storm-research

A Claude Code skill that generates comprehensive, Wikipedia-quality research reports using the STORM methodology from Stanford OVAL.

What It Does

/storm-research turns any topic into a fully cited research report through five phases:

Topic → Perspectives → Multi-Source Research → Outline → Grounded Writing → Fact-Checked Report

Phase	What Happens
Perspective Discovery	Identifies 3-8 diverse expert viewpoints on the topic
Multi-Perspective Research	Parallel web searches from each perspective, gathering sources
Outline Generation	Synthesizes findings into a hierarchical article structure
Grounded Writing	Writes each section with inline [N] citations from real sources
Verification & Polish	Adversarial fact-checking, gap detection, and refinement

Installation

Option A: Symlink (recommended for development)

# Clone the repo
git clone https://github.com/aruneshvv/claude-research-skill.git

# Symlink into your Claude Code skills directory
ln -s "$(pwd)/claude-research-skill" ~/.claude/skills/storm-research

Option B: Copy

# Copy the skill file directly
mkdir -p ~/.claude/skills/storm-research
cp SKILL.md ~/.claude/skills/storm-research/SKILL.md

Option C: Install as a Claude Code skill from URL

# Using Claude Code's skill installation (if supported)
claude skill install https://github.com/aruneshvv/claude-research-skill

After installation, restart Claude Code. The /storm-research command will be available.

Usage

# Basic research (5 perspectives, ~4k words)
/storm-research The impact of quantum computing on cryptography

# Quick overview (3 perspectives, ~2k words)
/storm-research renewable energy storage --depth shallow

# Deep dive (8 perspectives, ~8k words)
/storm-research the history and future of CRISPR gene editing --depth deep

# Get just the outline
/storm-research artificial general intelligence --outline-only

# Render as an interactive HTML artifact
/storm-research supply chain resilience --format artifact

# Skip fact-checking (faster, less thorough)
/storm-research transformer architectures --no-verify

How It Works

The skill implements the core STORM methodology adapted for Claude Code's tool ecosystem:

1. Perspective-Guided Question Asking

Instead of a single research pass, the skill discovers N diverse expert perspectives on the topic. Each perspective generates different questions, ensuring broad coverage:

Topic: "Nuclear Fusion Energy"
├── Nuclear Physicist      → "What are the current plasma confinement approaches?"
├── Energy Economist       → "What are the cost projections vs. renewables?"  
├── Policy Analyst         → "What regulatory frameworks exist?"
├── Environmental Scientist → "What is the environmental footprint?"
└── Industry Engineer      → "What engineering challenges remain?"

2. Parallel Multi-Source Research

Each perspective runs its own research thread using WebSearch and WebFetch, gathering real sources with URLs. Research agents run in parallel for speed.

3. Synthesized Outline

All findings are merged into a single hierarchical outline that organizes information by theme, not by perspective — producing a coherent article structure.

4. Citation-Grounded Writing

Every section is written with inline [N] citations pointing to the actual sources discovered during research. No fabricated references.

5. Adversarial Verification

Key claims are independently re-searched to verify accuracy. Claims are marked as CONFIRMED, MODIFIED, or UNVERIFIABLE.

Output Quality

• Every factual claim has at least one real citation
• All source URLs come from actual web searches (no fabrication)
• Opposing viewpoints are presented fairly
• Uncertainty and limitations are acknowledged
• Encyclopedic, neutral tone throughout

Comparison with STORM

Feature	Stanford STORM	storm-research
Research method	Simulated Wikipedia writer ↔ expert conversations	Multi-perspective parallel web research
Search	Uses external search APIs (You.com, Bing, etc.)	Uses Claude Code's WebSearch + WebFetch
Execution	Python package, requires API keys	Claude Code skill, zero setup
Outline	Generated from conversation transcripts	Synthesized from structured research briefs
Citations	Grounded in retrieved documents	Inline [N] with URL references
Verification	Limited	Adversarial fact-checking phase
Parallelism	Sequential pipeline	Multi-agent parallel research

Requirements

• Claude Code (CLI, desktop app, or web)
• WebSearch and WebFetch tools available in your Claude Code session

Project Structure

claude-research-skill/
├── SKILL.md              # The skill definition (this is what Claude Code reads)
├── README.md             # This file
├── LICENSE               # MIT License
└── docs/
    ├── architecture.md   # Technical architecture and design decisions
    ├── usage-guide.md    # Detailed usage guide with examples
    └── contributing.md   # Contribution guidelines

Contributing

See docs/contributing.md for guidelines.

License

MIT — see LICENSE.

Credits

• Inspired by STORM from Stanford OVAL
• Paper: Assisting in Writing Wikipedia-like Articles From Scratch with Large Language Models (Shao et al., NAACL 2024)
• Built for Claude Code by Anthropic