How to use the arXiv Research Assistant agent skill

What is an agent skill, and what does the arXiv skill let the agent do?

An agent skill is a folder of instructions and, optionally, scripts that an AI agent loads when a task matches the skill's description. This one is named arxiv and is released under the MIT licence.

Once installed, the agent can search arXiv by topic, author or category, pull abstracts and full metadata, generate BibTeX citations, download PDFs, list the latest submissions in a category, and look up who cites a paper, what it references and related papers via Semantic Scholar.

• Search papers by topic, author, category or arXiv ID
• Fetch abstracts and metadata, with versioned IDs and withdrawn-paper flags
• Generate BibTeX citations
• Download PDFs to a folder you choose
• List the newest papers in a field such as cs.AI or cs.CL
• Check citations, references and related papers through Semantic Scholar

How do you install the arXiv skill?

Download the ZIP of this skill from this page and unzip it. For a personal setup, place the unzipped folder in ~/.claude/skills/. For a single project, place it in the project's .claude/skills/ folder instead.

In the Claude apps you can instead upload the skill under Settings → Capabilities → Skills.

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

What does the skill need to run?

The bundled script, scripts/arxiv.py, uses only the Python 3 standard library, so there is nothing to pip install. The skill needs internet access to export.arxiv.org and api.semanticscholar.org. No API keys are required.

• Python 3 (standard library only, no pip installs)
• Internet access to export.arxiv.org and api.semanticscholar.org
• No API keys

How do you trigger the skill with natural requests?

You do not need to name the skill. Ask for paper work in plain language and the agent should pick it up when your request matches the skill's description. The description lists triggers such as mentions of arXiv, an arXiv ID like 2601.02780 or hep-th/0601001, an arxiv.org URL, paper search, literature review, cite this paper, or what's new in cs.LG.

Be specific about the topic, author, category or paper ID. The more concrete the request, the more precise the search.

• Find papers about GRPO reinforcement learning from the last year
• Give me a BibTeX entry for arXiv 1706.03762
• What's new in cs.CL today?
• Who cites 2601.02780, and what does it build on?
• Find related work to this paper: https://arxiv.org/abs/2601.02780

What commands does the script run?

The skill prefers scripts/arxiv.py over raw curl because the script parses the Atom XML, handles versioned IDs and withdrawn papers, and gives readable output. Add --json to any command for machine-readable output. The common flags are --max N, --sort relevance|date|updated, and --start N for pagination in search.

• search: search papers by topic, author or category
• get: full metadata and abstract for one or more IDs
• bibtex: BibTeX citation
• download: save a PDF to a folder
• new: latest papers in a category
• cites, refs, similar: Semantic Scholar links

What are the safety and privacy notes?

This skill runs a bundled script and uses the network. It sends your search terms and paper IDs to public external services, arXiv and Semantic Scholar. Do not use it for tasks that involve private data.

Withdrawn papers are flagged with [WITHDRAWN]. Do not cite them without noting that status. The skill also does not verify that a paper's claims are correct or reproducible.

• Search terms and IDs go to public external services
• Runs scripts and uses network access
• Withdrawn papers are marked [WITHDRAWN]
• Retrieval is not quality or reproducibility checking

What tips make the results more useful?

For a literature review, search twice: once sorted by date for recent work, and once by relevance for well-known papers. Then check the top candidates with get and cites, and generate BibTeX only for the papers you keep.

Space out repeated calls. arXiv asks for roughly one request every three seconds and Semantic Scholar roughly one per second without a key. The script exits with a clear message on HTTP 429. Old and new ID formats both work everywhere, and bibtex and download keep the version suffix so your citation matches the text you read.

• Search by date and by relevance for a full picture
• Pin versions like ...v2 when citing, since later versions can change
• Use the category taxonomy at https://arxiv.org/category_taxonomy to find category codes