| name | researchstack-idea-refine |
| description | Guided idea-refinement skill for research projects that are interesting but not yet submission-ready. Use when
Codex should improve a paper idea through structured questioning, narrowing, hypothesis building, contribution
shaping, and experimental reframing instead of only approving or rejecting it.
|
Idea Refine
Read ../references/workflow.md, ../references/conference-story-patterns.md, and ../references/novelty-boundary.md.
Use guided refinement, not free-form brainstorming. Move the project through these checkpoints:
- Replace broad ambition with a specific thesis.
- Turn the thesis into 2-4 falsifiable hypotheses.
- Separate core contribution from implementation support.
- Define the minimum believable experiment set.
- Decide what to cut.
First classify the idea as one of:
- measurement,
- mechanism,
- system,
- algorithm/control,
- architecture/co-design,
- artifact/evaluation,
- abstraction.
Prefer the most concrete truthful class. Do not recast the idea as an abstraction unless the abstraction passes the Abstraction Gate in conference-story-patterns.md.
Before refining the thesis, run a light Closest-Work Gate:
- closest direct competitor,
- closest adjacent competitor,
- industrial or open-source baseline,
- foundational predecessor.
If the closest work is unknown, make literature search the first refinement step instead of inventing a stronger story.
When improving an idea, choose one or more reframing moves:
- Narrow the scope to a single claim the evidence can support.
- Strengthen the mechanism so the contribution is not only "we tuned harder."
- Recast the contribution as measurement, mechanism, interface, or analysis if full system novelty is weak.
- Add a sharper setting or workload where the idea matters most.
- Convert a vague optimization claim into a principled tradeoff story.
- Convert a vague novelty claim into the Novelty Boundary Statement form from
novelty-boundary.md.
If the idea lacks a concrete mechanism, artifact, measurement, or falsifiable phenomenon, say that directly instead of inventing a conceptual framing.
End with:
- revised title candidates,
- revised abstract skeleton,
- novelty boundary and closest-work matrix,
- story candidate: venue story pattern, X/Y/Z statement if systems, closest-work contrast sentence, and first decisive figure or experiment,
- must-have experiments,
- stop-doing list.
