papers-research Skill
当用户想搜索学术论文、阅读研究文献、查找引用、下载 arXiv PDF 或执行任何文献综述类任务时使用。独立的 Python 工具包,无需 MCP 服务器。
安装方式:把技能目录放入 ~/.claude/skills/(Claude Code)或在 claude.ai 设置中启用;也可复制右侧安装命令一键添加。
技能指令原文(SKILL.md)
Papers Research
Skill-mode port of papers-mcp.
Orchestrates a bundled Python CLI (scripts/papers.py) via Bash to search
papers, fetch metadata/citations, and download + read PDFs.
One-time setup
Verify dependencies (only once per machine):
python -c "import httpx, arxiv, fitz" 2>&1 || pip install httpx arxiv PyMuPDF
If python isn't on PATH, use py (Windows) or the full interpreter path.
Invocation
The script lives at ${CLAUDE_PLUGIN_ROOT}/skills/papers-research/scripts/papers.py
— this variable is auto-substituted by Claude Code's plugin loader. Always
quote the path (it may contain spaces).
python "${CLAUDE_PLUGIN_ROOT}/skills/papers-research/scripts/papers.py" <subcommand> [args]
Subcommands
| Subcommand | What it does | Example |
|---|---|---|
| search [--limit N] | Semantic Scholar search (cap 20) | search "diffusion models" --limit 5 |
| detail | Full metadata, TL;DR, top 10 references | detail 10.48550/arXiv.2310.06825 |
| citations [--limit N] | Papers that cite this one (cap 20) | citations --limit 15 |
| arxiv [--max-results N] | arXiv preprint search (cap 10) | arxiv "RLHF" --max-results 5 |
| download [--save-dir D] | Save PDF locally | download 2310.06825 --save-dir ./pdfs |
| read [--max-pages N] | Extract PDF text via PyMuPDF | read ./pdfs/foo.pdf --max-pages 20 |
Paper ID conventions
detail and citations auto-detect ID type:
- DOI (
10.xxxx/...) → used as-is - arXiv (
ARXIV:2310.06825, or bare numeric ≥10 digits → auto-prefixed) - Semantic Scholar paperId (long hex string) → used as-is
download requires a plain arXiv ID like 2310.06825 (no prefix).
Standard workflows
A — Literature scan
- Run
search. - Present results as a ranked table: # | Title | Year | Citations | ID.
- Ask which paper(s) to dig into.
B — Deep-read one paper
detail→ confirm match, show abstract + TL;DR.- If arXiv ID present:
download --save-dir ./pdfs. read(default 10 pages = abstract + intro + conclusion;
raise to 30+ for full read but warn about context usage).
- Summarize: problem · method · key result · limitations.
C — Impact analysis
detailthe anchor paper.citations --limit 20.- Cluster citing papers by year/theme; highlight most-cited follow-ups.
D — Build a reading list
- Run Workflow A.
- For each chosen paper, also run
citationsto find related work. - Deduplicate by paperId; annotate each entry with one-line rationale.
Decision rules
Which search first?
- Generic topic, unknown venue →
search(broader, has citation counts). - User says "preprint" / "latest" / "arXiv" →
arxiv. - Ambiguous → run both, dedupe by title.
Before downloading a PDF:
- Always
detailfirst to confirm match. - If user request was vague, show abstract and confirm before downloading.
- Then
download.
Reading PDFs:
max_pages=10default is fine for skim.- Bump to 30+ only when user wants thorough analysis.
- If extraction returns
PDF无法提取文本(可能是扫描件), the PDF is a
scanned image — offer to find an alternative version on the publisher
site or use OCR (out of scope for this skill).
Failure handling
- HTTP 429 → script auto-retries 3× with exponential backoff. If it
still fails, wait 10s and retry once before reporting.
- Empty results → broaden query (drop quoted phrases, try synonyms)
before giving up.
- Unknown paper_id format → try as DOI first, then
ARXIV:, then raw. - Missing dependency → script returns
需要安装 X: pip install X;
surface this and offer to install.
Output conventions
- Always include the paper ID alongside titles so the user can re-query.
- Cite format:
[FirstAuthor et al., Year] Title (cites: N). - For downloaded PDFs, always report the absolute save path.
- Default
save-diris the current working directory; ask if the user
wants a specific location.