idea-gen

What problem does it solve? Turning a broad research direction into concrete, publishable ideas is slow and unstructured. This Skill automates the process by brainstorming 8-12 ideas via an external LLM, then filtering them through feasibility, novelty, impact, and Prof. He's 4-dimension evaluation framework to surface 4-6 high-quality directions. ## Core Features & Use Cases - External LLM Brainstorming: Uses Codex MCP with gpt-5.4 for divergent idea generation, with automatic fallback to Claude if the MCP tool is unavailable. - Multi-Stage Filtering: Applies feasibility checks, novelty quick-checks via web search, impact estimation, and a 4-dimension scoring framework (longevity, passion, application, uniqueness) with a 12/20 threshold. - Anti-Pattern Detection: Flags ideas that are overly trendy, overly niche, A+B stitching, or scale-dependent before final output. - Use Case: A researcher exploring "efficient transformers" runs the skill, which reads existing landscape files from a prior literature survey, generates 10 ideas, filters them down to 5 ranked directions, and writes IDEAS_RAW.md and IDEAS_FILTERED.md with scores, risk levels, and elimination reasons. ## Quick Start Ask the assistant to generate research ideas for a direction like "sample efficiency of offline RL with image observations" and it will produce ranked, scored idea reports in the outputs directory.