菜单
Use when you want to learn how experts would think about a design tradeoff, architecture choice, repeated failure, or domain question. Triggers on expert names, "mentor", "panel", "debate", "what would [X] say", "stuck on", style requests.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when installing, refreshing, or re-tuning sound taste rules in a project. Triggers on "sound setup", "install taste rules", "scaffold .claude/rules", "set up sound here", starting work in a repo with no .claude/rules/, or a taste-not-defined nudge. Also use to re-run after a stack change (new framework, new package in a monorepo).
Locks the session into DELEGATOR mode — route all work to subagents and workflows, do no work directly.
Use as load-bearing work lands, or when the user asks to debrief, hand off recent work, or check what they actually own versus what AI carried. Triggers on "debrief", "what do I own", "hand over", "hand it back", "over", and recently shipping a load-bearing decision the human has not reasoned through.
Use when a pipeline stage or the human meets facts that live outside the repo — a service, database, queue, third party, or live logs — and the work depends on their current state. Triggers on "what's the state of", "check prod", "is the service returning", "snapshot the data", external-system dependencies. Repo-local work skips it.
Turn rough ideas into clear work orders before planning or building. Use when request is vague like "add a button", "make it better", "fix the thing". Triggers on ambiguous or underspecified requests.
Resolves technical HOW decisions — architecture choices, technology selection, and design patterns. Use shape when the intent is clear but the technical path is not. Triggers on "shape this", "architecture for X", "how should I build", "which pattern", "technical approach".
| name | consult |
| description | Use when you want to learn how experts would think about a design tradeoff, architecture choice, repeated failure, or domain question. Triggers on expert names, "mentor", "panel", "debate", "what would [X] say", "stuck on", style requests. |
Surface expert knowledge — names, reasoning, and the tensions between them visible — so the user can learn and decide for themselves.
Two modes. MENTOR leads — the un-forced default. PANEL is opt-in.
| Mode | Breadth | Shape |
|---|---|---|
| MENTOR | one lens, deep | a single expert reasons through the question in their own voice |
| PANEL | multi-perspective | 2-4 experts, each reasoning visibly; the tensions between them are the point |
Refute is an orthogonal opt-in flag, never a mode and never the default. When the user asks to be challenged, experts attack the chosen position to find where it breaks instead of surveying alternatives. Available on either mode; off unless the user turns it on.
In PANEL, an optional goal tunes only how many experts and how diverse — nothing more. No goal forces a reasoning posture or pushes toward one answer.
| Goal | Count | Diversity |
|---|---|---|
| depth-novelty | 2-3 | same + adjacent domain |
| coverage | 3-4 | cross-domain |
| unblock | 2-3 | cross-lens |
Default: depth-novelty. Goal selection is optional — skip it and the default stands. (To challenge a position, turn on the refute flag; it is not a goal.)
These rules govern how consult communicates.
Domain-match the question into a pool — over-fetch to ~6-8 candidates so any mode has reshuffle slack. The pool is a fact about the question: fixed once, re-routed only if the question changes. No text output — go straight to Step 2.
One AskUserQuestion. Options: MENTOR (recommended — one expert, deep) and PANEL (multi-perspective). Each option's detail panel shows the candidate expert name(s) it would draw from the pool, so the user shapes the lens. Include refute as a selectable add-on in the prompt (off by default).
Count + Diversity (default depth-novelty).Fan out one Agent per selected expert, all launched in a single message so they run in parallel.
Each agent's prompt: read the profile at its absolute path; reason through the question from documented positions applied to the user's context, respecting the "Would NEVER Say" guardrails; for living figures prefer newer model knowledge past the Verified: date; if the refute flag is on, attack the user's position to find where it breaks. Return the expert's actual argument — a few sentences of reasoning the user can learn from — plus, for PANEL, one dissent line stating where this expert pushes back against the likely consensus. Attribution is preserved, not stripped — names stay attached.
Present each expert's reasoning as attributed prose (Presentation rules). MENTOR: one expert's argument, in depth. PANEL: each expert's position, then the tensions between them named explicitly.
Close with one AskUserQuestion for direction — never a forced recommendation:
Profiles live in profiles/. Route by domain:
| Domain | Profiles |
|---|---|
| React / State | abramov |
| CSS / Styling | wathan |
| Design Systems | frost |
| Web Animation | perry |
| TypeScript (type-level) | vergnaud |
| JavaScript | simpson, osmani |
| Go / Systems | pike, cox |
| Distributed Systems | lamport, kleppmann, helland |
| Formal Methods / Verification | lamport, kleppmann |
| Concurrency | pike, armstrong, lamport |
| Python | hettinger |
| Performance | gregg, osmani, muratori |
| Architecture / Patterns | fowler, martin, alexander |
| TDD / Testing | beck, freeman, hughes |
| DDD | evans, vernon |
| Event Sourcing / CQRS | young |
| Legacy / Refactoring | feathers, fowler |
| Microservices | newman |
| Rails / Monolith | dhh |
| DevOps / Observability | hightower, majors, humble, forsgren |
| REST / HTTP | fielding |
| API / Library Design | bloch |
| Product Management | cagan, jobs |
| UX / Design Psychology | norman |
| Design Leadership | zhuo |
| Startups | graham, dhh |
| Databases / Data Evolution | pavlo, helland, sadalage, young, kleppmann |
| Reliability / Stability | nygard, armstrong, cook |
| Team / Org Design | skelton-pais, forsgren, zhuo |
| Accessibility | soueidan |
| Simplicity / Data-Oriented | hickey |
| Category Theory / FP | milewski |
| FP in JS (pragmatic) | simpson |
| State Machines | khorshid |
| AI / LLMs | willison, karpathy, huyen, cherny, osmani |
| Note Systems / Memory | matuschak |
| Interactive Explanation | victor, case |
| Malleable / End-User SW | inkandswitch, litt, kleppmann |
| Knowledge Gardens | appleton, brander |
| Computing Visionaries | kay, papert |
| Decisions / Behavior | kahneman, klein, fogg, norman, simon |
| Systems Thinking | meadows, deming, snowden |
| Quality / Management | deming |
| Strategy | boyd, rumelt, goldratt |
| Constraints / Flow | goldratt |
| Communication | tufte, orwell, minto, jobs |
| Legibility / Emergent Order | geertz, jacobs, scott |
| Incentives / Metrics / Commons | goodhart, ostrom |
| Epistemology / Language | popper, kuhn, wittgenstein |
| Organizational Failure / Safety | perrow, vaughan, reason, cook |
| Evolution / Complexity | kauffman, dawkins |
| Learning | vygotsky, bruner |
| Security | schneier, shostack |
Consult surfaces knowledge — it does not execute or decide. The caller owns the decision; expert perspectives are understanding to learn from, not prescriptions to follow.