kw-question

/kw-question — from a vague idea to a good research question

This is the upstream step. /kw-explore assumes you already have a question worth pursuing and builds the technical route to attack it. /kw-question decides whether the question is worth asking at all, and sharpens it until it is. Run this first when the user has an idea, a gap, or a stalled project — not yet a real question.

Method distilled from public research-craft sources — Alon & Fischbach (problem choice as a trainable skill), Platt's strong inference (rival hypotheses + discriminating tests), Alvesson & Sandberg (challenge hidden assumptions, not just gap-spot), the Heilmeier Catechism (explicit goals/risks/success-failure criteria), Hamming & Nielsen (keep a list of important, attackable problems) — and the synthesis in the good-question project (https://github.com/Rimagination/good-question). Re-expressed here; credit to them.

Outputs go in the user's project (a Question Card markdown). Nothing is written to the knowledge base — a question is not yet a principle.

The bar: a question is "good" only if it has all seven

1. Stakes — answering it changes theory, method, practice, policy, or the next research step.
2. Specificity — evidence can directly touch it; it is not a topic.
3. Rivalry — at least 2-3 competing explanations exist.
4. Falsifiability — some achievable result could weaken, revise, or kill the claim.
5. Pilot feasibility — a credible proof-of-concept can start within ~2 weeks (or the user's stated constraint).
6. Negative learning — even a failed/negative result teaches something (a boundary, a mechanism, a method).
7. Grounding — claims trace to public sources, or are explicitly labeled as inference (do not invent field consensus).

Workflow

1. Diagnose the starting point. Classify what the user actually has: a broad interest, a literature gap, a half-formed idea, a draft proposal, or a stalled project. The starting category determines what's missing.
2. Ground a short domain brief (only if the question is knowledge-dependent and you're unsure). Keep facts (with sources) separate from inference. If the library is linked, kw search the topic to see what mechanisms already exist — but do NOT invent consensus you can't ground.
3. Diverge — generate candidate questions with structured lenses, e.g.:
   • challenge a hidden assumption the field takes for granted (Alvesson & Sandberg);
   • flip the default / reverse the causal arrow;
   • name the discriminating observation two rival explanations would disagree on (Platt);
   • shrink scope until evidence can touch it; widen until it matters. Produce several variants, not one.
4. Converge — filter by the seven-point bar. Drop candidates that are only novel, have no audience, cannot fail, or teach nothing when negative.
5. Rewrite weak forms into real questions. A topic, a method, a benchmark, or "there's a gap in X" is NOT a question. Convert each into a testable proposition.
6. Editor desk-reject pass. For each survivor, state the single strongest reviewer objection and either fix it or discard the candidate. Be the harsh reviewer, not the cheerleader.
7. Emit a Question Card for each surviving question.