research-superpower

内容来源：README.md（说明文档） · 原始地址 · 查看安装指南

原始内容

# Research Superpowers: Literature Search & Review

Give Claude Code superpowers for **systematic literature searching and review**. Search PubMed, screen papers, extract data, traverse citations, and synthesize findings.

**Focus:** Finding and reviewing scientific papers, not lab work or data analysis.

## What You Get

A complete toolkit for **systematic literature reviews**:

- **Literature Search** - PubMed and Semantic Scholar integration
- **Smart Paper Screening** - Abstract scoring + deep dive for relevant data extraction
- **Screening Rubrics** - Build and test custom relevance criteria collaboratively
- **ChEMBL Integration** - Check if medicinal chemistry papers have curated SAR data (~99k papers)
- **Open Access Finding** - Unpaywall API to find free versions of paywalled papers
- **Citation Traversal** - Intelligent backward and forward citation following
- **Large-Scale Screening** - Parallel subagent processing for 50+ papers
- **Research Orchestration** - End-to-end workflow from query to synthesized findings

## Quick Start

### Option 1: Install from Marketplace (Recommended)

```bash
# In Claude Code, add the marketplace
/plugin marketplace add https://github.com/kthorn/research-superpower

# Install the plugin
/plugin install research-superpowers@research-superpowers-marketplace

# Start asking literature search questions!
# Example: "Find papers on BTK inhibitor selectivity with IC50 data"
# Example: "Review literature on CRISPR gene editing in cardiomyocytes"
# Example: "What's known about MmpL3 inhibitors for tuberculosis?"
```

### Option 2: Manual Installation

```bash
# Clone to your config directory
git clone https://github.com/kthorn/research-superpower ~/.config/research-superpowers

# Start Claude Code and tell it your literature search question
```

**Note:** These skills help you **find and review papers**. For analyzing experimental data or designing experiments, you'll need other tools.

## How It Works (Literature Review Workflow)

1. **Parse your literature question** - Extract keywords, data types, constraints
2. **Build screening rubric** - Define what makes papers relevant (optional, for large searches)
3. **Search PubMed** - Find candidate papers
4. **Screen abstracts** - Score papers for relevance (0-10)
5. **Deep dive on promising papers** - Fetch full text, extract specific data and methods
6. **Traverse citations** - Follow relevant references and citing papers (via Semantic Scholar)
7. **Synthesize findings** - Organize results in SUMMARY.md with structured extraction
8. **Track everything** - Maintain research session folder with findings, PDFs, and deduplication

## Project Structure

```
research-superpowers/
├── skills/
│   ├── getting-started/        # Introduction and workflow overview
│   └── research/
│       ├── answering-research-questions/    # Main orchestration
│       ├── searching-literature/            # PubMed search
│       ├── evaluating-paper-relevance/      # Abstract screening + deep dive
│       ├── checking-chembl/                 # ChEMBL SAR data lookup
│       ├── traversing-citations/            # Citation network traversal
│       └── finding-open-access-papers/      # Unpaywall integration
├── scripts/
│   └── find-skills              # Search for relevant skills
└── hooks/
    └── session-start.sh         # Auto-load at Claude Code startup
```

## Research Session Output

Each research query creates a folder:

```
research-sessions/2025-10-11-btk-inhibitor-selectivity/
├── SUMMARY.md                   # Main findings organized by relevance
├── papers-reviewed.json         # Deduplication tracking
├── papers/
│   ├── 10.1234_example.pdf
│   └── ...
└── citations/
    └── citation-graph.json      # Citation relationships
```

## Skills Library

### Literature Search & Review Skills

- **answering-research-questions** - Main orchestration workflow (search → screen → extract → synthesize)
- **building-screening-rubrics** - Collaborative rubric design with test-driven refinement
- **searching-literature** - PubMed API integration with query optimization
- **evaluating-paper-relevance** - Two-stage relevance filtering (abstract + deep dive)
- **subagent-driven-review** - Parallel screening for large searches (50+ papers)
- **checking-chembl** - ChEMBL database lookup for medicinal chemistry papers
- **traversing-citations** - Smart citation following via Semantic Scholar
- **finding-open-access-papers** - Unpaywall API for finding free full text
- **cleaning-up-research-sessions** - Safe cleanup of intermediate files after research complete

## Philosophy

These skills focus on **systematic literature searching and review**:

- **Precision over breadth** - Find papers with specific data you need, not just topically related
- **Test-driven screening** - Build and validate relevance rubrics before bulk processing
- **Smart traversal** - Only follow relevant citations to avoid exponential explosion
- **Track everything** - Deduplicate papers, cache abstracts for re-screening, maintain provenance
- **Check in regularly** - Report progress every 10 papers, checkpoint every 50
- **Reproducible** - Save rubrics, queries, and methodology with each research session

**What this is NOT:** These skills don't analyze experimental data, design experiments, or perform statistical analysis. They help you **find and review published literature**.

