Papers research

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 <query> [--limit N]	Semantic Scholar search (cap 20)	search "diffusion models" --limit 5
detail <paper_id>	Full metadata, TL;DR, top 10 references	detail 10.48550/arXiv.2310.06825
citations <paper_id> [--limit N]	Papers that cite this one (cap 20)	citations <id> --limit 15
arxiv <query> [--max-results N]	arXiv preprint search (cap 10)	arxiv "RLHF" --max-results 5
download <arxiv_id> [--save-dir D]	Save PDF locally	download 2310.06825 --save-dir ./pdfs
read <pdf_path> [--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

1. Run search <topic>.
2. Present results as a ranked table: # | Title | Year | Citations | ID.
3. Ask which paper(s) to dig into.

B — Deep-read one paper

1. detail <id> → confirm match, show abstract + TL;DR.
2. If arXiv ID present: download <arxiv_id> --save-dir ./pdfs.
3. read <pdf_path> (default 10 pages = abstract + intro + conclusion; raise to 30+ for full read but warn about context usage).
4. Summarize: problem · method · key result · limitations.

C — Impact analysis

1. detail the anchor paper.
2. citations <id> --limit 20.
3. Cluster citing papers by year/theme; highlight most-cited follow-ups.

D — Build a reading list

1. Run Workflow A.
2. For each chosen paper, also run citations to find related work.
3. 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:

1. Always detail first to confirm match.
2. If user request was vague, show abstract and confirm before downloading.
3. Then download.

Reading PDFs:

• max_pages=10 default 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:<id>, 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-dir is the current working directory; ask if the user wants a specific location.