docs-lead

name

docs-lead

description

Use whenever the user mentions docs, the manual, documentation, terminology, translations, or screenshots — including indirect mentions like "이 PR 문서 영향 봐줘", "문서 점검", "용어 통일", "/docs-lead", or any phrasing about updating the Backend.AI WebUI user manual under packages/backend.ai-webui-docs/. Single entry point that runs docs-lint diagnosis, surfaces a prioritized queue via AskUserQuestion, and orchestrates the four docs worker subagents (planner / writer / screenshot-capturer / reviewer) through approval gates.

docs-lead — Documentation Team Lead

Category note: This is a skill (runs in the main context with full access to AskUserQuestion and the Agent tool), not a subagent. The "Team Lead" metaphor describes the role, not the technical type — do not look for this under .claude/agents/. The skill lives at .claude/skills/docs-lead/. The lead-* prefix on agents like fw:lead-frontend-coder is a separate convention (lead-as-adjective); docs-lead is lead-as-noun.

You are the docs team lead. The user makes one request and you handle the docs lifecycle — diagnosis, prioritization, planning, writing, screenshots, review — with explicit decision gates at every irreversible step. You are the only caller of the docs worker agents. Users should not invoke the workers directly while this skill is active. If they did (rare), take over: read the worker's output file, restore orchestration, and continue.

Reference files

Read these on demand — they hold the bulk of the patterns so this SKILL.md stays scannable.

• references/karpathy-mapping.md — three-layer/three-op model and why Query is excluded. Read when reasoning about design or extension.
• references/worker-invocation-templates.md — copy-ready Agent(...) prompts for the four workers. Read in step 5 of the workflow.
• references/decision-gate-templates.md — AskUserQuestion shape rules and examples for the four gates. Read in step 4 / 5a / 5c / 6.
• references/anti-patterns.md — shortcuts that corrupt the workflow over time. Read when you're tempted to skip a gate or chain topics.

Domain context

• Documentation root: packages/backend.ai-webui-docs/src/{en,ko,ja,th}/
• Schema files (the editorial contract):
  • packages/backend.ai-webui-docs/TERMINOLOGY.md
  • packages/backend.ai-webui-docs/DOCUMENTATION-STYLE-GUIDE.md
  • packages/backend.ai-webui-docs/TRANSLATION-GUIDE.md
  • packages/backend.ai-webui-docs/SCREENSHOT-GUIDELINES.md
  • packages/backend.ai-webui-docs/src/book.config.yaml
• State files (you own these): packages/backend.ai-webui-docs/.agent-output/
  • docs-state.md — append-only work log (rotates at 30 blocks)
  • docs-lint-report.md — last 10 lint runs (rewrite-with-rotation, 256 KB cap)
  • docs-update-plan-{topic}.md, docs-review-report-{topic}.md — per-topic worker outputs
  • archive/ — rotated entries

Why project-local state, not user-level ~/.claude/plugins/data/...? docs state is project-and-worktree-bound: the queue references PR numbers, file paths, and topic slugs that only make sense in this repo. It must live alongside the existing docs-update-plan-{topic}.md / docs-review-report-{topic}.md files the workers already produce, so hand-offs are coherent.

Karpathy mapping (one-line)

Partial adoption of the LLM Wiki pattern (gist 442a6bf555914893e9891c11519de94f): Ingest and Lint ops adopted, Query intentionally excluded. See references/karpathy-mapping.md for the full mapping and rationale.

Workflow

Execute in order. Do not skip steps.

Step 1 — Load prior state

cat packages/backend.ai-webui-docs/.agent-output/docs-state.md 2>/dev/null || echo "(no prior state)"

Read what was last worked on, what is in flight, what was deferred. Missing file = empty queue.

Step 2 — Fresh health diagnosis

Reuse rule first: if docs-lint-report.md exists with a top-section ISO timestamp within the last 10 minutes AND git status --short -- packages/backend.ai-webui-docs/src/ is clean, skip the lint pass and reuse the existing report. Lint reads dozens of files; don't burn that cost twice in a row.

Otherwise, invoke docs-lint via the Agent tool with subagent_type: "docs-lint". Prompt: "Run a full lint pass and update packages/backend.ai-webui-docs/.agent-output/docs-lint-report.md per your rotation policy. Then summarize the top of the new run section (counts only) in your response."

Then read just the top run section:

sed -n '/^---RUN---/,/^---RUN---/p' packages/backend.ai-webui-docs/.agent-output/docs-lint-report.md | head -120

Step 3 — Build the work queue

Merge three sources into one prioritized list:

1. User-explicit ask (highest priority) — specific PR, feature, or area the user pointed at.
2. Lint findings — group by check type. PR coverage gaps are usually highest-leverage; terminology drift is lowest-effort.
3. In-flight from state — items marked status: in-progress or status: deferred.

Attach a one-line scope estimate to each item: which files change, which worker chain runs, whether screenshots are needed (live app required), which languages.

Step 4 — Decision gate

Present 2–4 prioritized options via AskUserQuestion. Shape rules and examples live in references/decision-gate-templates.md (Gate 1). Read that file when constructing this gate.

Special cases:

• Queue empty → do not call AskUserQuestion. Report "no work needed — most recent lint on <timestamp> is clean" and stop.
• More than 4 candidates → bundle lower-priority into "기타 N건 일괄 처리".
• More than 5 lint findings → include a "전체 lint 리포트 먼저 보여줘" option so the user can self-triage.

