| name | cited-by |
| description | Search for papers that cite a given arXiv paper on INSPIRE-HEP, cache the citation list, filter by topic, and optionally read matching papers. Trigger when the user asks to find papers citing a specific arXiv paper, especially "在引用 arxiv:... 的文献中搜索XXX", "look for XXX from those cited arxiv:...", "find papers citing arxiv:... about XXX", etc. |
Cited-By Literature Search Protocol
Overview
Given an arXiv ID and an optional topic, fetch all papers that cite the given paper via INSPIRE-HEP, cache the full list as a compact markdown file, filter for topic relevance, and present candidate papers to the user. If the user approves specific papers, delegate to the arxiv-reading skill to read each one and produce a report.
Output format
The cache file is a markdown file with pipe-separated fields:
# Citing Papers for arXiv:0911.3380
**Total:** 500 papers | **Topic:** cosmology
1. Paper Title | Author1, Author2, Author3 | cited 42 | arXiv:1234.5678
2. Another Title | Author1 et al. | cited 15 | (no arXiv ID)
- Authors are comma-separated, truncated to first 6 + "et al." for large collaborations.
- No INSPIRE IDs or year fields are stored, to keep the file compact.
Workflow
Step 1 — Fetch citing papers
Use scripts/fetch_citations.py to search INSPIRE-HEP and cache all citing papers.
python3 <skill-dir>/scripts/fetch_citations.py --out-dir <user-working-dir>/arxiv-reading <arxiv-id>
Where <skill-dir> is ~/.claude/skills/cited-by/ and <user-working-dir> is the session's primary working directory.
If the user provides a topic filter, pass --topic:
python3 <skill-dir>/scripts/fetch_citations.py --out-dir <user-working-dir>/arxiv-reading --topic "<topic>" <arxiv-id>
The script outputs a markdown file at <user-working-dir>/arxiv-reading/arXiv_<arxiv-id>_citations.md.
Step 2 — Check cache
Before running fetch_citations.py, check if <user-working-dir>/arxiv-reading/arXiv_<arxiv-id>_citations.md already exists. If it does and the user hasn't asked for a refresh, skip the download and read the cached file directly.
Example:
cache_file="<user-working-dir>/arxiv-reading/arXiv_<arxiv-id>_citations.md"
if [ -f "$cache_file" ]; then
echo "Cache hit: $cache_file"
fi
Step 3 — Filter by topic and present candidates
Read the cached markdown file and filter papers whose title or authors match the topic the user requested:
- Read the
.md file to get all papers (each line after the header is N. Title | Authors | cited N | arXiv:ID).
- Search for topic keywords in paper titles and authors.
- Rank candidates by relevance (title match first, then citation count).
- Present the candidate list (up to ~15 papers) as a numbered table with columns: #, Title, Authors, Cited, arXiv.
- Ask the user which papers they want to read (by number, or "all", or refine the topic).
Step 4 — Read approved papers
For each paper the user approves, invoke the arxiv-reading skill to download and read it. Collect findings.
Step 5 — Generate report
After reading all approved papers, generate a report summarizing how each paper relates to the user's topic of interest. The report should be written to <user-working-dir>/arxiv-reading/<arxiv-id>_citing_report_<topic>.md.
Available scripts
| Script | Purpose | Usage |
|---|
scripts/fetch_citations.py | Fetch all citing papers from INSPIRE-HEP, output compact .md | python3 <skill-dir>/scripts/fetch_citations.py --out-dir <user-working-dir>/arxiv-reading [--topic "keyword"] <arxiv-id> |
Notes
- All cache files are read/written under
<user-working-dir>/arxiv-reading/.
<skill-dir> is ~/.claude/skills/cited-by/.<user-working-dir> is the session's primary working directory — use the absolute path from the conversation context, never $PWD or os.getcwd().- The INSPIRE API returns at most 250 results per page;
fetch_citations.py handles pagination automatically.
- Authors are stored as a comma-separated string (not a list); collaborations with >6 authors are truncated to first 6 + "et al.".
- No INSPIRE record IDs or publication years are stored — this keeps the cache file compact for LLM reading.
- The file format changed from JSON to markdown for ~50% smaller file size and easier LLM consumption.
- Always confirm with the user before delegating to arxiv-reading for individual papers (it involves downloading HTML).
- If
requests is not installed: pip install requests.
