Skip to content
Get packs

Guide · Agent skill

How to use the Gemini Literature Search agent skill (gemini-search)

Updated · written with AI from the project's own documentation and checked against our catalogue data

The gemini-search skill is a Claude agent skill that asks Google's Gemini model to find research papers on a topic, using a broader set of search angles than a plain keyword query. This guide explains what the skill does, how to install it, how to trigger it with natural requests, and what to check before you cite anything it returns. By the end you will be able to run a search, read the results table, and know which entries need checking at the source.

Gemini Literature Searchwanshuiyin/Auto-claude-code-research-in-sleep · MIT

Quick start

  1. 1Install the prerequisites. Install Node.js v16.0.0 or later, then install @google/gemini-cli and authenticate it with gemini auth.
  2. 2Set up the MCP bridge. Install gemini-mcp-tool and add the gemini-cli server to your Claude Code or Claude Desktop configuration.
  3. 3Provide credentials. Log in with your Google account or export GEMINI_API_KEY in the shell before starting Claude Code.
  4. 4Install the skill. Unzip the gemini-search folder into ~/.claude/skills/ or a project's .claude/skills/, or upload it under Settings, Capabilities, Skills.
  5. 5Run a search. Ask in plain language, for example "gemini papers on topic X", optionally adding an override such as year or code-only.
  6. 6Verify each result. Open the DOI or arXiv page for every paper and confirm the details before citing it.

What the gemini-search skill makes the agent do

When you ask for papers on a topic, the agent sends a detailed search request to Gemini through a bridge tool. Gemini breaks the topic into sub-problems, alternative names and benchmark or setting variants, then returns a numbered list of candidate papers with title, authors, year, venue, arXiv ID, DOI and code link where available.

The skill is meant to surface papers that keyword searches can miss. Its answers come from a language model, however, and are not checked against a database, so treat every result as a lead to verify.

  • Default target: up to 15 papers per search (MAX_RESULTS = 15)
  • Default minimum publication year: 2022, overridable with a year filter
  • Results are shown as a table with venue, year, code link and a one-sentence summary, plus arXiv ID and DOI where available

Install the gemini-search skill

Download the ZIP for skills/gemini-search from the directory page. The skill is a single file, SKILL.md. Install it in one of two ways.

Option 1: unzip into a skills folder. Put the unzipped gemini-search folder in ~/.claude/skills/ for personal use across projects, or in a project's .claude/skills/ folder to share it with one codebase.

Option 2: upload it in the Claude apps. Open Settings, go to Capabilities, then Skills, and upload the skill there.

Install and configure the prerequisites

The skill cannot run without a working Gemini setup. You need Node.js v16.0.0 or later, the Google Gemini CLI (@google/gemini-cli) installed and authenticated, and the gemini-mcp-tool MCP bridge configured in Claude Code or Claude Desktop. The source lists these install steps, which you run in a terminal:

  • Authenticate with a Google account login or set a GEMINI_API_KEY environment variable
  • Add the gemini-cli server to your MCP configuration: in ~/.claude.json, or in %APPDATA%\Claude\claude_desktop_config.json for Claude Desktop
  • Set the API key in the same shell before you start Claude Code, so the MCP process inherits it. A key from Google AI Studio is free to obtain.
npm install -g @google/gemini-cli gemini auth npm install -g gemini-mcp-tool claude mcp add gemini-cli -- npx -y gemini-mcp-tool export GEMINI_API_KEY='your-key-here' claude

Trigger the skill with natural requests

The skill activates when you use phrases such as "gemini search", "gemini papers" or "search with gemini", or when you ask for AI-powered literature discovery beyond arXiv and Semantic Scholar. You can also call it directly with a query.

