Ask Azhou

Ask Azhou

🦊 阿舟 · Ask Azhou

🚪 说清你想做什么，我告诉你敲哪扇门。

Invocation class: user-invoked orchestrator, declared in frontmatter; the axis, its semantics and the composition rules are defined once in docs/skill-standard.md §2.2 (spec #126 ID-9) — invoked by name or by an explicit "which skill" request; it recommends and never invokes another skill on the user's behalf.

It is guidance, not a script. This is the hub's front door: sixteen canonical skills, organized by what you are trying to do. Answer with a recommendation (skill, mode, and the boundary that matters), not an invocation.

Brand protocol

Emit this exact display event once:

🦊 阿舟 · Ask Azhou 启动｜mode=route｜scope=<catalog>

Use ✅ 验证通过 only after the recommendation names a real canonical skill whose boundary fits the request. Use ❌ 验证失败 when no canonical skill fits and 🔒 阿舟暂停这一项 when the request is outside this hub's catalog entirely. Emoji is display-only; keep JSON keys, schema values, digests, paths, commands, test names, and raw evidence emoji-free. A host without Unicode may remove the leading emoji while preserving the fixed text, ｜ separators, fields, and values. Raw evidence stays out of receipts unless the user supplied it.

Route by intent

Ask which family the request belongs to, then the branch question.

Install, diagnose, or prove the hub itself (发行基建)

• "Set this hub up / repair a broken install / migrate or uninstall a skill" → azhou-setup — plan first, apply the exact reviewed plan.
• "Is my checkout healthy / what broke / is a lease stuck?" → azhou-doctor — read-only diagnosis, no mutation.
• "What version / commit / branch / installable inventory is this?" → azhou-info — only provable repository facts.
• "Prove this repository state passes its gate before I claim done" → azhou-verify — the full authoritative gate, exit code and all.

Produce or polish a visible artifact (内容生产力)

• "Explain this to someone who knows nothing, with pictures" → eli5 — one self-contained HTML, big pictures, few words; refuses precision-critical asks.
• "Draw an accurate, editable diagram (architecture, flow, sequence)" → excalidraw-diagram — real render, deterministic style and layout gates.
• "Turn this complex result into a rich, reviewable HTML artifact" → lavish — artifact, relay, review, export, or share mode.
• "Make my agent's replies terse / action-first / commit-message-ready" → super-caveman — output-behavior modes, explicit enable/disable.
• "Write an architecture design document from upstream sources" → arch-doc — end-to-end authoring with calibration and review gates.

Keep knowledge and quality honest across sessions (工程纪律)

• "Persist this verified decision / debug fact / convention for future sessions" → llm-wiki — private local store; decisions carry a lifecycle and archived pages are tamper-evident.
• "Reconcile docs, rules, and memory at task close" → repo-pedant — the closeout protocol with inventory proof.
• "Write or trim prose without losing facts / audit it for leaked reasoning" → prose-standard — the complete-proposition rule and the leakage taxonomy.
• "Make my tests reliable under real CI concurrency / diagnose a flake" → ci-test-reliability — topology-aware design guidance and the flake-masking blacklist.
• "How am I actually using my agent / give me a usage report from my local session history" → session-insights — fact-bound aggregate reports from local session stores (Claude Code adapter; Codex and zcode fail closed until verified); it analyzes local stores only, never monitors, never sends anything out.
• "Which skill for this?" → this skill, ask-azhou.

Long-running experiments

• "Run, resume, or report nanochat training experiments in my pinned checkout" → autoresearch — user-owned environment, explicit boundaries.

Boundaries

• Recommendation only: name the skill and its mode; the user stays in control of what runs.
• Catalog edges: questions about mattpocock-skills workflows (tdd, grilling, specs) belong to the locally installed ask-matt, not this map — point there instead of duplicating it.
• No request fits everything: when no canonical skill owns the intent, say so plainly and name the nearest boundary.

Verification

The routing map must name every canonical skill; the repository gate enforces that parity (add or rename a skill without updating this map and the gate goes red (removal is caught by discovery parity)). End with a stable receipt carrying schema, status, current truth, artifacts/changes, verification, holds, next action, and learning signal; artifacts is empty for a routing answer. The router-pattern source pin and update path live in provenance.md.