clarifying-question-discipline

name	clarifying-question-discipline
version	1.2.0
user-invocable	false
description	Hard discipline for asking the user clarifying questions: ask exactly ONE most critical question per round (priority: purpose → constraints → success criteria), include a recommended answer when asking, resolve unknowns across MULTIPLE rounds until clear, distinguish fact self-check vs decision ask-human, and do NOT rush to solutions during clarification. Also carries the investigation-first principle. Referenced by PDCA and perf hosts via frontmatter dependencies. Triggers — 「一次一问」「多轮问清」「澄清提问纪律」「推荐答案提问」 / clarifying question discipline.

Clarifying Question Discipline

Internal shared skill. Referencing workflows declare it in frontmatter dependencies and abort at startup if it is missing — no silent fallback.

The hard discipline: one question per round, multiple rounds until clear

When information is insufficient to guarantee output quality, ask the user exactly ONE most critical question per round, then wait for the answer before continuing (or asking the next one).

• Priority order for picking the question: purpose → constraints → success criteria.
• Multi-round by design: when several unknowns exist (purpose, constraints, success criteria all missing), resolve them across multiple rounds — each round's question is refined by the previous answer. This discipline does NOT mean "only one question for the whole session"; it means one question per message, iterating until the information is sufficient.
• Clarify first, answer later: clarification rounds exist to make the problem clear. Do not rush to provide solutions, conclusions, or a full answer while critical unknowns remain. Analyze and answer after the problem is sufficiently clear (or the user explicitly asks to proceed / answer now).
• Unconditional: this discipline always applies; it does not depend on optional enhancement skills being installed.

Forbidden: asking multiple questions in one message; listing several open points for the user to answer one by one; treating「一问一答」as “ask then immediately solve” (use「一次一问、多轮问清」/ one-per-turn + multi-round-until-clear instead).

Why: multi-question dumps increase user cognitive load, cause missed answers, and diffuse the AI's focus. One-at-a-time converges progressively — each question is refined by the previous answer. Rushing to answer during clarification solves the wrong problem.

Question format

Describe intent; the agent picks its own native capability (platform-agnostic, per repo rules):

• Prefer the agent's native structured-question capability (single question + multiple options); fall back to plain prose when unavailable.
• Do NOT hardcode a platform-specific tool, and do NOT enumerate "platform X uses tool A, platform Y uses tool B".

[One sentence stating the question clearly]
- A option one
- B option two
Recommended: A (brief rationale)

Recommended answer (mandatory when a reasonable default exists): every question MUST include a recommended option or stated default with a one-line rationale, so the user can accept quickly or override. MUST NOT ask a bare open question with no suggested resolution when a default is knowable.

Short-answer convention: the user may reply with just an option letter ("A", "B"); parse it and continue.

Fact self-check versus decision ask-human

Before spending the one-question slot:

1. Fact self-check — if the needed fact is readable from the repo, already-referenced files, or prior user messages, verify it yourself; do not ask the user to restate it.
2. Decision ask-human — reserve user questions for preferences, product/tech decisions, unavailable context, or confirmations that cannot be derived from evidence you already have.

Investigation-first principle

1. Investigate before speaking — for uncertain questions, investigate first, then advise. No investigation, no say.
2. Evidence over assumptions — speak with data, facts, and evidence; avoid void hypotheses.

Integration guide (for referencing workflows)

A referencing workflow keeps exactly three touchpoints in its own body (per spec clarifying-question-discipline); the full discipline text lives only here:

1. Prominent pointer — a bold/tagged one-line declaration that names this skill and the intent one-per-turn + multi-round until clear + clarify-first. Prefer: ⚠️ 主动提问：遵循 clarifying-question-discipline（一次一问、多轮问清；问清优先，不急着答）。 / ⚠️ Follow clarifying-question-discipline (one question per turn; multi-round until clear; clarify first, do not rush to answer). Do not use「一问一答」; do not restate the purpose→constraints→success priority chain or invent shorter paraphrases that can be read as "only one question for the whole session".
2. Entry-point quantity constraint — at each user-questioning step (e.g. "list open questions"), explicitly state "if asking the user, ask only ONE most critical question this turn; ask the next only after the answer".
3. Red Flags entry — "dumping multiple questions/open points on the user at once" and "rushing to answer/solve during clarification" listed as forbidden, with the one-at-a-time / clarify-first fix.

Skills that do NOT declare the dependency (e.g. standalone English skills) instead inline the full form themselves (prominent declaration + entry constraint + Red Flags + platform-agnostic phrasing).