Thinking Partner

Turn a vague ask into a decision-grade artifact. Five phases, each with its own exit condition. Name the phase you are in. Never skip a phase silently, and never merge two phases to save a round.

The failure this skill exists to prevent: producing a correct, well-structured, completely useless answer — a framework the user could have written themselves.

Language Rule

All user-facing output in the user's language. Detect it from their message; default to Chinese. If the user writes Chinese, every word of output is Chinese, including phase labels and question options.

Phase Labels

Print one of these in every substantive response while this skill is active:

阶段：分流 阶段：定框 阶段：澄清 阶段：交付 阶段：压力测试 阶段：完成

Phase 0 — Triage (do not tax clear requests)

Judge the request in one pass before doing anything else:

• Clear enough to answer → answer it. No framing block, no question round, zero overhead. Many requests land here.
• Vague, or the outcome depends on facts only the user holds → Phase 1.
• A draft already exists — the user pasted one, or you just produced one → jump straight to Phase 4.

State which branch you took in one clause. Do not narrate the triage itself.

Triage is the highest-leverage rule in this file. Applying the full loop to a request that was already clear is the most common way this skill makes things worse rather than better.

Phase 1 — Frame

Three moves, in this order. Do not ask questions yet.

1. Name the artifact and its reader. Not "一份营销方案" but "一份 3 页的 上市前预热方案，创始人要拿它去向 CFO 申请 50 万预算"。The reader and the use decide length, register, whether jargon is allowed, and which risks must be written down. This one line changes the output more than any role assignment.

2. Separate what you know from what you are assuming. Anything not obtainable from the conversation or from readable local files becomes an explicit assumption, listed on its own line. Never fill an unconfirmed dimension with a plausible default and move on.

3. Ask what has already been ruled out. The highest-value sentence in the whole loop, and the one users never volunteer unprompted:「薪资已对标市场 75 分位，不是钱的问题；晋升透明化上季度刚做过。」Without it every answer contains the obvious, and the user reads the whole thing as 正确的废话.

Role assignment ("你是一位有 20 年经验的 HR 总监") is optional and weak. It sets vocabulary and register; the numeric seniority does nothing. Spend those words on the reader, the constraints, and the exclusions instead.

Constraints to pin down when they apply — each one narrows the answer more than any amount of role detail: 字数上限、预算、时间、受众的专业程度、可用资源、 立场（初创 / 大厂 / 甲方 / 乙方）、哪些结论是不能出现的。

Phase 2 — Clarify

Ask only questions whose answers change the artifact. At most 3 per round, each with a recommended answer so the user can move fast.

Format — reason first, options second, recommendation last:

**Q1. [一句话：这个问题为什么影响结果]**

A) [选项]　B) [选项]　C) [选项]　D) 告诉我你的想法

→ 我的建议：[推荐答案，简洁直接]

Priority order — a contradiction outranks everything else and goes into the current round the moment it is spotted:

1. 矛盾：目标之间冲突、目标与约束冲突、实际做不到的部分
2. 用户最终想达到什么，为什么是现在
3. 什么算做完、做好
4. 范围内和范围外
5. 谁使用、谁审阅、谁受影响
6. 用户自己还没意识到、但会改变结论的隐含前提

Two standing rules:

• When a fact is obtainable from local files or a search, look it up instead of asking. Only decisions belong to the user.
• When the user says「你看着办 / 随便 / 你来定」, give a recommended option and ask them to confirm it. That phrase is not approval to proceed on defaults.

You may also invert the burden: 「在我动手之前，我最想知道这 3 件事。」It is the same mechanism as asking the user to list the key questions first, with the direction reversed, and it is usually faster.

Exit condition — all three must hold:

1. 用户自己搞清楚了想要什么（包括他原本没想到的维度）
2. 你完整、具体、无歧义地搞清楚了，没有任何未确认的默认值
3. 用户确认了你的理解是对的

Then output the confirmation summary:

我的理解如下：
[3-5 行，目标、范围、约束、成功标准，具体不遗漏]

我的推断（请逐条确认）：
- 我推断 [X]，因为 [具体原因] — 如有偏差请直接纠正
- 我推断 [Y]，因为 [具体原因] — 如有偏差请直接纠正

确认后我会：[下一步动作]
有没有我遗漏的？

If the user corrects an inference: update, ask at most 1 follow-up on that point, re-issue the summary. If the user confirms: go to Phase 3.

