geoff-ji/jiharn/daily-paper-digest

daily-paper-digest

Trigger

Use this skill when the user asks for a one-time or scheduled digest of recent papers in some topic (e.g. "每天 8 点给我发 AI Agent 最新论文", "每周一总结机器学习论文"). When the schedule fires through cron, this skill is loaded automatically via skill_hint.

Do NOT use this skill for one-off "查一下某篇论文" questions — that path goes through web_search / read_url directly without scheduling.

Inputs

• topic: subject area or keywords (e.g. "AI Agent / LLM"). When the

cron instruction or chat does not name a topic, fall back to the most recent topic the user mentioned in memory; if still empty, ask once instead of inventing a topic.

• window: time window — default "last 24h" for daily, "last 7d" for

weekly digests.

• max_results: default 5, hard cap 8.
• sources: arXiv first, then web_search for venue / press. Prefer

arxiv.org and well-known preprint mirrors.

Steps

1. Resolve topic. If missing and not recoverable, send a single short

ask back instead of running a search.

1. Pull candidate papers. Reuse the arxiv skill as the primary source;

fall back to web_search + read_url for non-arxiv venues.

1. Rank by relevance to topic, then recency. Drop duplicates by arxiv

id or canonical URL.

1. For each paper produce:

• title
• primary authors (max 3, then et al.)
• year
• source URL
• one-line plain-language takeaway
• optional "为什么值得看" line when the paper is genuinely notable

1. Output as a compact IM message — short bullets, newest first, no

markdown headings per paper.

Verification

• Every paper has a real source URL.
• Title, authors, year are never invented.
• If the result set is empty or thin, say so plainly with the count

rather than padding with low-confidence hits.

• For scheduled runs, only emit [SILENT] when the topic itself is

unresolved and the user has not yet provided one.

Failure Signals

• topic missing and not recoverable from memory or chat.
• All candidate sources return errors.
• Same paper repeats across consecutive runs (dedupe is failing).
• Output reads like an abstract dump rather than a takeaway summary.