An upright machined rack of narrow slots holding stacks of dark glass leaves, with an overhead rail whose carriage lifts one leaf out of a slot; a diagonal cut turns the rest of the rack into bone-grey outline drawing, and a cyan line runs from the lifted leaf up through the carriage, along the rail and through a pair of rollers.

Research

ArXiv

Search and retrieve arXiv academic papers by topic, category, or paper ID — with AlphaXiv-enriched AI-generated overviews.

Workflows

Tools

References

Triggers

The Problem

Ask a generic AI to find you the latest research on a topic and you get a vague summary that might be months out of date, with no paper IDs, no actual abstracts, and a real chance the citations are hallucinated. Academic research moves fast. If you're trying to track what's happening in LLMs, AI safety, or security, you need the actual papers — real IDs, real dates, real authors — not a prose summary that sounds confident but can't be verified.

How This Skill Approaches It

Hit the arXiv Atom API directly, no API key needed, with category-specific queries across cs.AI, cs.LG, cs.CL, cs.CR, cs.MA, cs.SE, and cs.IR. Three workflows cover the main use cases: Latest pulls the most recently updated papers in a category sorted by lastUpdatedDate; Search runs fielded queries against title, abstract, and author fields with boolean operators; Paper fetches a specific paper by ID and enriches it with an AlphaXiv AI-generated overview when one exists. The AlphaXiv enrichment layer gives you a readable markdown summary of the paper without needing to parse the PDF — useful for quick triage before you commit to reading the full thing.

Uses arXiv Atom API across cs.AI/cs.LG/cs.CL/cs.CR/cs.MA/cs.SE/cs.IR
Three intents: Latest, Search, Paper

Not for general research (Research), URL parsing, or annual reports

In Action

What you say to your DA, and what the ArXiv skill actually does.

You say "what are the newest papers on multi-agent systems this week"

Runs the Latest workflow against cs.MA sorted by lastUpdatedDate, returns real paper IDs, titles, authors, and publish dates — no hallucinated citations.
You say "find arxiv papers about LLM prompt injection vulnerabilities"

Runs the Search workflow with a fielded query across cs.CR and cs.AI abstracts, then enriches top results with AlphaXiv overviews where available.
You say "explain this paper to me — arxiv.org/abs/2401.12345"

Runs the Paper workflow: fetches the Atom entry for the ID, pulls the AlphaXiv overview markdown if it exists, and returns a readable summary with the original abstract as a fallback.

Inside the Skill

The thinking, frameworks, and architecture that distinguish this skill from a generic version of the same task.

Ideal State

Every paper ID, title, author list and date matches arXiv, and newness is judged by <published>, so an edited old paper is never shown as a new submission.
An AlphaXiv overview is used only when its fetch returned 200; otherwise the answer says it worked from the abstract or full text.
Every result links to arXiv and says in one line why it matters to the principal's interests (the focus areas in <configRoot>/LIFEOS/USER/CUSTOMIZATIONS/SKILLS/ArXiv/PREFERENCES.md; with no file, judge relevance from the request), most relevant first.

Intent Routing

Intent	Trigger	File
Latest	"latest papers in X", "new papers on X", "what's new in AI research"	`Intents/Latest.md`
Search	"search arxiv for X", "find papers about X", "arxiv papers on X"	`Intents/Search.md`
Paper	arxiv URL, paper ID like `2401.12345`, "explain this paper"	`Intents/Paper.md`

Quick Reference

Shared by all three intents. Each intent's exact query is in its own Tool contract.

arXiv API (no auth):

Base: https://export.arxiv.org/api/query
Search fields: ti: (title), au: (author), abs: (abstract), cat: (category), all: (everything)
Booleans: AND, OR, ANDNOT
Sort: sortBy=submittedDate&sortOrder=descending returns new submissions, newest first
Pagination: start=0&max_results=10

AlphaXiv overview (no auth):

curl -s "https://alphaxiv.org/overview/{PAPER_ID}.md"
A 200 returns a structured markdown analysis optimized for language models. Not all papers have overviews: a 404 means the analysis is not yet generated.
AlphaXiv uses the latest version, so the lookup takes the bare numeric ID (e.g., 2603.12345, never 2603.12345v2).

Categories (topic words map a user's topic to a category; an ambiguous topic uses several joined with OR):

Category	Name	Topic words
`cs.AI`	Artificial Intelligence	AI, artificial intelligence
`cs.LG`	Machine Learning	machine learning, ML, deep learning
`cs.CL`	Computation and Language (NLP/LLMs)	LLMs, NLP, language models
`cs.CR`	Cryptography and Security	security, cybersecurity
`cs.MA`	Multi-Agent Systems	agents, multi-agent: query `cs.MA+OR+cs.AI`
`cs.SE`	Software Engineering	software engineering
`cs.IR`	Information Retrieval	information retrieval, RAG
`cs.RO`	Robotics	robotics
`cs.CV`	Computer Vision	computer vision

The principal's priority categories are listed in PREFERENCES.md (see Data & Config); with no file, use the categories the request names.

Data & Config

User customizations live at <configRoot>/LIFEOS/USER/CUSTOMIZATIONS/SKILLS/ArXiv/ (configRoot is $CLAUDE_CONFIG_DIR when set, else ~/.claude). When that directory exists, its PREFERENCES.md, configurations and resources override the defaults; when it is absent, the defaults stand.

Gotchas

arXiv API requires HTTPS and -L (follows redirects). HTTP 301s to HTTPS silently.
The arXiv API returns Atom XML, not JSON. Parse with text processing, not jq.
lastUpdatedDate includes edits to old papers. For truly new submissions, check <published> dates.
AlphaXiv overviews are AI-generated summaries. Great for quick understanding, but verify claims against the actual paper for anything you'd cite.
arXiv API rate limit is 3 seconds between calls. Batch your queries.
max_results caps at 2000. For broader sweeps, paginate with start.
Category search (cat:cs.AI) returns papers with that as primary OR cross-listed category.

Examples

Example 1: Latest papers in a category

User: "what's new in AI safety papers this week"
→ Latest intent
→ This week's new cs.AI submissions by <published> date: titles, authors, abstracts, links, and a Papers worth reading pick

Example 2: Topic search

User: "search arxiv for prompt injection defenses"
→ Search intent (query all:"prompt injection" with boolean refinement)
→ Ranked recent matches with abstracts, plus related searches

Example 3: Single paper lookup

User: "explain this paper: 2401.12345"
→ Paper intent
→ Summary from the AlphaXiv overview (or the abstract when it 404s), key contributions, a worth-reading verdict, and the PDF link

Workflows & Routing · 3

Each workflow is one job the skill runs. The trigger phrases route your request to the right one — this is the skill's routing table.

01

Latest Intents/Latest.md

latest papers in X, new papers on X, whats new in AI research
02

Search Intents/Search.md

search arxiv for X, find papers about X, arxiv papers on X
03

Paper Intents/Paper.md

arxiv URL, paper ID like 2401.12345, explain this paper