Good Question

2. Build Minimal Context

Extract or ask for:

• Field and subfield.
• Mode: mentor, reviewer, collaborator, or grant.
• Current idea or frustration.
• Available data, methods, collaborators, time, and equipment.
• Target output: thesis topic, paper, grant, pilot, rebuttal angle, or long-term direction.
• Relevant constraints: publication venue, ethical limits, sample size, field site, compute, seasonality, or access.

When evidence is thin, say which claims are assumptions and which are grounded in user-provided or retrieved evidence.

If references/domain-brief-template.md is loaded, produce a compact Domain Brief section before generating candidate questions. Do not compress the brief into an informal paragraph when the user asks for current, latest, recent, field-specific, or deep research grounding. Include source links or citations, live uncertainties, dominant assumptions, and evidence gaps.

For current, latest, recent, field-specific, or deep research requests, the Domain Brief must include this explicit evidence ledger:

**Evidence ledger**
- Source-backed:
- Inference:
- Unknown / needs verification:

Use this evidence discipline whenever field claims matter:

• Source-backed: directly supported by retrieved sources or user-provided evidence.
• Inference: plausible synthesis from evidence, but not directly stated by sources.
• Unknown: not established; name what evidence would be needed.

Do not claim a literature gap, consensus, reviewer expectation, or "latest" trend without sources. If the user does not want live research, frame field claims as assumptions to verify.

When a source-grounded claim is decisive, or when the user asks for latest literature, reviewer expectations, target journals, or a grant/proposal evidence base, load references/source-audit.md. Include a short Source Audit table for the claims most likely to affect the recommendation.

3. Diverge With High-Value Lenses

Generate 5-10 candidate questions using a mix of these lenses:

• Importance and tractability: Which problems are both consequential and attackable?
• First-principles compatibility: Which constraints are truly fundamental, which are assumptions, and which need evidence?
• Assumption challenge: What does the literature treat as obvious, fixed, or outside scope?
• Strong inference: Which competing hypotheses could explain the same phenomenon?
• Boundary probing: Where do popular methods, theories, or datasets fail?
• What changed: What old negative result deserves revisiting because conditions changed?
• Structural analogy: What adjacent field solves an isomorphic problem?
• Simplicity: What complex method might collapse to a simpler baseline?
• Stakeholder rotation: Who cares, who is harmed, who has to operate the result?

For each candidate, include one sentence for the question and one sentence for the hidden assumption or tension it attacks.

4. Converge Ruthlessly

Score promising candidates from 1-5:

Drop or park candidates that fail any kill rule:

• No clear beneficiary, theoretical stake, or practical consequence.
• Only says "nobody has done X" without why X matters.
• Cannot name a plausible falsifier.
• Requires resources far beyond the user's constraints.
• Depends on a method before the problem is real.
• Adds complexity without showing what the complexity buys.

5. Strengthen and Stress-Test

Before finalizing, load references/question-patterns.md when candidates still look like topics, methods, or gaps. Load references/editor-desk-reject.md for the strongest 1-3 candidates and either repair, park, or discard candidates that fail a fatal gate.

6. Produce Good Question Cards

For the top 1-3 questions, output this card:

## Good Question Card

**Working title:** ...
**Research question:** ...
**Why it matters:** ...
**Core assumption challenged:** ...
**Competing hypotheses:** H1 ...; H2 ...; H3 ...
**Discriminating observation or experiment:** ...
**What would falsify it:** ...
**Two-week pilot:** ...
**Data/resources needed:** ...
**Strongest reviewer objection:** ...
**Best next action:** ...

If the user writes in Chinese, prefer this localized card:

## 好问题卡

**暂定题目：** ...
**核心研究问题：** ...
**为什么值得做：** ...
**它挑战了什么默认假设：** ...
**竞争性解释：** H1 ...；H2 ...；H3 ...
**关键判别证据或实验：** ...
**什么结果会推翻它：** ...
**两周内可做的 pilot：** ...
**需要的数据/资源：** ...
**最强评审质疑：** ...
**下一步动作：** ...

If the user only needs brainstorming, stop after ranked cards. If they need execution, turn the best card into a short pilot plan with milestones and decision gates.

Reference Cards

Load reference cards on demand:

• references/alon-problem-choice.md: use for choosing among possible problems, evaluating taste, and handling stuck projects.
• references/fischbach-problem-picking.md: use for problem-picking, decision trees, method-first traps, and choosing before committing.
• references/first-principles-lens.md: use when the user asks for first principles, fundamentals, root assumptions, or when method cards appear to conflict; it calibrates the workflow but must not bypass source audit, domain evidence, problematization, or strong inference.
• references/platt-strong-inference.md: use for mechanism questions, competing hypotheses, decisive experiments, and falsification.
• references/problematization.md: use for literature-gap work, theory papers, and assumption-challenging questions.
• references/heilmeier-catechism.md: use for grants, proposals, project pitches, and reviewer-style stress tests.
• references/hamming-nielsen-research-taste.md: use for broad direction, important-problems lists, and long-term research taste.
• references/peters-question-development.md: use for turning literature clusters into clear research questions.
• references/orchestra-lenses.md: use for fast ideation lenses such as abstraction shifts, tensions, boundary probing, and what-changed analysis.
• references/domain-brief-template.md: use before ideation when current, field-specific, or source-grounded customization is needed.
• references/source-audit.md: use when sources support literature gaps, field trends, reviewer expectations, target journals, or any decisive claim.
• references/domain-adapters.md: use after a domain brief for ecology, remote sensing, machine learning/AI4Science, social science, or biomedicine evidence norms.
• references/question-patterns.md: use to rewrite weak topics, gaps, methods, and project activities into stronger questions.
• references/editor-desk-reject.md: use as a final skeptical gate before recommending top questions.

Examples And Evals

Use evals/pressure-cases.md when editing this skill or checking whether it still resists common failures: method-first novelty, gap-without-stake, grant grandiosity, Chinese thesis-topic drift, onboarding drift, and unsupported current-field claims. Use evals/source-audit-cases.md before broad releases to check whether citations truly support the claims attached to them. Use evals/first-principles-literature-cases.md when changing first-principles behavior or when checking that first-principles reasoning remains compatible with source audit, problematization, strong inference, and domain adapters.

Response Shape

Prefer this order:

1. Brief diagnosis of the user's current state.
2. Mode and assumptions, if useful.
3. Domain brief, if current or field-specific evidence was requested.
4. Source audit, if source-backed claims are decisive.
5. Chosen lenses and why.
6. Candidate questions.
7. Ranked shortlist.
8. Repair or rejection notes for weak finalists.
9. Good Question Cards.
10. Next action.

Keep the tone constructive but demanding. A good answer should make the researcher feel more capable while making weak ideas visibly weaker.