geoff-ji/jiharn/arxiv

arxiv

Overview

Scan arxiv for new papers matching the user's request and produce a compact digest. Favour titles, authors, and a one-sentence takeaway per paper over copy-pasting abstracts wholesale.

Steps

1. Decide the scope: a category (e.g. cs.AI), a specific arxiv id, or a

keyword query. Default to cs.AI for the last 24 hours when the user is vague.

1. Fetch the list via read_url against the appropriate

https://arxiv.org/list/<category>/new or https://arxiv.org/abs/<id> URL. For keyword queries prefer https://export.arxiv.org/api/query?search_query=....

1. Parse the titles / authors / abstracts. Skip cross-listings the user

already implicitly excluded.

1. Produce 3–5 short bullets per paper: title, primary authors, one-line

plain-language takeaway. No markdown headings per paper — the chat surface is a messenger, not a document.

1. When nothing useful shipped in the requested window, say so plainly

rather than padding the reply.

Output format

• Line 1: one-line summary of the scope and count (e.g. "cs.AI, 3 new

papers in the last 24 hours").

• One bullet per paper, newest first.
• End with a short "note:" line only when there is a non-obvious caveat

(rate-limit hit, partial results, timezone mismatch).

Pitfalls

• The arxiv HTML listing is long; ask read_url to truncate at the

listing section rather than loading every byte.

• The arxiv API occasionally returns 503s — retry once, then report

plainly instead of inventing content.

• Do not speculate on experimental claims the abstract does not make.