decision-prompt-builder

name	decision-prompt-builder
description	At a judgment point, emit the 2-3 questions only the human modeler can answer — framed as trade-offs, not answers — and refuse to answer them. The inverse of "AI answers, human confirms": here the AI asks, the human answers, then the AI assists with the consequences.

Purpose

At every modeling-judgment point, surface the minimal set of questions only the modeler can legitimately answer, before the AI shows any suggestion — and refuse to answer them. This is the reusable form of the one good pattern that already exists in method-selector (the "前置颗粒度对齐 — ask the ONE most load-bearing question" step): make it available at every judgment gate, not just method selection.

The posture this enforces: instead of "AI proposes a verdict → human confirms it" (which collapses into rubber-stamping), the flow becomes "AI asks the question the human owns → human answers → AI assists with what follows". The AI is allowed to lay out the trade-off; it is not allowed to pick the side.

This skill does NOT answer its own questions, pre-select a recommended option, or fill any decision artifact. It produces questions; the human's answers flow into the modeler_decision / modeler_rationale fields of the relevant decision artifact or the decision log.

When to use

Call this skill FIRST, before the judgment-bearing skill shows its evidence/suggestion, at any gate whose verdict is a modeling judgment:

• G1 framing — before problem-classifier shows ai_suggested_type: what output does the team want to defend?
• G2 / G2.5 method choice + baseline — before method-selector shows its ai_suggestion.
• post-experiment (G4.5) — before result-report-generator / robustness-checker show their suggested verdicts: which result is the headline, how confident, is it robust enough?
• G5 claim scope / figure role — before paper-section-writer / figure-table-planner finalize: what is the contribution, what claim does this figure defend?

Any skill whose tier is makes_modeling_judgment (the five C-layer skills) should trigger this skill first. B-layer skills may use it for their load-bearing span (the rubric, the framing, the physical meaning).

Do NOT use this skill:

• During mechanical stages (code generation, review, freeze, render-check, the auditors) — there is no modeling judgment to ask about, and a question there is friction theatre.
• To ask a question whose answer is mechanically determinable (e.g. "should the seed be fixed?" — that is AI-owned; never ask it).

Inputs

• The current gate / judgment point and the candidate options the downstream skill will present (so the questions are grounded in the actual trade-off).
• planning/session_config.json — the mode (learning | speed), which controls verbosity and anchor-timing only.
• The problem parse / classification for context.

Workflow

1. Identify the single judgment the human must own at this gate (method choice / framing / result verdict / confidence / claim scope / figure claim).

2. Derive the 2-3 questions that, if answered, narrow the decision space the most. Rank them by how load-bearing they are. Each question MUST:

   • Be framed as a trade-off, not a leading question: "Q3 can be cast as optimization or as evaluation — optimization gives you a defensible 'optimal' claim but needs a clean objective; evaluation is safer but weaker. Which does your team want to defend?"
   • Name the consequence of each side, not the recommended side.
   • Be answerable only by a human with modeling judgment (not by computing something).

3. Emit at most 3 questions. If everything looks like a question, you have failed to prioritize — pick the 3 that matter. (Anti-fatigue: if the human is asked everything, they tune out, and the gate degrades into rubber-stamping.)

4. Hand the human's answers to the downstream skill / decision artifact. Do not author the answers. Do not mark any artifact DECIDED.

Mode behavior (learning vs speed — scaffolding only, never gate strictness)

The mode in planning/session_config.json changes ONLY how this skill prompts — never the floors, never the copy-detection, never any gate.

	learning mode	speed mode
Question count & framing	full 2-3 questions, each option's trade-off explained	one terse question, trade-off compressed to a clause
AI suggestion timing	withheld until after the human has written their answer (anti-anchoring)	shown alongside, since the expert wants a draft to react to
Term definitions	inline (defines TOPSIS, RMSE, etc.)	assumed known
Post-answer reflection	"would this convince a judge? what's the weakest part?"	omitted

Anchor suppression is the key learning-mode mechanic. In learning mode, the downstream skill's ai_suggestion must be withheld until the human commits their own answer, then revealed for comparison ("you chose ARIMA for trend-fit; the AI also flagged seasonality you didn't mention"). This directly attacks the anchoring bias that makes a human "decision" just an echo of the AI's pick. In speed mode anchoring is an accepted trade-off for an expert who can critically evaluate a draft.

The mode is recorded per decision (captured_in_mode) by modeler-decision-logger, so a post-contest review can show which decisions were made on autopilot (speed) vs deliberated (learning).

Outputs

A decision_prompt block injected at the top of the judgment skill's output (not a saved file unless the skill persists it):

## Decision prompt — your call, not the AI's  (mode: learning)

Before you see the AI's suggestion, answer these (most load-bearing first):

1. **[output form]** Q3's answer can be a *plan* (optimization — you can claim "optimal", but you need a clean
   objective and constraints) or a *score/ranking* (evaluation — safer, but you can't claim optimality). Which
   does your team want to defend, and why?
2. **[baseline]** The simplest honest reference here is greedy allocation. Do you accept that as the baseline, or
   is there a more meaningful one for this problem?

(The AI's suggestion is withheld until you've written your answer above — so it doesn't anchor you.)

Rules

• Ask, never answer. Refuse to state or pre-select the answer to your own questions, even when the human is unsure.
• Never pre-fill a modeler_decision / modeler_rationale field. Never mark an artifact DECIDED.
• At most 3 questions per gate. Fewer is better.
• Only ask genuine modeling-judgment questions — never one whose answer is mechanically determinable.
• Frame trade-offs (name both consequences); never frame a leading question that telegraphs the "right" answer.
• The mode toggle changes verbosity and anchor-timing ONLY. It NEVER relaxes a char floor, a copy-detection, or any gate. If mode == speed, the same human-authored decision-log records with the same floors are still required.
• Do not run during mechanical stages.

Verification

Before handing off, verify:

• ≤ 3 questions, each a genuine human-judgment question framed as a trade-off.
• No question has a pre-selected or AI-supplied answer.
• In learning mode, the downstream ai_suggestion is withheld until after the human answers.
• No decision artifact field was authored or marked DECIDED by this skill.

Failure modes

Stop and report if:

• The user asks this skill to answer its own question — refuse; narrow the trade-off instead, never resolve it.
• There is no genuine modeling judgment at this point (it's a mechanical stage) — do not invent a question.

Handoff

• To the judgment-bearing skill (method-selector, result-report-generator, etc.): the human's answers, to flow into the decision artifact.
• To modeler-decision-logger: the captured_in_mode value for the decision record.

Relationship to the posture shift

This skill operationalizes the "AI asks, human answers" inversion. The C-layer decision artifacts make rubber-stamping a dead end after the fact (empty/copied rationale fails the gate); this skill attacks the same problem before the fact, by making the human supply the judgment as an answer to a real question rather than as a confirmation of an AI verdict. Learning mode's anchor suppression is what makes the human's answer genuinely theirs rather than an echo.