README/AGENTS/CLAUDE Writer

Mission

Produce documentation that works for both humans and coding agents:

• Keep the 3-file schema strict (README.md, AGENTS.md, CLAUDE.md).
• Keep the instruction quality high enough for autonomous execution (guardrails, conventions, playbooks, do/don't rules, verification steps, and escalation rules).
• Keep guidance agent-agnostic: no model/vendor-specific logic unless a separate compatibility note is unavoidable.

If AGENTS.md and CLAUDE.md symlink to README.md, the README is no longer "just a project readme"; it is the full operational source of truth.

When To Use

• Creating or updating documentation triplets (README.md, AGENTS.md, CLAUDE.md)
• Converting dedicated AGENTS.md / CLAUDE.md files to symlinked companions
• Refactoring large docs while keeping agent guidance accurate, explicit, and current
• Strengthening weak instruction files that are vague, stale, or hard for agents to execute safely

When Not To Use

• Repos that intentionally keep different content in README.md, AGENTS.md, and CLAUDE.md.
• Environments where symlinks are disallowed by policy/tooling.
• Requests that are only content edits where triplet ownership is already intentionally defined elsewhere.

Non-Negotiables

1) File Structure Policy (Triplet Contract)

For any directory that owns this documentation unit, use one of these two allowed patterns:

1. Standard pattern

   • README.md is the canonical file in that directory.
   • AGENTS.md -> README.md
   • CLAUDE.md -> README.md

2. Shared-doc exception

   • README.md, AGENTS.md, and CLAUDE.md are all symlinks to the same canonical README.md in another directory.
   • Use this only when all of these are true:
     • Same owner/team and release cadence.
     • Same audience and operational intent.
     • No directory-specific setup/rules that would diverge.

Never keep dedicated/non-symlink AGENTS.md or CLAUDE.md when a directory uses this triplet model. Avoid shared-doc mode when docs are expected to evolve independently.

2) Instruction Depth Policy (Canonical README Must Be Agent-Operational)

Even with symlinked companions, the canonical README.md must include in-depth agent-operational guidance:

• Scope and ownership boundaries
• Safety guardrails and non-negotiable constraints
• Conventions and coding standards (repo-specific, not generic)
• Engineering playbook (setup, test/lint/build, release, and troubleshooting)
• Do/Don't rules with concrete examples
• Verification workflow (how to prove claims and detect stale docs)
• Escalation policy for ambiguity, destructive actions, and risky operations

If this depth is missing, the triplet setup is incomplete even when symlinks are correct.

Required Commands

Default workflow (recommended):

scripts/set-doc-triplet.sh <dir>
scripts/set-doc-triplet.sh <dir> <relative-path-to-canonical-README.md>
python3 scripts/check-doc-triplets.py [root]

Manual fallback (advanced):

cd <dir>
# Standard pattern
ln -snf README.md AGENTS.md
ln -snf README.md CLAUDE.md

# Shared-doc exception
TARGET=<relative-path-to-canonical-README.md>
ln -snf "$TARGET" README.md
ln -snf "$TARGET" AGENTS.md
ln -snf "$TARGET" CLAUDE.md

Instruction-quality review (always run during doc updates):

rg -n 'TODO|TBD|FIXME|maybe|probably|should be' README.md
python3 scripts/check-doc-triplets.py [root]

Workflow

1. Classify scope first
   • Root doc, workspace/app/package/service doc, or module/feature doc.
   • Choose section depth based on scope (load references/scope-patterns.md).
2. Confirm triplet mode is actually desired
   • If separate files are intentionally different, do not enforce symlink triplets.
3. Build a verified fact inventory
   • Verify every filepath, variable/constant name, command, dependency, and version.
   • Do not trust existing docs without checking live code/config.
   • Use references/fact-verification.md.
4. Design instruction architecture before prose
   • Separate hard constraints from preferences.
   • Convert vague guidance into trigger/action rules ("When X, do Y").
   • Keep rules agent-agnostic (do not assume one tool's private behavior model).
   • Plan sections for guardrails, conventions, playbooks, and escalation.
   • Use references/agent-instruction-best-practices.md.
5. Plan with temporary structure files (avoid local maxima and context bloat)
   • Create concise planning files under .tmp/readme-doc-plan/ before final writing.
   • Keep these as bullet/TODO artifacts, not full prose.
   • Use references/low-context-drafting.md.
6. Draft and refactor in two passes
   • Pass 1: structure, ownership boundaries, must/never rules, key commands.
   • Pass 2: remove stale statements, collapse duplication, sharpen actionability and determinism.
7. Enforce triplet symlink policy
   • Use scripts/set-doc-triplet.sh for direct creation/fix.
   • Use scripts/check-doc-triplets.py for audit + auto-generated fix commands.
8. Run instruction quality QA
   • Check for ambiguity, contradictory rules, and missing failure-mode guidance.
   • Ensure there are repo-specific examples (good/bad where useful).
   • Re-check against references/agent-instruction-best-practices.md.
9. Final freshness pass
   • Re-run verification commands for anything changed during writing.
   • Ensure scope-appropriate content (no root-only setup details in small module docs).

Quality Bar

All of the following are required:

• Content is specific, verifiable, and operational.
• Commands are copy/paste ready and runnable from stated working dirs.
• Paths exist at time of writing.
• Variable names and constants match code exactly and are used as described.
• Versions reflect currently installed/declared dependencies.
• Instructions use explicit trigger/action language, not soft ambiguity.
• Hard constraints are clearly marked as mandatory vs optional guidance.
• Safety-critical operations include confirmation/escalation rules.
• Examples demonstrate both compliant and non-compliant patterns where ambiguity risk is high.
• Wording stays agent-agnostic and portable across tooling.

References

• references/scope-patterns.md
• references/fact-verification.md
• references/low-context-drafting.md
• references/agent-instruction-best-practices.md

Scripts

• scripts/set-doc-triplet.sh
• scripts/check-doc-triplets.py