How to use the Research Idea Creator agent skill (idea-creator)

What the idea-creator skill makes the agent do

An agent skill is a folder of instructions (a SKILL.md file) that tells an AI agent how to carry out a specific task. This one covers research ideation. Given a broad area, the agent surveys recent papers using web search and any local PDFs you have, then asks a separate Gemini model, through a local MCP bridge, to brainstorm 8-12 concrete research ideas.

Each idea comes with a hypothesis, a cheap test, a risk level and an effort estimate. The agent then groups near-duplicate ideas, drops only those that are clearly over budget, and attaches notes on related prior work for later review.

• Surveys the landscape: local PDFs in papers/ and literature/, then web search over recent venues and arXiv.
• Generates 8-12 ideas through the gemini-review MCP bridge.
• Annotates ideas with prior-work and impact notes rather than discarding them on taste.
• Writes output to an idea-stage/ directory.

Install the idea-creator skill

Download the skill ZIP from this page and unzip it. You can place the skill folder in one of two locations: your personal skills folder at ~/.claude/skills/, or a project's .claude/skills/ folder if you want it only for one repository.

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

This skill is written as an override for Codex users. Install the companion package skills/skills-codex first, then install idea-creator after it.

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

Trigger the skill with natural requests

You do not need a special command. The skill's description lists phrases that should activate it, so ask in plain language. The source lists the trigger phrases "找idea", "brainstorm ideas", "generate research ideas", "what can we work on", and requests to explore a research area for publishable directions.

Give a specific broad direction rather than a single word. The skill passes your direction into its brainstorming prompt, so a clearer area produces more relevant ideas.

• "Brainstorm research ideas on sparse attention for long-context language models."
• "What can we work on in graph-based fraud detection?"
• "Generate research ideas on test-time adaptation for vision models."

Understand the workflow phases

The skill runs in phases. Phase 1 maps the field: it scans papers/ and literature/ first, then runs at least five different web search queries over recent top-venue and arXiv papers, reading the abstracts and introductions of the top 10-15 results. It groups papers by approach and flags gaps.

Phase 2 sends the landscape and gaps to the Gemini reviewer, which returns the brainstorm. Phase 3 is mechanical: it clusters near-duplicates and drops only ideas that are objectively out of budget, such as needing more than one week of available GPU time or a provably unavailable dataset. Phase 4 runs deeper validation on the survivors, including novelty checks and critical review through the Gemini bridge.

• Phase 1: landscape survey (about 5-10 minutes according to the source)
• Phase 2: idea generation through gemini-review
• Phase 3: duplicate clustering and budget-only feasibility gate
• Phase 4: novelty checks and critical review of top ideas

Set up the Gemini reviewer and what the skill needs

Phase 2 depends on the gemini-review MCP tools. The skill will not run as written without a working Gemini MCP bridge. The reviewer model is named gemini-review; if you need a specific Gemini model, set the GEMINI_REVIEW_MODEL environment variable.

The skill also expects web search access for the landscape survey. GPU compute is optional and only needed for pilot experiments, which the source mentions in terms of 8x RTX 3090 or less. Pilots refer to other sub-skills, so they need separate setup.

• Gemini reviewer through the local gemini-review MCP bridge (optional GEMINI_REVIEW_MODEL override)
• Web search access for the landscape survey
• Optional GPU compute for pilot experiments
• Companion skills from skills/skills-codex, installed first
• Referenced skills: /novelty-check, /research-lit, /run-experiment and /monitor-experiment

Safety notes: scripts and network

The skill does not run scripts. It does use the network: it performs web searches for papers and sends prompts to the Gemini reviewer through the MCP bridge. Review what you send, since your research direction and landscape notes are included in those prompts.

Pilot experiments are capped in the skill's constants: pilots longer than 2 hours per GPU are skipped and flagged for manual pilot, hard-stopped at 3 hours, limited to 3 in parallel, and limited to 8 total GPU hours. You can override these by argument if you understand the cost.

• Runs scripts: no
• Uses network: yes (web search and the Gemini bridge)
• Pilot limits are set in the skill's constants and can be overridden by argument

Tips for better results

Point the skill at a focused area and mention the papers you already know, since the local library scan reads your existing PDFs first. Treat the output as a structured starting list, not a verdict. The skill states that novelty and impact are judged later by a reviewer and search step, so ideas may overlap existing work.

Use the ideas as input to a deeper pass. The source suggests running /idea-discovery for the full pipeline, which orchestrates this skill.

• Put PDFs you already have into papers/ or literature/ before asking.
• Name a concrete area and, if you have one, a compute budget.
• Check each idea's prior_work note before investing time.
• Run the novelty check before committing to an idea.