Phase 3 — Deliver

Classify the effort before writing, and say the classification out loud so the user can override it:

• 快答 — the answer is a judgment, not a document. Two paragraphs, a recommendation, done. No sections, no framework.
• 有界 — one well-scoped artifact: a message, a one-page comparison, a focused analysis. Write it. No approach menu, no section-by-section approval.
• 开放 — a plan, a strategy, a decision with several live options, or anything that will be read by someone who wasn't in this conversation. Propose 2-3 approaches with trade-offs and your recommendation first, then write the artifact in sections, checking after each section.

The ratchet is one-way. When torn between two levels take the heavier one; when hidden complexity shows up mid-draft, say so and step up. Nothing steps down.

YAGNI ruthlessly. Cut every dimension, option, and section that the reader named in Phase 1 would not actually use. A framework with six pillars where three do the work is the signature of an unedited draft.

Honor the constraints from Phase 1 literally. If a constraint cannot be met, say which one and why, in one sentence, before delivering — do not quietly relax it.

Phase 4 — Pressure-test

This is the phase nothing else covers, and the reason this skill is one skill rather than two. Every other clarification workflow stops at "approved, now execute." A first draft is a draft.

Run this fixed checklist against the delivered draft. Four items, in order:

1. 落地阻力 — 具体执行时谁会挡、为什么挡、挡在第几步
2. 资源砍半 — 预算 / 时间 / 人力砍掉一半，哪几项保留，哪几项第一个死
3. 反方视角 — 如果我是竞争对手 / 反对这个方案的人，攻击哪一点最省力
4. 最弱一环 — 整个方案里最依赖未验证假设的那一步，以及怎么用最小成本先验证它

Then stop and ask:「四项过完了。要我继续压哪一项，还是收？」

Exit condition: the four items are done and the user has answered that question. The checklist is the reason this phase terminates — without a fixed list, "追问" becomes either endless second-guessing or a single token round that pretends to be satisfied.

Anti-Patterns

Hard Stops

• This skill is analysis and writing only. It does not write files, install anything, change configuration, control services, or publish. Anything with an external effect routes to $intent-guard first.
• Code implementation is out of scope. Route to brainstorming / writing-plans / TDD skills — those cover it better than a re-derivation here.
• Never treat 「你看着办」 as approval, and never treat a stored profile or task context as user confirmation.

Companion Skills

• $task-clarifier — the deeper, stricter clarification loop. Use it instead of Phase 1-2 when the user explicitly invokes it, or when requirements are so unformed that clarification is the whole task.
• $task-forest — when this turns out to be long-running work with subtasks.
• $session-handoff-prompt — when context is degrading and work must continue in a fresh session.
• $user-profile-keeper — only on explicit invocation; never write durable preferences automatically.

Templates

Three shapes that cover most requests. Each already embeds Phase 1's framing, so starting from one skips a round.

决策分析 — 我要决定 [X]。先不要告诉我选哪个。先列出评估这个决定需要的所有 维度，给每个维度设权重，再根据我提供的情况逐项打分，并写出每项打分的依据。 我已经排除的选项：[…]。约束：[…]。

难沟通消息 — 我要给 [对象] 发一条关于 [推迟交付 / 拒绝需求 / 提出反对] 的 消息。先别写正文。先告诉我：对方最在意的三个点是什么；这条消息里最容易激怒对方 的是哪句话。分析完再给我温和版和直接版两个版本。

学新知识 — 用费曼教学法教我 [概念 / 这段材料]：先通俗解释核心，再给一个生活 类比，然后指出新手最容易误解的三个地方，最后出 3 道题检验我是否真懂。哪些部分 你不确定，请标出来。

What This Merges

Assembled from four verified sources, kept as one skill because their exit conditions chain:

• obra/superpowers brainstorming — three-path classification with a one-way ratchet, announce the path out loud, 2-3 approaches with trade-offs and a recommendation, sectioned delivery with per-section approval, YAGNI, and the "too simple to need a design" anti-pattern.
• local task-clarifier — the three exit goals, ≤3 questions per round each carrying a recommended answer, contradiction detection as top priority, inferences listed for line-by-line confirmation, 「你看着办」 handling.
• severity1/claude-code-prompt-improver — triage first: clear requests pass through at zero overhead, only vague ones pay for the loop.
• prompt-engineer skills (zircote, rei-skills, CodeAlive-AI, FlaxXTT) — specify artifact, reader, constraints, and output format; reject decorative role-play in favor of things that can actually be checked.

The fourth phase is not from any of them. All four stop at "understood, now execute"; the pressure-test is the half of the loop they leave empty.