| name | paper-watch |
| description | Show the newest arXiv research papers on a topic that the user has NOT seen before, by running this skill's bundled script — never from memory. The script reads progress.md (the spine) first to learn what was already shown, fetches the latest papers from arXiv, prints only the genuinely new ones, and writes progress.md last. Use it whenever the user asks what is new on arXiv, wants new papers or fresh research on a topic, wants a paper feed, or runs a scheduled "what's new since yesterday" job. Only new-since-last-time counts — showing a paper twice defeats the point of a watch. |
| allowed-tools | Bash, Read |
Paper watch
A watch is only useful if it tells you what is new. Showing the same paper
again tomorrow is noise. So this skill remembers what it has already shown — in a
file, progress.md, the loop's spine — and each run reports only papers newer
than that.
The papers are not guessable. arXiv publishes new research every day, and only
arXiv knows today's list. So the papers always come from the bundled script, run
fresh, every time — never from memory.
Getting today's new papers
Run the script. It reads progress.md first, fetches arXiv, prints only what is
new, then updates progress.md:
python3 .claude/skills/paper-watch/scripts/paperwatch.py
Pick a topic with --topic:
python3 .claude/skills/paper-watch/scripts/paperwatch.py --topic "diffusion models"
For the raw data to reason with (e.g. to write an email or a summary), ask for JSON:
python3 .claude/skills/paper-watch/scripts/paperwatch.py --json
The spine is the whole point of this project
- The script reads
progress.md first ("what have I already shown?") and
writes it last ("remember what I showed this time"). That file is the memory
that survives between runs.
- arXiv refreshes about once a day, so a daily schedule is the natural
heartbeat: run it each morning and get only that day's new papers. A fast
in-session loop would just see "nothing new" — the papers only change overnight.
- If the fetch fails, say so — never invent papers. A made-up "here's what's
new" is the one answer a watch must never give.
"Nothing new since last run" is correct, not broken
If you run it twice close together and the second time says nothing new, that is
the spine working: it remembered. To show a beginner why the spine matters,
delete it and run again — every paper comes back as new:
rm progress.md
