idea-evaluator
Idea Evaluator
Overview
This skill evaluates one preliminary research idea the way a top-venue reviewer would. It returns one of three advisory verdicts: Strong Accept, Accept with Revisions, or Reject and Pivot. The target bar is a top systems or ML venue such as NeurIPS, ICML, ICLR, OSDI, NSDI, or SOSP. Novelty and significance are the two axes that decide acceptance there.
The point is to kill a doomed idea cheaply, before the author spends weeks drafting. A weak idea caught at evaluation costs an afternoon. The same idea caught at submission costs the whole project.
Control-flow core: fatal-flaw short-circuit. Run the fatal-flaws audit FIRST. If any flaw is CRITICAL, stop and emit the rejection directly. Do not score a doomed idea. Scoring is decoration on a rejection that is already certain.
Core principle: never fabricate. Do not invent prior work, DOIs, arXiv IDs, benchmark numbers, or SOTA results to justify a verdict. If a novelty claim or a baseline-recency claim rests on literature you have not checked, mark it [UNVERIFIED] and route the user to a literature check. A confident verdict built on a hallucinated baseline is worse than an honest "unverified".
Verdicts are advisory. This skill surfaces a recommendation. It never auto-rejects an idea, never edits .writing/ state, and never blocks the user from proceeding. The user owns the decision. Present the verdict, the evidence, and the actions, then let the user choose.