## Requirements

- Claude Code CLI
- Internet connection (for PubMed and Semantic Scholar APIs)
- Optional: API keys for higher rate limits (free tier works for most use cases)

## Reducing Command Prompts

When Claude runs API calls (curl commands), you may see approval prompts. To eliminate these for research sessions:

### Option 1: Pre-configure permissions (Recommended)

Copy the template permissions file to your research project:

```bash
# In your research project directory (e.g., research-sessions/YYYY-MM-DD-query/)
mkdir -p .claude
cp ~/.config/research-superpowers/.claude/settings.local.json.template .claude/settings.local.json
```

This pre-approves:
- All PubMed/NCBI API calls
- All Semantic Scholar API calls
- All Unpaywall API calls
- DOI resolution calls
- Basic file operations

### Option 2: Per-command approval

When prompted, choose option 2: "Yes, and don't ask again for similar commands in this directory"

**Note:** If this doesn't work (keeps prompting), use Option 1 instead. Claude Code may add specific commands instead of patterns.

## API Information

**PubMed/E-utilities:**
- Free, no API key required
- Rate limit: 3 requests/second without key, 10 req/sec with key

**Semantic Scholar:**
- Free tier: 100 requests per 5 minutes
- API key (free): 1000 requests per 5 minutes
- [Get API key](https://www.semanticscholar.org/product/api#api-key)

**Unpaywall:**
- Free: 100,000 requests per day
- No API key required (just provide email)
- [Learn more](https://unpaywall.org/products/api)

**ChEMBL:**
- Free, no API key required
- ~99,000 curated medicinal chemistry papers
- Structured SAR data (IC50, MIC, structures, assays)
- [Learn more](https://www.ebi.ac.uk/chembl/)

## Contributing

This is an experimental project! Contributions welcome:
- Try it with real research questions
- Report issues and edge cases
- Suggest new skills
- Improve existing workflows

## License

MIT License - see LICENSE file

## Acknowledgments

Inspired by [Superpowers](https://github.com/obra/superpowers) by Jesse Vincent

Research Superpowers: Literature Search & Review

Give Claude Code superpowers for systematic literature searching and review. Search PubMed, screen papers, extract data, traverse citations, and synthesize findings.

Focus: Finding and reviewing scientific papers, not lab work or data analysis.

What You Get

A complete toolkit for systematic literature reviews:

Literature Search - PubMed and Semantic Scholar integration
Smart Paper Screening - Abstract scoring + deep dive for relevant data extraction
Screening Rubrics - Build and test custom relevance criteria collaboratively
ChEMBL Integration - Check if medicinal chemistry papers have curated SAR data (~99k papers)
Open Access Finding - Unpaywall API to find free versions of paywalled papers
Citation Traversal - Intelligent backward and forward citation following
Large-Scale Screening - Parallel subagent processing for 50+ papers
Research Orchestration - End-to-end workflow from query to synthesized findings

Quick Start

Option 1: Install from Marketplace (Recommended)

# In Claude Code, add the marketplace
/plugin marketplace add https://github.com/kthorn/research-superpower

# Install the plugin
/plugin install research-superpowers@research-superpowers-marketplace

# Start asking literature search questions!
# Example: "Find papers on BTK inhibitor selectivity with IC50 data"
# Example: "Review literature on CRISPR gene editing in cardiomyocytes"
# Example: "What's known about MmpL3 inhibitors for tuberculosis?"

Option 2: Manual Installation

# Clone to your config directory
git clone https://github.com/kthorn/research-superpower ~/.config/research-superpowers

# Start Claude Code and tell it your literature search question

Note: These skills help you find and review papers. For analyzing experimental data or designing experiments, you'll need other tools.

