router

Project-agnostic. No hardcoded tree — discover, then route. Works in any repo.

1 — Discover the docs (do this first, once per session)

Run a cheap scan to learn what this project has. Cache the result mentally for the session.

bash
# docs roots + entry points
ls README* CONTRIBUTING* 2>/dev/null
fd -t d -d 2 -i 'docs?|spec|adr|decisions|wiki|reference' 2>/dev/null \
  || find . -maxdepth 3 -type d \( -iname 'docs' -o -iname 'spec*' -o -iname 'adr' -o -iname 'decisions' \) -not -path '*/node_modules/*'
# the markdown tree under the main docs dir (swap <DOCS> for what you found)
fd -e md . <DOCS> 2>/dev/null | head -80 || find <DOCS> -name '*.md' | head -80

If a hand-maintained index exists (<DOCS>/README.md, SUMMARY.md, mkdocs.yml, docusaurus.config.*), read that — it's the authoritative map. Don't rebuild what the maintainer already wrote.

2 — Route a question → one doc

Match the question's shape to a folder by its conventional name, then Read that one doc. Don't bulk-read the tree.

| Question shape | Look in (by conventional name) | |----------------|-------------------------------| | "What does feature X do?" (user-facing) | product-spec/, docs/features/, top-level README | | "How is X implemented / data flow / API contract?" | specs/, docs/architecture/, design/ | | "Why did we choose X?" | decisions/, adr/, docs/adr/ (ADR files) | | "How do I do X / convention for X?" | references/, docs/guides/, CONTRIBUTING.md | | "What's planned / status of X?" | plans/, roadmap/, docs/roadmap/, project board | | "Known issues / open bugs?" | reviews/bugs/, ISSUES.md, GitHub issues | | "What shipped recently?" | CHANGELOG.md, changelog/, releases |

If names don't match these, fall back to the discovered index from step 1.

Rules

• Map question → doc → Read exactly that doc. Grep the tree only when no index resolves it.
• Docs reflect state when written (check dates / filenames) — verify against code before acting on stale detail.
• For how-to-code questions, prefer the project's own coding-rules / conventions doc over re-deriving from spec.
• Before fixing a bug, check the bug catalog / issues for an existing entry + severity.

3 — Share docs + this skill with collaborators (GitHub)

Three layers — do whichever the user asked for.

A. Ship the skill in the repo (teammates auto-get it)

Project skills live at .claude/skills/<name>/SKILL.md and are checked into git — anyone who pulls + uses Claude Code gets them automatically, zero setup.

bash
mkdir -p .claude/skills/router
cp ~/.claude/skills/router/SKILL.md .claude/skills/router/SKILL.md   # or author a project-tuned copy
git add .claude/skills/router && git commit -m "docs: add docs router skill"

• Global copy (~/.claude/skills/) = only you, every project.
• Project copy (.claude/skills/ in repo) = whole team, this repo. Commit it to share.
• A project copy can hardcode the real tree (faster, no discovery) — keep this generic one global as the template.

B. Make docs browsable on GitHub (no Claude Code needed)

• Minimum: a <DOCS>/README.md index with relative links to every doc — GitHub renders it; links are clickable in the web UI.
• Link from root: add a "## Documentation" section in the top-level README.md pointing at <DOCS>/.
• Full site (optional): GitHub Pages via MkDocs or Docusaurus, or a GitHub Wiki for free-form pages. Pages = versioned with code; Wiki = separate, easier for non-devs.

C. Shareable onboarding link (Claude Code)

For a teammate who'll use Claude Code: create an ONBOARDING.md at repo root (point them at <DOCS>/ + the skills), then use the ShareOnboardingGuide tool to upload it and get a link they open in Claude Code. Generic and project-agnostic.

Reuse note

This skill is intentionally project-agnostic — it lives in ~/.claude/skills/ so it loads in every repo. To specialize it for one project, copy it into that repo's .claude/skills/ and replace step 1's discovery with the project's actual doc tree (like a hand-written router).