Research Digest

research-digest

This skill generates a weekly digest of academic papers relevant to your research interests and delivers it via Telegram.

When to invoke

• When the user asks for a research digest or paper summary
• When triggered by a cron schedule (e.g., "every Monday 8am")
• When the user wants to know what's new in their field

Steps to execute

1. Read research interests from memory

   • Check MEMORY.md for the user's recorded research areas, current projects, and relevant keywords
   • If no interests are recorded, ask the user to describe their research focus before continuing

2. Search for recent papers

   • Use web_search to search arXiv for papers published in the last 7 days matching their interests
   • Search query pattern: site:arxiv.org [research area] [current year-month]
   • Run 2–3 searches with different keyword combinations to maximize coverage
   • Also search for papers via Google Scholar or Semantic Scholar if arXiv yields < 3 results

3. Fetch and parse paper abstracts

   • Use web_extract on each promising result to get the full abstract and metadata
   • Extract: title, authors, institution, date, abstract, link
   • Focus on papers that are directly relevant — filter out tangential results

4. Select top 5 papers

   • Rank by relevance to the user's stated interests
   • Prefer papers with strong empirical results or novel methodology
   • Include at least 1 paper that might be a "surprising finding" outside the user's usual scope

5. Synthesize the digest

   • For each paper: 2-sentence summary + 2–3 bullet points of key contributions
   • Write in plain language, avoiding unnecessary jargon
   • Note any papers that directly build on or contradict the user's current projects (check MEMORY.md)

6. Deliver via Telegram

   • Format for readability in Telegram (use bold headers, line breaks)
   • Send as: "Research Digest — Week of [date]" followed by numbered entries
   • Include arXiv links for each paper
   • End with: "Reply to discuss any of these papers or ask for a deeper dive"

7. Update memory

   • Save the digest summary (titles + key points only, not full text) to MEMORY.md under "Recent digests"
   • This enables future sessions to reference what was covered

Output format

Research Digest — Week of [date]

1. [Paper Title] — [authors], [institution]
[2-sentence summary]
• [Key contribution 1]
• [Key contribution 2]
• [Key contribution 3]
arXiv: [link]

2. [Paper Title] ...

Error handling

• If Telegram delivery fails: save digest to a file at /sandbox/digest-[date].md and notify via CLI
• If fewer than 3 relevant papers found: widen the search timeframe to 14 days and retry
• If arXiv is unreachable: try Semantic Scholar as fallback

Notes

• The quality of this skill improves as the user records more research interests in memory
• After 5+ successful runs, Hermes will auto-optimize the search queries and summarization style for this user's preferences