How It Works (Literature Review Workflow)

Parse your literature question - Extract keywords, data types, constraints
Build screening rubric - Define what makes papers relevant (optional, for large searches)
Search PubMed - Find candidate papers
Screen abstracts - Score papers for relevance (0-10)
Deep dive on promising papers - Fetch full text, extract specific data and methods
Traverse citations - Follow relevant references and citing papers (via Semantic Scholar)
Synthesize findings - Organize results in SUMMARY.md with structured extraction
Track everything - Maintain research session folder with findings, PDFs, and deduplication

Project Structure

research-superpowers/
├── skills/
│   ├── getting-started/        # Introduction and workflow overview
│   └── research/
│       ├── answering-research-questions/    # Main orchestration
│       ├── searching-literature/            # PubMed search
│       ├── evaluating-paper-relevance/      # Abstract screening + deep dive
│       ├── checking-chembl/                 # ChEMBL SAR data lookup
│       ├── traversing-citations/            # Citation network traversal
│       └── finding-open-access-papers/      # Unpaywall integration
├── scripts/
│   └── find-skills              # Search for relevant skills
└── hooks/
    └── session-start.sh         # Auto-load at Claude Code startup

Research Session Output

Each research query creates a folder:

research-sessions/2025-10-11-btk-inhibitor-selectivity/
├── SUMMARY.md                   # Main findings organized by relevance
├── papers-reviewed.json         # Deduplication tracking
├── papers/
│   ├── 10.1234_example.pdf
│   └── ...
└── citations/
    └── citation-graph.json      # Citation relationships

Skills Library

Literature Search & Review Skills

answering-research-questions - Main orchestration workflow (search → screen → extract → synthesize)
building-screening-rubrics - Collaborative rubric design with test-driven refinement
searching-literature - PubMed API integration with query optimization
evaluating-paper-relevance - Two-stage relevance filtering (abstract + deep dive)
subagent-driven-review - Parallel screening for large searches (50+ papers)
checking-chembl - ChEMBL database lookup for medicinal chemistry papers
traversing-citations - Smart citation following via Semantic Scholar
finding-open-access-papers - Unpaywall API for finding free full text
cleaning-up-research-sessions - Safe cleanup of intermediate files after research complete

Philosophy

These skills focus on systematic literature searching and review:

Precision over breadth - Find papers with specific data you need, not just topically related
Test-driven screening - Build and validate relevance rubrics before bulk processing
Smart traversal - Only follow relevant citations to avoid exponential explosion
Track everything - Deduplicate papers, cache abstracts for re-screening, maintain provenance
Check in regularly - Report progress every 10 papers, checkpoint every 50
Reproducible - Save rubrics, queries, and methodology with each research session

What this is NOT: These skills don't analyze experimental data, design experiments, or perform statistical analysis. They help you find and review published literature.

Requirements

Claude Code CLI
Internet connection (for PubMed and Semantic Scholar APIs)
Optional: API keys for higher rate limits (free tier works for most use cases)

Reducing Command Prompts

When Claude runs API calls (curl commands), you may see approval prompts. To eliminate these for research sessions:

Option 1: Pre-configure permissions (Recommended)

Copy the template permissions file to your research project:

# In your research project directory (e.g., research-sessions/YYYY-MM-DD-query/)
mkdir -p .claude
cp ~/.config/research-superpowers/.claude/settings.local.json.template .claude/settings.local.json

This pre-approves:

All PubMed/NCBI API calls
All Semantic Scholar API calls
All Unpaywall API calls
DOI resolution calls
Basic file operations

Option 2: Per-command approval

When prompted, choose option 2: "Yes, and don't ask again for similar commands in this directory"

Note: If this doesn't work (keeps prompting), use Option 1 instead. Claude Code may add specific commands instead of patterns.

API Information

PubMed/E-utilities:

Free, no API key required
Rate limit: 3 requests/second without key, 10 req/sec with key

Semantic Scholar:

Free tier: 100 requests per 5 minutes
API key (free): 1000 requests per 5 minutes
Get API key

Unpaywall:

Free: 100,000 requests per day
No API key required (just provide email)
Learn more

ChEMBL:

Free, no API key required
~99,000 curated medicinal chemistry papers
Structured SAR data (IC50, MIC, structures, assays)
Learn more

Contributing

This is an experimental project! Contributions welcome:

Try it with real research questions
Report issues and edge cases
Suggest new skills
Improve existing workflows

License

MIT License - see LICENSE file

Acknowledgments

Inspired by Superpowers by Jesse Vincent