STORM Deep Research

STORM for Claude Code

Type one command. Get a cited, multi-perspective research report.

_{Architecture figure from the Stanford OVAL STORM project. This plugin reproduces the top (STORM) loop as a Claude Code prompt pipeline.}

Stanford built a research system called STORM. It writes Wikipedia-style, cited articles by interviewing a topic from several perspectives. You can try the official version at storm.genie.stanford.edu.

This plugin runs the same idea inside Claude Code. One command, no GitHub clone, no API keys. It discovers perspectives, interviews each one against real web sources, maps where they disagree, writes a cited report, and grades its own work.

Before and after

Without it, you ask once and get the majority view with no sources. Or you paste four separate prompts and keep track of the thread yourself.

With it, you type one command and get a sourced, multi-perspective report saved to a file:

/storm:storm the impact of generative AI on education

How it works

/storm:storm runs the STORM pipeline step by step.

1. Perspective discovery. Five lenses (practitioner, academic, skeptic, economist, historian) plus STORM's "basic fact writer", adapted to the topic.
2. Grounded interviews in parallel. One storm-researcher subagent per perspective runs a writer-and-expert interview over three rounds, backed by real web search. Every claim is cited.
3. Contradiction map. Where the perspectives clash, what they all agree on (so probably true), and the blind spot none of them covered.
4. Outline. Drafted first, then refined with what the interviews found.
5. Cited article. Written section by section (storm-writer subagents in parallel), with [n] citations globalized and de-duplicated by URL.
6. Synthesis briefing. A 60-second summary for a decision maker, five findings ranked by reliability, a non-obvious connection, a concrete action, and the open question that would change the picture.
7. Self peer-review. Confidence scores from 1 to 10, the weakest claim, a bias check, a missing perspective, and a final grade. This is the fix for STORM's known weakness: the system does not critique itself.

The report is saved to storm-<topic>.md with a clickable References section.

Commands

Command	Web	Subagents	Output
/storm:storm	yes	yes, in parallel	Full cited report, briefing and peer-review, saved to disk
/storm:storm-brief	no	no	Fast four-phase briefing, not sourced, shown in chat

/storm:storm-brief runs Nav Toor's four-prompt method in a single pass with no web calls. Use it when you want the five-minute briefing without waiting on research.

Install

Claude Code, from GitHub

/plugin marketplace add hadufer/claude-storm
/plugin install storm@storm-marketplace
/reload-plugins

Claude Code, from a local folder

/plugin marketplace add /path/to/claude-storm
/plugin install storm@storm-marketplace
/reload-plugins

Try it for one session, without installing

claude --plugin-dir /path/to/claude-storm

Validate the plugin before publishing:

claude plugin validate .

Usage

# Full grounded, cited research
/storm:storm the impact of generative AI on education

# Optional flags
/storm:storm electric vehicles --depth deep --lang en
/storm:storm quantum computing --depth quick

# Fast one-pass briefing, no web
/storm:storm-brief SaaS pricing strategy --role founder

Flags

• --depth quick | standard | deep: research depth. Default is standard, meaning five perspectives plus the basic writer and three interview rounds.
• --lang <code>: language of the deliverable. Default is the language of your query.
• --no-web: skip web search. Avoid this; use storm-brief instead.
• --role <role> (brief mode): your role, used to tailor the concrete action.