How to use the Research Idea Creator agent skill

What the idea-creator skill makes the agent do

Given a broad research direction, the skill asks the agent to brainstorm possible research ideas, check them against literature found through web search, and rank them by how publishable they look. The skill's description says it is for generating and ranking research ideas when you want to explore an area for publishable directions.

The skill describes a multi-phase workflow. Its visible SKILL.md is mostly a description of that workflow and its configuration. Parts of the full behaviour depend on other helper skills and tools that the workflow can call, including ones for novelty checks, small pilot experiments and reviews by a second AI model.

How to install the idea-creator skill

You can install the skill in one of two ways. Download the ZIP file from the skill's page on this directory and unzip it into your personal skills folder at ~/.claude/skills/, or into a project's .claude/skills/ folder if you want the skill to apply only to that project.

If you use the Claude apps instead of Claude Code, you can upload the skill under Settings → Capabilities → Skills.

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

How to trigger the skill with natural requests

You do not need a special command to start the skill. The skill's description lists phrases that should activate it, so you can ask in ordinary language. The listed triggers include 找idea, brainstorm ideas, generate research ideas, what can we work on, and any request to explore a research area for publishable directions.

The skill also accepts a research direction as an argument, shown in its hint as [research-direction]. Be specific about your area, since the broader the topic, the more general the output will be.

• brainstorm ideas on efficient fine-tuning for small language models
• what can we work on in graph neural networks for drug discovery?
• generate research ideas about robustness of vision transformers
• 找idea: 多模态检索

What you need before running it

The skill is written for Claude Code or another runtime that reads SKILL.md agent skills. Its tool list includes web search and fetch, and Bash execution. Without web search, the literature check described in the workflow cannot run as written.

Some things are optional. Second-opinion reviews can use Codex, Oracle or a manual-review MCP server. Pilot experiments need GPU compute, and the workflow also relies on the companion skills /novelty-check, /run-experiment and /monitor-experiment. If these are missing, the skill cannot run the full workflow, so expect to rely on its written analysis alone.

• Required: Claude Code or a compatible SKILL.md runtime
• Required: web search and fetch tools, and Bash execution
• Optional: Codex, Oracle or manual-review MCP servers for second opinions
• Optional: GPU compute and the companion skills /novelty-check, /run-experiment, /monitor-experiment

Safety notes: scripts, network and GPU budget

This skill runs scripts and uses the network. Its workflow can search and fetch web pages, and it can launch GPU pilot runs if the required helper tools are present. Review the skill's files before you install it, and only grant it access to the folders and accounts you are comfortable with.

The configuration sets a total GPU budget of 8 GPU-hours for all pilots combined. It also limits pilots to at most 3 ideas in parallel, and skips any pilot estimated to take more than 2 hours per GPU. Ideas outside these limits are validated on paper only. You can change these limits by passing an override in your request.

• Runs scripts: yes
• Uses network: yes
• Default total GPU budget: 8 GPU-hours
• Pilots: at most 3 ideas in parallel; pilots over 2 hours per GPU are flagged for manual pilot

Tips for getting useful ideas

Start with a focused direction and a short list of constraints, such as the kind of data or compute you have. Treat every output as a set of candidates, not findings. The workflow's novelty check is a step to run, but it can still miss work that has already been published, so verify the most promising ideas yourself.

If you want a second opinion, you can choose a reviewer backend in your request, for example with an override such as — reviewer: oracle-pro or — reviewer: manual. The default reviewer is Codex. If you want the full pipeline, the skill notes that the /idea-discovery workflow orchestrates this skill along with literature research and review.

• Name a specific sub-area rather than a whole field
• State your compute limits so pilots stay realistic
• Check the top ideas against recent papers yourself
• Use a reviewer override only if the matching tool is installed