Guide · Agent skill
How to use the Idea Discovery Pipeline agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The idea-discovery skill is a workflow for an AI agent that takes a broad research direction and works toward a shortlist of testable ideas. It chains literature survey, idea brainstorming, novelty checking and critical review, and can run small pilot experiments along the way. This guide explains what the skill does, how to install it, how to trigger it, and what it needs before you run it.
Idea Discovery Pipelinewanshuiyin/Auto-claude-code-research-in-sleep · MIT Download the skill Quick start
- 1Check the sibling skills. Confirm that the skills-codex sub-skills (research-lit, idea-creator, novelty-check, research-review, research-refine-pipeline) are installed before this one.
- 2Set up the Gemini reviewer. Make sure the local gemini-review MCP bridge is running, since the reviewer steps depend on it.
- 3Download and unzip the skill. Download the ZIP from this page and unzip it into ~/.claude/skills/ or your project's .claude/skills/ folder.
- 4Make a natural request. Ask the agent for the idea discovery pipeline on a specific research topic, using a trigger phrase such as "idea discovery pipeline".
- 5Set budgets and checkpoint mode. Optionally state pilot budgets in your request, and choose whether checkpoints should pause for your approval.
- 6Review the output. Read idea-stage/IDEA_REPORT.md for the ranked ideas, then check the refined proposal and experiment plan for the top idea.
What is an agent skill?
An agent skill is a folder containing a SKILL.md file. The file tells an AI agent how to carry out a specific task, and the agent loads it when your request matches the skill's description. The skill can call other skills and, in this case, run scripts and access the network.
What the idea-discovery skill makes the agent do
The skill runs a pipeline in phases. It first surveys recent papers on your topic through the research-lit sub-skill, looking at arXiv, Google Scholar and Semantic Scholar to map sub-directions, approaches, open problems and gaps. It then brainstorms candidate ideas with idea-creator, checks each one for novelty with novelty-check, and sends the top idea to research-review for critical feedback from a reviewer model.
- Phase 1: literature survey (research-lit)
- Phase 2: idea generation, filtering and pilot experiments (idea-creator)
- Phase 3: deep novelty verification (novelty-check)
- Phase 4: external critical review (research-review)
- Phase 4.5: method refinement and experiment planning (research-refine-pipeline)
What you get at the end
The main output is idea-stage/IDEA_REPORT.md, a ranked report of ideas. The pipeline also produces refine-logs/FINAL_PROPOSAL.md and refine-logs/EXPERIMENT_PLAN.md for the top idea. Ideas that fail a pilot or turn out to be already published are eliminated and noted in the report.
Install the idea-discovery skill
Download the ZIP from this page and unzip it. Then place the folder in one of these locations:
- Personal skills: unzip into ~/.claude/skills/
- Project skills: unzip into your project's .claude/skills/ folder
- Claude apps: upload the skill under Settings → Capabilities → Skills
Install the required sibling skills first
This skill depends on several other skills. The source says the sibling skills in skills/skills-codex/* must be installed first, and that this package should be installed after them. The required sub-skills are research-lit, idea-creator, novelty-check, research-review and research-refine-pipeline.
This package is an overlay for Codex users who want Gemini as the reviewer. It also needs a local gemini-review MCP bridge to reach the Gemini reviewer.
How to trigger the skill with a natural request
The skill's description lists phrases that activate it. You can ask in plain language, for example asking for the idea discovery pipeline or for the complete idea exploration workflow. The source lists these trigger phrases: "找idea全流程", "idea discovery pipeline", and "从零开始找方向".
- "Run the idea discovery pipeline on [your topic]"
- "I want the complete idea exploration workflow for [your topic]"
- "从零开始找方向" (start from zero and find a research direction)
Set limits before you run it
The pipeline can run pilot experiments on GPUs, which uses compute time. The source sets these defaults and lets you override them in your request.
- PILOT_MAX_HOURS = 2 per pilot; longer pilots are skipped and flagged for manual pilot
- PILOT_TIMEOUT_HOURS = 3; running pilots past this are killed
- MAX_PILOT_IDEAS = 3 pilots run in parallel
- MAX_TOTAL_GPU_HOURS = 8 across all pilots
- ARXIV_DOWNLOAD = false by default, so only metadata is fetched
- AUTO_PROCEED = true by default, so checkpoints are informational and the run continues
/idea-discovery "your topic" — pilot budget: 4h per idea, 20h totalSafety notes: scripts and network access
The skill runs scripts and uses the network. It searches arXiv, Google Scholar and Semantic Scholar, and sends material to a Gemini reviewer through a local MCP bridge. Pilot experiments use GPU access. Review the limits above and the checkpoint setting before starting an unattended run, especially on shared GPU or cloud budgets.
- Set pilot budgets that match your compute allowance
- Set AUTO_PROCEED to false if you want to approve each selection
- Keep ARXIV_DOWNLOAD off unless you need the PDFs
Tips for better results
Start with a specific topic. The survey phase works from a broad direction, but the more concrete your starting point, the more useful the landscape map. If the survey misses the mark, ask the agent to adjust the scope and it will re-run the search with updated queries.
- Use AUTO_PROCEED=false if you want to steer the direction at each checkpoint
- Give pilot budgets explicitly when your compute is limited
- Treat novelty results as a starting point and verify important claims yourself
What people use it for
- A graduate student turning a broad topic into a shortlist of testable research ideas
- An ML engineer using pilot experiment results to decide which idea to pursue
- A team already using the Auto-claude-code-research-in-sleep skill set that wants a Gemini-reviewed variant
- Stress-testing a promising idea with a literature check and reviewer feedback before committing time
Frequently asked questions
Is the idea-discovery skill free?
The source material for this page does not state pricing or licence terms beyond the listed licence, which is MIT. Check the repository's licence file for the exact terms. Note that pilot experiments need GPU access, which may have its own cost.
What do I need to install before idea-discovery?
You need the sibling skills in skills/skills-codex/* installed first: research-lit, idea-creator, novelty-check, research-review and research-refine-pipeline. You also need a local gemini-review MCP bridge for the Gemini reviewer.
How do I install the idea-discovery skill?
Download the ZIP from this page and unzip it into ~/.claude/skills/ for personal use or a project's .claude/skills/ folder. In the Claude apps, you can upload it under Settings → Capabilities → Skills.
Does idea-discovery need a GPU?
Pilot experiments need GPU access. The default budget is 2 hours per pilot and 8 GPU-hours in total, and you can change these in your request.
Can idea-discovery verify that an idea is novel?
Not with certainty. The novelty check relies on literature search results and can miss prior work, so you should verify important claims yourself.
Is idea-discovery better than a simple brainstorm prompt?
It is designed for structured, multi-step work, from survey through review. For quick one-off brainstorming the pipeline is heavy, so a simpler prompt is probably a better fit.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use this skill if you are starting from a broad research topic and want a structured, reviewed shortlist of testable ideas, and you already use the Codex sibling skills and can run a Gemini bridge. Skip it for quick brainstorming, or if you cannot set limits on GPU and cloud compute.
Download the skill