Skip to content
Get packs

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 Pipeline is an agent skill that runs a multi-step research workflow, starting from a broad research direction and ending with ranked, reviewed research ideas and a refined proposal and experiment plan for the top idea. It is aimed at researchers and PhD students starting a project, and at ML teams who want an automated first pass before spending GPU time. After this guide you will know what the skill does, how to install it, how to trigger it with a natural request, and what it needs and risks before you run it.

Idea Discovery Pipelinewanshuiyin/Auto-claude-code-research-in-sleep · MIT

Quick start

  1. 1Download and unzip. Download the skill ZIP from this page and unzip it into ~/.claude/skills/ or the project's .claude/skills/ folder.
  2. 2Or upload in the Claude apps. In the Claude apps, open Settings → Capabilities → Skills and upload the skill.
  3. 3Check the requirements. Confirm the sub-skills, the helper scripts, reviewer model access, and network access are available.
  4. 4Make a request. Ask for the idea discovery pipeline or the complete idea exploration workflow and state your research direction.
  5. 5Review checkpoints. Read the reported selections at each checkpoint; set AUTO_PROCEED to false if you want to approve them.
  6. 6Check the report. Open idea-stage/IDEA_REPORT.md and the refine-logs files, and apply your own expert judgement.

What is an agent skill and what does this one do?

An agent skill is a folder containing a SKILL.md file that tells a coding agent how to carry out a particular kind of task. When the skill is installed and your request matches its description, the agent follows its instructions.

This skill chains sub-skills into one pipeline: literature survey (research-lit), brainstorming (idea-creator), novelty verification (novelty-check), external critical review (research-review), and method refinement with experiment planning (research-refine-pipeline). Each phase builds on the previous phase's output.

  • Final report: idea-stage/IDEA_REPORT.md with ranked ideas.
  • Refined proposal for the top idea: refine-logs/FINAL_PROPOSAL.md.
  • Experiment plan for the top idea: refine-logs/EXPERIMENT_PLAN.md.
  • Optional small pilot experiments on GPUs, within time and budget limits.
  • Optional HTML rendering of the final report.

Install the Idea Discovery skill

Download the ZIP for this skill from this page and unzip it. For a personal install, unzip it into ~/.claude/skills/. For a single project, unzip it into the project's .claude/skills/ folder.

If you use the Claude apps instead, open Settings, go to Capabilities, then Skills, and upload the skill there.

  • Personal install: ~/.claude/skills/
  • Project install: .claude/skills/ inside your project folder
  • Claude apps: Settings → Capabilities → Skills, then upload

Trigger the skill with a natural request

The skill's description lists trigger phrases. You can use them directly, or ask for the complete idea exploration workflow in your own words. Put your research direction in the request so the agent knows the topic.

The source also describes overrides you can add to the request, such as a reference paper or compact mode. Defaults apply when you do not specify them.

  • "idea discovery pipeline: sparse autoencoders for interpretability"
  • "找idea全流程" followed by your topic
  • "从零开始找方向" followed by your topic
  • "Run the complete idea exploration workflow on [your direction]"
  • Override example: "idea discovery pipeline on [topic] — compact: true"

What you need before running it

The skill depends on other sub-skills and helper scripts. It was written for Codex or a compatible agent environment, and it needs access to an OpenAI reviewer model through a secondary Codex agent. The default reviewer model is gpt-6-astra.

Pilot experiments need GPU compute, and literature search needs network access.

  • Sub-skills installed: research-lit, idea-creator, novelty-check, research-review, research-refine-pipeline
  • Helper scripts run_state.py and idea_discovery_gate.py from the wider repository
  • Access to an OpenAI reviewer model via a secondary Codex agent (default gpt-6-astra)
  • GPU compute for optional pilot experiments
  • Network access for literature search and optional arXiv downloads

Safety notes: scripts and network access

This skill runs scripts and uses the network. Literature search uses network access, and downloading arXiv PDFs happens only if the ARXIV_DOWNLOAD setting is true; by default it only fetches metadata.

Pilot experiments consume GPU hours. Per the source, a pilot can take up to hours each, so set budgets before you start. The source sets a per-pilot limit, a timeout, a cap on parallel pilots and a total GPU budget.

  • PILOT_MAX_HOURS = 2: pilots estimated above 2 hours per GPU are skipped and flagged for manual pilot.
  • PILOT_TIMEOUT_HOURS = 3: running pilots are killed after 3 hours, with partial results collected if available.
  • MAX_PILOT_IDEAS = 3: at most 3 top ideas are piloted in parallel.
  • MAX_TOTAL_GPU_HOURS = 8: remaining pilots are skipped once the total GPU budget is exceeded.
  • ARXIV_DOWNLOAD = false by default.

Tips for getting useful results

Give a specific direction rather than a single word, and read the report's ranked ideas and review notes before you commit compute. The output still needs expert judgement.

AUTO_PROCEED = true (the default) means checkpoints are informational and the run continues in the same turn. Set it to false if you want to approve each selection checkpoint yourself, which makes the run stop and wait for your reply.

  • Start with a narrow direction and, if you have one, pass a reference paper.
  • Use compact: true when your session has a short context window.
  • Set AUTO_PROCEED to false if you want to choose the idea at each checkpoint.
  • Check the run record and the evidence gate result before trusting the final report.

What people use it for

  • Turning a broad topic such as sparse autoencoders for interpretability into a shortlist of ranked ideas.
  • Checking whether candidate ideas have already been published before committing to one.
  • Drafting a proposal and experiment plan for the top idea.
  • Running small pilots on GPUs for up to three top ideas, within the set budget.
  • Producing an HTML version of the final report for sharing.

Frequently asked questions

Is the Idea Discovery Pipeline skill free?

The skill is published under the MIT licence, as stated on this page. Running it may still cost you GPU compute for pilot experiments and access to the reviewer model.

How do I install the idea-discovery agent skill?

Download the ZIP from this page and unzip it into ~/.claude/skills/ for a personal install or into a project's .claude/skills/ folder. In the Claude apps, you can instead upload it under Settings → Capabilities → Skills.

What do I need to run the idea discovery pipeline?

You need Codex or a compatible agent environment, the sub-skills research-lit, idea-creator, novelty-check, research-review and research-refine-pipeline, the helper scripts run_state.py and idea_discovery_gate.py, and access to an OpenAI reviewer model. GPU compute is needed only for pilots, and network access is needed for literature search.

Does the skill run experiments automatically?

It can run small pilot experiments on GPUs within time and budget limits. Pilots estimated above 2 hours per GPU are skipped and flagged for manual pilot, and running pilots are stopped after 3 hours.

Why does the skill need an OpenAI reviewer model?

The source says the external critical review uses a secondary Codex agent with a reviewer model, with a default of gpt-6-astra. The workflow depends on this reviewer for its review phases.

Can I use this skill without a GPU?

The source says GPU compute is needed for optional pilot experiments. Without GPUs you can still use the survey, brainstorming, novelty and review phases, but the source does not describe a non-GPU pilot option.

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 a research project and want a structured path from a broad topic to shortlisted, reviewed ideas, or if your ML team wants an automated first pass before spending GPU time. Pick something else if you need a hands-off answer you will not check, or if you cannot run pilot experiments on your own compute.

Download the skill