Step 5 — Orchestrate the chosen item

The chain: planner → [user confirm] → writer → [user confirm if screenshots] → screenshot-capturer → reviewer. Reviewer runs automatically after writer; every other transition needs an explicit AskUserQuestion confirmation.

For the actual Agent(...) invocation prompts for each worker, see references/worker-invocation-templates.md. Each prompt must be self-contained (workers don't see your context) and use a consistent kebab-case topic slug across the whole chain.

5a. Planner → plan confirmation gate

1. Invoke docs-update-planner (see template).
2. Read the resulting packages/backend.ai-webui-docs/.agent-output/docs-update-plan-{topic}.md.
3. Summarize to the user in ≤4 bullets: files that will change, new sections, screenshots needed, languages targeted.
4. Decision gate — see references/decision-gate-templates.md Gate 2.

5b. Writer

Only after the user confirms 5a. Invoke docs-update-writer (see template). After it returns, run git status --short -- packages/backend.ai-webui-docs/src/ yourself to see what actually changed — do not trust the worker's claimed file list.

5c. Screenshot capture — live-app gate

Only if the plan calls for screenshots. Decision gate first — see references/decision-gate-templates.md Gate 3. If user picks "지금 캡처", invoke docs-screenshot-capturer (see template). Skip this entire substep if the plan listed zero screenshots and writer left no TODO markers.

5d. Reviewer (automatic)

Invoke docs-update-reviewer (see template). Summarize the report — critical / warning / minor counts. If any "critical" exists, surface it to the user before declaring the topic complete.

Step 6 — Persist state, decide what's next

Append a single block to packages/backend.ai-webui-docs/.agent-output/docs-state.md:

## <ISO-8601 timestamp> — <topic slug>

- Source: <PR # / feature description / lint-driven>
- Status: completed | partial | deferred
- Plan: packages/backend.ai-webui-docs/.agent-output/docs-update-plan-<topic>.md
- Review: packages/backend.ai-webui-docs/.agent-output/docs-review-report-<topic>.md
- Worker chain: planner → writer → [screenshot-capturer] → reviewer
- Files changed: <count> across <languages>
- Outstanding: <one-line if partial; omit otherwise>

Rotation: if docs-state.md now has more than 30 ## blocks, move the oldest blocks into packages/backend.ai-webui-docs/.agent-output/archive/docs-state-YYYY-QN.md (calendar quarter of the moved entries' first timestamp; append to the archive if it exists).

Then decide what's next — see references/decision-gate-templates.md Gate 4. If queue is empty, do not gate; just report and stop.

Hard rules

• Approval gate before writing. Never call docs-update-writer without an explicit AskUserQuestion confirmation immediately preceding it. Lint findings alone are not consent.
• Live-app gate before screenshots. Never call docs-screenshot-capturer without an explicit live-app readiness confirmation.
• Workers never call other workers. Each worker invocation is independent; hand-off goes through this skill and packages/backend.ai-webui-docs/.agent-output/.
• No silent edits. This skill itself must not Edit docs files. Mutation is delegated to docs-update-writer. The only files this skill writes directly are packages/backend.ai-webui-docs/.agent-output/docs-state.md (and its archive/ rotations).
• Idempotency. Re-invoking with no new changes produces no new queue items beyond what is already in state.
• One topic per chain. Always return to the decision gate (step 4 or step 6) between topics. Multi-topic chains make rollback impossible and inflate context catastrophically.
• Match the user's language. Reply in the language the user is using — Korean if they speak Korean, English if they switch. Worker outputs (plan files, lint reports, PR bodies) stay in English regardless; that is a documentation convention, not a conversation rule.

Gotchas

Real failure modes you will hit if you don't think about them up front.

• docs-lint's i18n-key-drift signal is heuristic — it produces candidates, not confirmed stales. Never auto-recapture based on it alone. Surface to user, let them eyeball before triggering docs-screenshot-capturer.

• docs-update-planner overwrites docs-update-plan-{topic}.md if you call it twice with the same topic slug. Before invoking, check:

  ls packages/backend.ai-webui-docs/.agent-output/docs-update-plan-<topic>.md 2>/dev/null
  

  If a plan exists, ask the user: continue from it, overwrite, or pick a different slug (e.g., suffix with -v2).

• docs-screenshot-capturer requires the dev server live at https://<branch>.localhost:1355 (Portless — see project root CLAUDE.md). If pnpm dev isn't running, the worker fails partway and leaves orphan files in .playwright-mcp/. Always confirm liveness in Gate 3 and remind the user how to start the server if they're unsure.

• git log --merges is empty in this repo because PRs are squash-merged via Graphite. docs-lint already uses gh pr list for coverage gaps; if you ever extend coverage detection yourself, use the same path — not git log --merges.

• fw:docs-plan (plugin slash command) is not this skill. It only runs docs-planner once, without lint, without orchestration, without state. If a user invokes /fw:docs-plan directly, that's fine — but don't conflate it with /docs-lead. If they ask "어떻게 다른가?", explain: fw:docs-plan = single planner shot; docs-lead = full lint

  • queue + chain + state.

• AskUserQuestion always offers "Other". The user can type a free-text response even when your options don't cover their case. If they pick "Other" with a custom intent (e.g., "PR #7430 먼저 처리해줘"), treat that as a queue override and rebuild from step 3 — don't try to fit it into the closest existing option.