Overrides go after the topic. The source lists these options:

  • max: request up to 20 papers
  • year: 2020- for papers from 2020 onward
  • code-only: only papers with open-source code
  • venues: a comma-separated list such as NeurIPS,ICML,ICLR
  • model: for example gemini-3-flash-preview or auto-gemini-3
/gemini-search "topic" — max: 20 /gemini-search "topic" — year: 2020- /gemini-search "topic" — code-only /gemini-search "topic" — venues: NeurIPS,ICML,ICLR

Example requests

Plain requests work well. State the topic, and add a filter if you need one. Here are some phrasings that match the skill's trigger wording:

  • "Gemini search for papers on retrieval-augmented generation evaluation, code only"
  • "Search with Gemini for surveys on graph neural networks for molecules, from 2021 onward"
  • "Gemini papers on speculative decoding, venues NeurIPS and ICML"

Check and use the results

Results arrive as a table. The skill also lists arXiv IDs for cross-reference with an arXiv search and DOIs as the canonical link for published work. Gemini can invent or misattribute titles, authors, years or DOIs, so open each paper at its source before adding it to a bibliography.

If Gemini returns fewer papers than requested, the skill notes this and does not re-query.

  • Open the DOI or arXiv page and confirm title, authors, year and venue
  • Follow the code link and check that it is the official repository
  • Use a second source such as arXiv or Semantic Scholar to cross-check

Safety notes: scripts and network access

This skill runs scripts and uses the network. Its allowed tools include Bash, Read, Write and the Gemini MCP tools, and its search request is sent to Google's Gemini service. Your query text leaves your machine as part of that request.

The skill includes a sandbox option on one Gemini tool for safe code execution. Review the skill's SKILL.md before installing, and avoid putting confidential material in queries.

  • Do not paste unpublished or confidential text into search queries
  • Review SKILL.md, the only file in the package, before you install it

What people use it for

  • Starting a literature review and covering topic variants and alternative names
  • Mapping an unfamiliar field by looking for surveys and related work
  • Adding a second discovery source alongside arXiv or Semantic Scholar
  • Finding papers that have open-source code for a method you plan to test

Frequently asked questions

Is the gemini-search skill free?

The skill's source is published under the MIT licence. Gemini access depends on your setup: you can log in with a Google account or use a GEMINI_API_KEY, and the source notes that a key from Google AI Studio is free to obtain. Check current Google terms for any usage limits.

What do I need to run gemini-search?

You need Node.js v16.0.0 or later, the Google Gemini CLI installed and authenticated, and the gemini-mcp-tool MCP bridge configured in Claude Code or Claude Desktop. You also need a Google account login or a GEMINI_API_KEY. The skill cannot run without this setup.

How do I install the gemini-search skill?

Download the ZIP from the directory page and unzip it into ~/.claude/skills/ for personal use or into a project's .claude/skills/ folder. You can also upload it in the Claude apps under Settings, Capabilities, Skills.

Can Gemini's paper results be trusted for citations?

Not without checking. Gemini answers are generated by a language model and are not checked against a database, so titles, authors, years or DOIs can be wrong. Verify each paper at its DOI or arXiv page before citing it.

How does gemini-search compare with arXiv or Semantic Scholar search?

The skill is designed for broad discovery beyond keyword matching, and it returns sub-topics and naming variants. arXiv and Semantic Scholar index real records, so they are better for exact lookup. Using both gives you wider coverage and a way to cross-check.

Why does the skill say my Gemini model was not used or the search failed?

The source notes that an explicit gemini-3-pro-preview can be silently downgraded to gemini-2.5-pro on some accounts when capacity runs out. It also says that an API key exported in a separate tool shell does not reach an already running host, so set it before starting Claude Code. If the MCP tool fails, the skill falls back to the Gemini CLI.

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 gemini-search if you already have Gemini CLI and gemini-mcp-tool working and want an AI-driven discovery source alongside arXiv or Semantic Scholar. Skip it if you need final, citable bibliographies without manual checking, or if you have no Google account or Gemini API key.

Download the skill