arXiv Research Skill

Core Directives

1. Never scrape arxiv.org directly. Always use the arxic-mcp tools.
2. Token Efficiency FIRST. Never use arxiv_get_paper_details or pull full abstracts as your first step unless the user explicitly provided a paper ID. Start broad.
3. Rate Limits. The tools enforce a mandatory 3-second delay. Do not retry if a call is taking slightly longer than expected.

Tool Conditionals

IF the user asks to search for a topic, keyword, or author (e.g., "papers on quantum gravity", "find papers by LeCun"): THEN use arxiv_search with parameters query and max_results to retrieve a list of IDs. AND THEN present the titles and IDs to the user, pausing for them to select an ID for deeper analysis.

IF the user asks for the newest or most trending papers in a general field without a specific keyword (e.g., "what's new in AI?"): THEN use arxiv_search_trending_papers with a standard category prefix (e.g., cs.AI).

IF the user provides a paper_id and asks for a quick overview or TL;DR: THEN use arxiv_get_paper_summary to get an AI-optimized synopsis.

IF the user provides a paper_id and asks for full, deep-dive analysis of the methodology and results: THEN use arxiv_get_paper_details to ingest the complete, unabridged abstract and metadata into context.

IF the user provides a paper_id and asks for more papers like it, or to go down a "rabbit hole": THEN use arxiv_search_related_papers to automatically cross-reference the primary category and find adjacent research.

IF the user asks you to cite a paper or generate a bibliography: THEN use arxiv_get_paper_citations passing the paper_id and optional format (bibtex, apa, mla). Display the output in a codeblock.

IF the user explicitly asks to "download the PDF" or "save the paper to my machine": THEN use arxiv_download_paper passing the paper_id and an absolute path for destination_dir.