Generate, validate, and rank research ideas with pilot experiments for a given direction.
npx skills add https://github.com/zjunlp/Mechanist --skill idea-creator-zjunlp
Please help me install this Agent Skill. Skill: idea-creator Source: https://github.com/zjunlp/Mechanist/tree/main/skills/idea-creator Command: npx skills add https://github.com/zjunlp/Mechanist --skill idea-creator-zjunlp
idea-stage/LANDSCAPE.md (produced by /research-lit) and uses an external LLM reviewer to generate 8-12 differentiated ideas. - Multi-Stage Filtering: Applies feasibility, novelty, and impact checks, then runs devil's-advocate critique via the external reviewer to narrow candidates to the top 2-3. - Parallel Pilot Experiments: Launches minimal GPU experiments (with strict time and GPU-hour budgets) to gather empirical signal before committing to full research. - Use Case: A researcher says "brainstorm ideas on the factorized gap in discrete diffusion LMs" and receives a ranked IDEA_REPORT.md with hypotheses, pilot results, risk levels, and a suggested execution order. ## Quick Start Ask the assistant to run /idea-creator with a specific research direction, such as "generate research ideas for sample efficiency of offline RL with image observations".High-intent search queries and answers about installing and using this skill.
Run /idea-creator with a specific direction of 1-2 sentences covering problem, domain, and constraint. The skill brainstorms 8-12 ideas with an external LLM, filters them for feasibility and novelty, and outputs a ranked report in idea-stage/IDEA_REPORT.md.
Ideas pass through four stages: landscape-aware generation, first-pass filtering on feasibility and impact, deep novelty checks with devil's-advocate review, and parallel pilot experiments on GPUs. Only ideas with positive pilot signal are recommended.
Yes, the pilot experiment phase is skipped when no GPU is available or ideas are purely theoretical. Such ideas are flagged as "needs pilot validation" in the report so you can validate them later.
Each pilot is capped at 4 hours per GPU with a 6-hour hard timeout, at most 6 ideas are piloted in parallel, and total GPU consumption is limited to 10 hours. These constants can be overridden via arguments.
Directions that are too broad, like "NLP" or "computer vision", produce ideas too vague to experiment on. The skill stops and asks for a 1-2 sentence direction specifying the problem, domain, and constraint before proceeding.
JeanDiableAnchor a literature survey on one paper and map its research landscape.
MACJONSWOrchestrates autonomous research lifecycle from idea discovery to submission-ready output.
goupup-aiVerify research method novelty against recent academic literature.
Biajin-PKUExtract claims, evidence, assumptions, and limitations from scientific papers.