs hierarchical documentation strategy.
2. Select structure (single root file or root + sub-files). Output: file layout plan.
3. Generate AGENTS.md content with concrete commands and paths. Output: actionable docs with JIT indexing.
4. Validate command correctness and duplication boundaries. Output: clean, copy-paste-safe instruction files.
5. Re-check after major repo changes. Output: synchronized documentation.

Structure Decision

• Simple project: one AGENTS.md when stack and patterns are uniform.
• Complex monorepo: root AGENTS.md + scoped subdirectory AGENTS.md files.

Quick Commands

# Discovery baseline
rg --files

# Detect core config/tooling
rg -n "workspaces|nx|turbo|pnpm|yarn|packageManager|tsconfig|pytest|playwright" .

# Locate existing AGENTS files
find . -name AGENTS.md -o -name AI-DOCS.md

Anti-Patterns

NEVER assume a technology stack without discovery

WHY: incorrect assumptions produce unusable instructions. BAD: generate React/Jest guidance without evidence. GOOD: run discovery commands and map docs to detected stack.

NEVER dump encyclopedic content into root AGENTS.md

WHY: oversized docs increase token cost and reduce usability. BAD: embed full framework manuals. GOOD: keep root concise and link to scoped files/references.

NEVER duplicate the same instructions across root and sub-files

WHY: duplication creates drift and maintenance overhead. BAD: copy/paste identical conventions in every file. GOOD: keep universal rules at root and package-specific rules locally.

NEVER provide unverified commands

WHY: broken commands erode trust and block contributors. BAD: include hypothetical commands. GOOD: include only validated copy-paste commands.

Quick Reference

Verification

bunx markdownlint-cli2 "**/AGENTS.md" "skills/agents-md/**/*.md"
sh skills/skill-quality-auditor/scripts/evaluate.sh agents-md --json