idea-generation · lingzhi227 — Claude Code skill: research ideas with novelty

What idea-generation does

idea-generation takes a research area or task description and produces 3–5 distinct ideas. Each one comes with a name, a title, an experiment plan and ratings. Then comes iterative refinement: the idea gets critiqued and sharpened while its core stays intact. The loop runs until convergence or 5 rounds.

Novelty checking runs against Semantic Scholar. The bundled novelty_check.py script performs a multi-round literature search and returns a binary verdict — novel or not — with justification. You can also search Semantic Scholar and arXiv by hand. Finally, ideas are ranked on three 1–10 scales: interestingness, feasibility and novelty.

The script needs Python, and the search itself goes out to Semantic Scholar. The skill's own rules insist ideas stay feasible with available resources — no new datasets, no massive compute. It advises against overfitting an idea to one dataset or model, and tells the agent to be a harsh critic on novelty.