paper-searchlisted

# Paper Search Two chained capabilities: **find papers**, then **deep-read** any one of them. Both are grounded in real data — search hits come from public scholarly APIs, and reading reports are built only from the extracted PDF text. ``` search a topic ──▶ ranked real papers ──▶ deep-read the open-access ones ``` > **Running the scripts:** run them **by their full path from your current > working directory — do NOT `cd` into the skill folder.** That way `--save` > writes `search-results.md` into the user's workspace, where Claude can open it > as a clickable preview. (Replace `scripts/…` in the examples below with the > skill's real path, e.g. `~/.claude/skills/paper-search/scripts/search_papers.py`.) > If `python` isn't found, use `python3` (needs Python 3.8+). ## Capability 1 — Search **When:** the user wants to find papers / prior work / sources on a topic, or asks what the research says about something. 1. **Default = run immediately. Only ask if the topic is missing.** Don't gate the search behind a menu. If the message already contains a **topic** (e.g. *"查论文 玩家共鸣机制 200"* or *"find papers on X"*), search **right away** — take any params they included (count, years, sort, open-access) and sensible defaults for the rest (20 results, all years, best match). **Do not show the setup menu and do not wait.** After the results, add ONE optional line so they can still refine: *"Showing 200, all years, best match — say e.g. 'since 2021, open access