brycewang-stanford/Searching Scientific Literature

<!-- ╔══════════════════════════════════════════════════════════════╗ ║ 本文件为开源 Skill 原始文档，收录仅供学习与研究参考 ║ ║ CoPaper.AI 收集整理 | https://copaper.ai ║ ╚══════════════════════════════════════════════════════════════╝ 来源仓库: https://github.com/kthorn/research-superpower 项目名称: research-superpower 开源协议: MIT License 收录日期: 2026-04-02 声明: 本文件版权归原作者所有。此处收录旨在为社会科学实证研究者 提供 AI Agent Skills 的集中参考。如有侵权，请联系删除。 -->

Searching Scientific Literature

Overview

Search PubMed for scientific literature using optimized queries. Extract metadata and prepare papers for relevance evaluation.

Core principle: Cast a wide enough net to find relevant papers, but use targeted keywords to keep results manageable.

When to Use

Use this skill when:

• Starting a new research question
• User asks "find papers about..."
• Need initial paper set for evaluation
• Searching for specific methods, compounds, diseases, techniques

Search Strategy

1. Parse User Query

Extract:

• Keywords: Main concepts (e.g., "BTK inhibitor", "selectivity", "kinase")
• Data types: What user needs (IC50 values, methods, structures, results)
• Constraints: Date ranges, specific journals, author names
• Synonyms: Alternative terms (e.g., "Bruton's tyrosine kinase" = "BTK")

2. Construct PubMed Query

Boolean operators:

• AND - narrow results (must have both terms)
• OR - broaden results (either term)
• NOT - exclude terms

Example queries:

"BTK inhibitor"[Title/Abstract] AND selectivity[Title/Abstract]

("kinase inhibitor" OR "protein kinase") AND (selectivity OR "off-target")

"ibrutinib"[Title/Abstract] AND ("IC50" OR "inhibitory concentration")

Field tags:

• [Title/Abstract] - search title and abstract only
• [Title] - title only (more precise)
• [Author] - specific author
• [Journal] - specific journal
• [Date] - date range

3. Execute Search

API endpoint:

https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esearch.fcgi?\
db=pubmed&\
term=YOUR_QUERY&\
retmax=100&\
retmode=json&\
sort=relevance

Parameters:

• db=pubmed - search PubMed database
• term= - your query (URL encode spaces and special chars)
• retmax=100 - max results (start with 100)
• retmode=json - return JSON
• sort=relevance - most relevant first (or pub_date for newest)

Example bash:

curl "https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esearch.fcgi?db=pubmed&term=BTK+inhibitor+selectivity&retmax=100&retmode=json&sort=relevance"

Response format:

{
  "esearchresult": {
    "count": "156",
    "retmax": "100",
    "idlist": ["12345678", "87654321", ...]
  }
}

4. Fetch Paper Metadata

API endpoint:

https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esummary.fcgi?\
db=pubmed&\
id=12345678,87654321&\
retmode=json

Extract from response:

• Title
• Authors (list)
• Journal name
• Publication date
• Abstract (via separate efetch call or use esummary)
• PMID
• DOI (if available in articleids)

Getting DOI from PMID:

"articleids": [
  {"idtype": "pubmed", "value": "12345678"},
  {"idtype": "doi", "value": "10.1234/example.2023"}
]

If DOI missing:

• Use PMID as fallback identifier
• Try to resolve DOI via PubMed Central or publisher APIs later

Output Format

Create list of paper objects:

[
  {
    "pmid": "12345678",
    "doi": "10.1234/example.2023",
    "title": "Selective BTK inhibitors for autoimmune diseases",
    "authors": ["Smith J", "Doe A", "Johnson B"],
    "journal": "Nature Chemical Biology",
    "year": "2023",
    "abstract": "We developed a series of...",
    "source": "pubmed_search"
  }
]

Error Handling

Rate limits (CRITICAL - shared across all processes/subagents):

• No API key: 3 requests/second (official limit)
• With API key: 10 requests/second
• Single agent/script: Use 500ms delays (2 req/sec, safe margin)
  • 350ms is theoretically sufficient but causes ~20% HTTP 429 errors in practice
• Multiple parallel subagents: Use longer delays to share capacity
  • 2 parallel: 1 second each (2 total req/sec)
  • 3 parallel: 1.5 seconds each (2 total req/sec)
  • 5 parallel: 2.5 seconds each (2 total req/sec)
  • Formula: delay_seconds = (num_parallel / rate_limit) + safety_margin
• If you get HTTP 429 errors: Wait 5 seconds, resume with doubled delays

Empty results:

• Try broader terms
• Remove field tags
• Check for typos
• Use OR to add synonyms

Too many results (>500):

• Add more specific terms
• Use field tags to narrow
• Add date constraints
• Consider splitting into sub-queries

Integration with Other Skills

After search completes:

1. Save results to research folder as initial-search-results.json
2. For each paper, call evaluating-paper-relevance skill
3. Track in papers-reviewed.json (use DOI as key, fallback to PMID)

Quick Reference

| Task | Command | |------|---------| | Search PubMed | curl "https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esearch.fcgi?db=pubmed&term=QUERY&retmax=100&retmode=json" | | Get metadata | curl "https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esummary.fcgi?db=pubmed&id=PMID1,PMID2&retmode=json" | | URL encode query | Replace spaces with +, special chars with %XX | | Narrow results | Use AND, add field tags, more specific terms | | Broaden results | Use OR, remove field tags, add synonyms |

Common Mistakes

Too narrow: Only 5 results → Use OR, remove constraints Too broad: 5000 results → Add AND terms, use field tags Missing abstracts: Use efetch instead of esummary for full abstract text DOI not found: Many older papers lack DOI - use PMID as fallback Rate limiting: Add 500ms delays (single agent) or longer (parallel subagents sharing rate limit)

Next Steps

After completing search:

• Announce: "Found N papers matching query"
• Begin evaluation using skills/research/evaluating-paper-relevance
• Update user with progress as papers are screened