documentation-specialist

Full-project documentation maker for orchestrator template or any forked app repo. Produces a GitHub Docs-style navigable site (docs/index.md + sections) plus agent cache. Uses .github/skills/chain/SK

SKILL.md

GitHub

Documentation Specialist

Purpose: Produce a complete, navigable documentation set for this repo or any project — human-readable like GitHub Docs (hub → sections → step-by-step guides) plus agent cache in docs/codebase/.

Docs only. Do not edit application source, migrations, or tests unless the user explicitly expands scope.

Blueprint: references/doc-site-blueprint.md

Content policy (non-negotiable)

Allowed Forbidden
Step-by-step procedures, verify steps, Next steps links Secrets (keys, tokens, passwords, real .env values)
Chain/skill/loop usage, manifest shapes, branch workflow Trade secrets (undisclosed business logic, proprietary algorithms)
Architecture at appropriate abstraction Coding tips, hacks, clever shortcuts, opinionated micro-optimizations
Config field names and purpose with placeholders (YOUR_API_KEY) Production credentials, private URLs with auth

When unsure → omit or use placeholder. Reference CONVENTIONS.md for standards, do not invent new "tips."

Phase 0 — Manifest & scope (required)

  1. Read .github/project-manifest.yaml (or .claude/project-manifest.yaml) — stack, paths, token_policy.

  2. Read docs/codebase/README.md, docs/index.md (if present), docs/codebase/.codebase-scan.txt, latest TODO/*.md.

  3. Classify repo:

    • Template (stack.framework: generic) — orchestrator prompts.github/skills/chain/SKILL.mds/loops/skills
    • App fork (laravel, django, etc.) — product guides, runtime, APIs, runbooks
  4. Parse user scope ($ARGUMENTS):

    • full — full doc site + cache + optional codebase scan if stale
    • refresh — update from cache (no full scan)
    • architecture | onboarding | api | integrations — targeted section

Phase 1 — Choose chain (opt-out allowed)

User intent Chain Steps
Full docs / onboarding / stale cache .github/skills/chain/SKILL.md documentation-full load-cache → read-codebase → readme-specialist → documentation-specialist
Update docs from current cache .github/skills/chain/SKILL.md documentation-refresh load-cache → readme-specialist → documentation-specialist
Single section only No chain — minimal cache load, then this skill —

Respect no chain / skip chain opt-out.

Emit when chaining: Chain: <id> (<n> steps, tier <token_tier>)

Phase 2 — Chain handoffs (when chaining)

Pass ≤80 tokens between steps:

{"doc_scope":"full","stack":"generic","cache_stale":false,"site_targets":["docs/index.md","docs/guides.github/skills/chain/SKILL.mds-and-skills.md"]}
Step Delivers
load-project-cache-first Cache cited; staleness flag
read-codebase Updated docs/codebase/*, scan marker, repo memory
readme-specialist Root README.md with link to docs/index.md
documentation-specialist (execute) Full doc site + cache alignment + gap report

Phase 3 — Execute (two layers)

Layer A — Human doc site (GitHub Docs style)

Required for full scope. Create or refresh:

docs/
  index.md                    # Navigable hub (sidebar-style lists)
  getting-started/
    index.md
    quickstart.md
    project-overview.md
  guides/
    index.md
    daily-workflow.md
    chains-and-skills.md
    documentation.md
  reference/
    index.md
    manifest.md
    chains.md
    skills.md
  operations/
    index.md
    testing.md
    delivery.md

Every page must have: Overview → Before you begin → numbered Steps → Verify → Next steps (2–3 links).

Every section must have index.md listing child pages with one-line descriptions.

See references/doc-site-blueprint.md for templates and app-repo extensions.

Layer B — Agent cache (docs/codebase/)

Keep cache lean for AI sessions (may overlap summaries with Layer A but cache stays section-oriented):

File Contents
docs/codebase/README.md Cache index; link to docs/index.md for humans
ARCHITECTURE.md Manifest → cache → chains/loops
STRUCTURE.md Directory map
CONVENTIONS.md Branch, sync, script-first, token rules
TESTING.md Audit scripts, CI gates
INTEGRATIONS.md GitHub, sync targets
CONCERNS.md Numbered risks + mitigations (no secret detail)
STACK.md Tooling (app repos)

App repos — add when applicable

docs/guides/local-runtime.md, docs/reference/api.md, docs/operations/deploy.md, existing docs/guides/, docs/runbooks/.

Quality rules

  • [UPDATED YYYY-MM-DD] on every touched file
  • Relative links only; hub reachable in ≤2 clicks from any page
  • Factual claims cite path from cache or read-codebase — no invented behavior
  • .github/skills/script-not-shell/SKILL.md for multi-line shell (CONCERNS §6)
  • After .grok/ edits: remind python3 scripts/sync_grok_to_github_claude.py
  • Red-team pass: scan draft for secrets, trade secrets, coding tips before marking complete

Phase 4 — Gap report

## Documentation complete: <scope>

- Chain used: <id or "none">
- Human site: docs/index.md + <N> pages across <sections>
- Cache updated: <list>
- Content policy: pass (no secrets/tips flagged)
- Gaps: <numbered or "none">
- Next: <1–3 actions>

Optional: reports.github/skills/chain/SKILL.mds/YYYY-MM-DD-documentation-<scope>.md

Anti-patterns

  • Flat markdown dump with no docs/index.md hub
  • Pages without Next steps
  • Embedding real credentials "for convenience"
  • Coding tip sidebars ("pro tip: use reflection to…")
  • Full codebase scan on refresh when cache is fresh
  • Skipping readme-specialist on full runs
  • Forcing a chain after user opted out

Related

  • Blueprint: references/doc-site-blueprint.md
  • Chains: documentation-full, documentation-refresh
  • README polish: /readme-specialist
  • Cache rebuild: .github/prompts/read-codebase.prompt.md
  • Composition: .github/skills/chain/SKILL.md