What problem does it solve? Starting a research project from a broad direction is hard: you risk duplicating existing work, picking infeasible experiments, or pursuing ideas reviewers will reject. This Skill systematically converts a research direction into a ranked, empirically validated shortlist of concrete ideas. ## Core Features & Use Cases - Landscape Survey: Scans local paper libraries and recent literature from top venues (NeurIPS, ICML, ICLR, arXiv) to map gaps, contradictions, and untested assumptions. - LLM-Augmented Brainstorming: Uses an external OpenAI model via Codex MCP to generate 8-12 candidate ideas with hypotheses, minimum viable experiments, and risk levels, then filters by feasibility, novelty, and impact. - Parallel Pilot Experiments: Runs cheap GPU pilots (bounded by configurable hour budgets) on the top 2-3 ideas and re-ranks based on empirical signal. - Use Case: A researcher says "find ideas on factorized gaps in discrete diffusion LMs" and receives an IDEA_REPORT.md with ranked ideas, pilot results, eliminated dead ends, and a suggested execution order. ## Quick Start Ask the assistant to generate and rank research ideas for your specific direction, for example by saying "find ideas for sample efficiency of offline RL with image observations".
