/subagent-driven-review

name: Subagent-Driven Literature Review
description: Use parallel subagents for large-scale paper screening and deep dive analysis
when_to_use: Large literature searches (50+ papers), parallel paper screening, deep dive analysis on multiple papers, citation network exploration, when main context is getting full
version: 1.0.0

<!-- ╔══════════════════════════════════════════════════════════════╗ ║ 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。 -->

Subagent-Driven Literature Review

Overview

**Core principle:** Fresh subagent per batch + consolidation between batches = fast parallel screening with quality control

For large literature reviews (50+ papers), dispatching parallel or sequential subagents dramatically speeds up screening while maintaining quality through consolidation checkpoints.

When to Use

Use subagent-driven approach when:

• **Large searches:** 50+ papers to screen
• **Parallelizable work:** Papers are independent, can be screened separately
• **Deep dive tasks:** Multiple papers need detailed extraction (data tables, methods, datasets)
• **Citation exploration:** Following citation networks recursively
• **Context management:** Main context getting full, need fresh context
• **Time pressure:** Need results faster than sequential screening

**Do NOT use when:**

• Small searches (<20 papers) - overhead not worth it
• Need real-time user visibility into every paper
• Papers require cross-comparison during screening
• Simple, fast screening tasks

Use Cases

1. Parallel Paper Screening (Most Common)

**Scenario:** You have 100 papers from PubMed search to screen for relevance

**Pattern:**

Main agent:
1. Splits 100 papers into 5 batches of 20
2. Dispatches 5 subagents IN PARALLEL (single message, multiple Task calls)
3. Each subagent:
   - Fetches abstracts for its batch
   - Scores using rubric
   - Returns JSON with results
4. Main agent consolidates results into papers-reviewed.json

Time savings: 5x faster than sequential!

**Prompt template for subagent:**

I need you to screen papers 1-20 from this PMID list for relevance to [QUERY].

PMIDs to screen: [PMID list]

Use the evaluating-paper-relevance skill to:
1. Fetch abstract for each PMID
2. Score 0-10 based on:
   - Keywords: [list]
   - Data types needed: [measurements, protocols, datasets, etc.]
3. Return JSON:

{
  "screened_papers": [
    {"pmid": "12345", "score": 8, "status": "relevant", "reason": "..."},
    ...
  ],
  "stats": {"highly_relevant": 3, "relevant": 5, "not_relevant": 12}
}

Do NOT update papers-reviewed.json - return results only.

**Rate limiting (CRITICAL - PubMed limits are SHARED across all parallel subagents):**
- If you are the ONLY subagent running: Use 500ms delays (2 req/sec, safe)
- If running with OTHER parallel subagents: Use longer delays to share capacity
  - You are 1 of 2 parallel: Use 1 second delays
  - You are 1 of 3 parallel: Use 1.5 second delays
  - You are 1 of 5 parallel: Use 2.5 second delays
- If you get HTTP 429 errors: Wait 5 seconds, then use 5-second delays for remaining requests

2. Deep Dive on Priority Papers

**Scenario:** Initial screening identified 15 highly relevant papers, need detailed data extraction from each

**Pattern:**

Main agent:
1. Creates TodoWrite with 15 tasks (one per paper)
2. For each paper, dispatches subagent to:
   - Fetch full text (PMC, Unpaywall)
   - Extract relevant data (tables, figures, methods)
   - Identify key findings
   - Return structured findings
3. Main agent consolidates into SUMMARY.md
4. Reviews and adds to papers-reviewed.json

Can dispatch in parallel (5 at a time) or sequentially

**Prompt template for subagent:**

Deep dive analysis for paper PMID [12345] / DOI [10.xxxx/yyyy]

Use evaluating-paper-relevance skill to:
1. Check for curated data sources (if applicable to domain)
2. Fetch full text (try PMC, then Unpaywall if paywalled)
3. Extract relevant data based on research domain:
   - Data tables and measurements
   - Methods and protocols
   - Key results and findings
   - Figures with relevant information
4. Return structured JSON:

{
  "pmid": "12345",
  "doi": "10.xxxx/yyyy",
  "full_text_source": "PMC" or "Unpaywall" or "paywalled",
  "data_sources": ["Table 1", "Figure 3", "Supplementary Data"],
  "key_measurements": ["specific values or ranges found"],
  "methods_summary": "Brief description of methods",
  "key_findings": ["Finding 1", "Finding 2", ...],
  "data_availability": "GEO: GSE12345" or "Code: github.com/..." or null
}

Do NOT update papers-reviewed.json - return findings only.

3. Citation Network Exploration

**Scenario:** Found one highly relevant paper, need to explore forward and backward citations

**Pattern:**

Main agent:
1. Dispatches two subagents IN PARALLEL:
   - Subagent A: Fetch and screen forward citations
   - Subagent B: Fetch and screen backward citations
2. Each returns list of promising PMIDs with scores
3. Main agent:
   - Consolidates results
   - Removes duplicates
   - Adds to screening queue
   - Updates papers-reviewed.json

**Prompt template for subagent:**

Find and screen forward citations for PMID [12345].

Use traversing-citations skill to:
1. Fetch forward citations from PubMed or OpenCitations
2. Screen abstracts for relevance to [QUERY]
3. Score each citation (0-10)
4. Return JSON with promising papers (score ≥7):

{
  "seed_pmid": "12345",
  "direction": "forward",
  "citations_found": 45,
  "relevant_citations": [
    {"pmid": "67890", "score": 8, "title": "...", "reason": "..."},
    ...
  ]
}

Do NOT update papers-reviewed.json - return results only.

4. Domain-Speci