Agent Skills: Documentation Refresh Skill

Documentation Refresh Skill

Use this skill when asked to overhaul, refresh, modernize, or clean up a repository's documentation — especially for a system that is already in production and has accumulated stale README content, dead analysis docs, outdated AI-agent knowledge bases (AGENTS.md, CLAUDE.md, etc.), or misleading code comments.

This skill is repo-agnostic. It does not assume any particular language, framework, or doc tool. It encodes a phased process, a reusable checklist, and a set of anti-patterns learned from running this refresh on real production repos.

When to use this skill

• The user says something like "our docs are out of date," "refresh the README," "clean up AGENTS.md," "audit our documentation," or "set up a docs/ folder with architecture diagrams."
• A repo has legacy analysis files, dead feature references, or docs that no longer match the live code.
• The user wants a repeatable framework they can reuse across repos.

How to use this skill

1. Read CHECKLIST.md in this skill directory for the phase-by-phase process (Phase 0 through Phase 6).
2. Read ANTI-PATTERNS.md for common mistakes to avoid while auditing and rewriting docs.
3. Adapt the checklist to the target repo:
   • Confirm with the user which phases apply (a small repo may not need every phase; a large one may want each phase in its own session to protect context windows).
   • Identify the repo's actual doc tooling (Markdown + Mermaid is the default assumption below, but adjust if the repo uses something else, e.g. Sphinx, Docusaurus, or a wiki-only workflow).
4. Always produce (or update) a running "roadmap" file in the repo root (e.g. DOCUMENTATION_ROADMAP.md) that captures the plan, audit findings, and a "Lessons Learned" scratchpad. This gives future sessions (human or agent) a clear handoff point and turns the current refresh into reusable material for the next one.
5. At natural session boundaries, write or update a HANDOFF.md summarizing what was done, what was verified, and exactly what the next session should do next. Treat this as mandatory when the work will span more than one session.
6. Before declaring any phase complete, run the repo's actual lint/test command (discover it from AGENTS.md, README.md, justfile, package.json, Makefile, or CI config) to make sure doc-only changes didn't break anything, and re-check for stale references introduced by the phase's own edits.

Core principles (apply unless the user overrides them)

1. Repo docs are the source of truth. Any external wiki mirrors the repo, never the reverse.
2. Diagrams as code. Prefer Mermaid (or another text-based diagram format already used in the repo) so diagrams stay version-controlled and diffable.
3. Separate, focused guides over one monolithic doc. Split by audience and purpose: architecture, data flow, developer onboarding, operations, deployment, etc.
4. Delete stale docs outright rather than leaving them "for reference." Git history is the archive.
5. Update code-level docs (docstrings/comments) alongside Markdown. A refreshed README next to a misleading docstring is a half-finished job.
6. Phase the work and hand off explicitly. Long doc refreshes should be broken into independently reviewable phases, each with its own commit(s) and a handoff note for whoever (or whatever) picks up next.

See CHECKLIST.md for the concrete phase breakdown and ANTI-PATTERNS.md for pitfalls to avoid.