What problem does it solve? Turning a broad research direction into concrete, publishable ideas is slow and error-prone: researchers must survey the literature, spot gaps, check novelty, and guess which ideas are worth GPU time. This Skill automates that pipeline, producing a ranked idea report backed by literature evidence, cross-model review, and pilot experiments. ## Core Features & Use Cases - Landscape survey and gap analysis: Scans local paper libraries and recent literature (top venues, arXiv) to map sub-directions, contradictions, untested assumptions, and open problems. - Cross-model idea generation and jury: Fans out candidate generation across analytic lenses, then uses an external reviewer backend (Codex or manual review MCP) for brainstorming, devil's-advocate triage, and ranking. - Pilot experiments with GPU budgets: Runs cheap parallel pilots (max 2 hours per GPU, 8 GPU-hours total) via /run-experiment and /monitor-experiment, then re-ranks ideas on empirical signal. - Use Case: A researcher says "find ideas for sample efficiency of offline RL with image observations" and receives an IDEA_REPORT.md with ranked ideas, novelty scores, pilot results, eliminated dead ends, and a suggested execution order. ## Quick Start Ask the assistant to run /idea-creator with a specific research direction such as "factorized gap in discrete diffusion LMs" to generate a ranked report of validated research ideas.
