How to use the Idea Discovery Pipeline agent skill

What an agent skill is and what this one makes the agent do

An agent skill is a folder with a SKILL.md file that gives an AI agent instructions for a specific workflow. When the skill is available, the agent can follow those instructions when a request matches the skill's description.

The idea-discovery skill runs a chained pipeline: literature survey (/research-lit), brainstorming (/idea-creator), novelty verification (/novelty-check), critical feedback from a second reviewer model (/research-review), and refinement of the method and experiment plan (/research-refine-pipeline). Each phase builds on the previous one's output.

• Survey the literature on the topic
• Brainstorm and rank candidate ideas
• Check whether the ideas are novel
• Get critical feedback from a second reviewer model
• Refine the best idea into a proposal and experiment plan

Install the idea-discovery skill

Download the ZIP for the skill from the directory page. Unzip it into your personal skills folder, ~/.claude/skills/, or into a project's .claude/skills/ folder if you want the skill available only for that project.

Alternatively, in the Claude apps, open Settings, go to Capabilities, and upload the skill under Skills.

The skill's files include SKILL.md. Installing the skill alone is not enough to run the full pipeline; see the requirements section below.

• Personal install: unzip into ~/.claude/skills/
• Project install: unzip into the project's .claude/skills/
• Claude apps: Settings → Capabilities → Skills, then upload

Trigger the skill with a natural request

The skill's description names the phrases that should activate it. You can say things like 'idea discovery pipeline' or '从零开始找方向' (start from zero to find a direction), or ask for 'the complete idea exploration workflow'. The Chinese trigger '找idea全流程' also activates it.

The skill takes a research direction as its argument. Give it a topic in plain language so the survey phase knows what to search for.

• 'Use the idea discovery pipeline for sparse-attention efficiency in small language models'
• '从零开始找方向: continual learning for code models'
• 'Run the complete idea exploration workflow on retrieval-augmented evaluation'

Optional settings you can override

The skill has defaults that you can change by adding an override to your request. For example, compact mode writes a shorter summary file for short-context models and session recovery. Reference-paper mode bases the ideas on a paper you provide (a local PDF path, an arXiv URL or any paper URL).

Key defaults in the source include: pilot experiments are skipped if they are estimated to take more than 2 hours per GPU, running pilots are killed after 3 hours, at most 3 ideas get pilots, and the total GPU budget is 8 hours. Checkpoints run automatically by default, and the final report is rendered to HTML.

• Compact summaries: add compact: true
• Reference paper: add a paper path or URL in the request
• Pilot limits: 2 hours per pilot, 3 hour timeout, 3 ideas, 8 GPU-hours total
• Automatic checkpoints are on by default; set AUTO_PROCEED to false to be asked at each selection point

What it needs and what it writes

The skill expects Claude Code or a compatible agent with the Skill tool and the other sub-skills installed: research-lit, idea-creator, novelty-check, research-review and research-refine-pipeline. It also needs Bash execution and web search or fetch access. For cross-model review it uses Codex MCP (mcp__codex__codex) or another configured reviewer backend; the source names gpt-6-astra as the default reviewer model. GPU access is optional and only needed for pilot experiments.

Results go into a report folder. The main outputs are idea-stage/IDEA_REPORT.md with ranked ideas, refine-logs/FINAL_PROPOSAL.md and refine-logs/EXPERIMENT_PLAN.md for the top idea. Run-state records are kept under .aris/runs/ so the workflow leaves an audit trail.

• Required: Claude Code or compatible agent with the Skill tool, plus the referenced sub-skills
• Required: Bash execution and web search/fetch access
• Needed for cross-model review: Codex MCP or another configured reviewer backend
• Optional: GPU access for pilot experiments
• Helper scripts run_state.py and idea_discovery_gate.py come from the wider repository

Safety notes: scripts and network access

This skill runs scripts and uses the network. Its allowed tools include Bash(*), and the workflow makes many tool calls and external requests, including web searches and fetches. Review the run before you rely on it, and be aware that it can take a long time.

The helper scripts run_state.py and idea_discovery_gate.py must be present for the final report to be marked complete. If they are unavailable, the source says the final report is BLOCKED.

• Bash(*) is in the allowed tools, so the agent can run shell commands
• Web search and fetch are used during the literature phase
• Reviewer calls go to the configured backend, such as Codex MCP
• If the helper scripts are missing, the final report is BLOCKED

Tips for getting useful results

Give a specific research direction rather than a single word, since the survey phase searches on it. Treat novelty and review outputs as model judgements, not verified facts, and check the cited papers yourself.

For a quick brainstorm, the full pipeline is probably more than you need, because it is long and makes many calls. Use the report files as a starting point for your own reading and planning.

• Name a concrete direction and, if you have one, a reference paper
• Read the novelty and review sections critically
• Keep the report folder as a record of how the idea was validated