storm-research — Claude Code Skill

English | 한국어

A STORM + Co-STORM-inspired research skill for Claude Code. It turns one topic into a sourced, Wikipedia-style long-form article — discovering perspectives dynamically, running grounded writer↔expert conversations with a moderator that surfaces blind spots, building an outline/mind-map, and writing sectioned prose with numbered citations. It then adds two sections that patch STORM's known gaps: a Debates & Open Questions section and a Reliability & Review appendix (self-critique). Tuned by default for Physical AI / robotics, with a generic seed pack for any topic.

This is NOT the official Stanford STORM system (Synthesis of Topic Outlines through Retrieval and Multi-perspective Question Asking, Stanford OVAL Lab — https://github.com/stanford-oval/storm). It is a prompt/agent-orchestration adaptation of STORM's and Co-STORM's ideas. It approximates STORM's retrieval index, embedding-based knowledge base, and persistent state with a logged evidence pack and an LLM-maintained markdown mind-map — faithful in method, not in infrastructure. Output always requires your own verification.

Version: 1.1.0 · see CHANGELOG.md

---

What it does

An 8-phase pipeline (see storm-research/SKILL.md and storm-research/references/storm-pipeline.md):

0. Scope — topic, role, goal, seed pack (physical-ai | generic), local sources on/off, depth.
1. Perspective discovery (STORM) — derive topic-specific perspectives from analogous sources' outlines (not fixed personas).
2. Grounded conversation + moderator (STORM + Co-STORM) — writer↔expert Q&A answered only from retrieved sources; a moderator mines retrieved-but-unused evidence for blind spots.
3. Outline + mind-map (STORM 2-pass + Co-STORM KB) — structural prior → conversation-refined outline, with a node→evidence map.
4. Article generation — sectioned prose with inline numbered [n] citations + a References list.
5. Polish — lead/summary written last + a faithful dedup pass.
6. Grafted sections — a Debates & Open Questions section + a separate-lane Reliability & Review appendix (confidence, weakest link, bias check, missing perspective, citation fact-check).
7. Output — saves ./storm-research/YYYY-MM-DD-<topic-slug>.md and posts a chat summary.

Modes: article (default) · adaptive depth (scale perspectives/turns to the topic) · prompts-only (raw prompt scaffolding only — a simplified path that skips Phase 1 discovery).

Requirements

• Claude Code with skills support (the Skill tool / /skills).
• For the full pipeline: an environment where Claude can dispatch subagents and use web search/fetch tools. Without those, run a shallower single-context pass or prompts-only mode.
• No external services, API keys, or Python. Pure markdown orchestration.

Install

Quick install from GitHub

git clone https://github.com/sageraii/storm-research-skill.git
cd storm-research-skill
./install.sh             # personal scope — available in all projects
# or
./install.sh project     # current repository only

Then restart Claude Code (or reload skills) and invoke /storm-research.

Option A — Personal skill (every project) · recommended

./install.sh            # copies to ~/.claude/skills/storm-research

or manually: mkdir -p ~/.claude/skills && cp -R storm-research ~/.claude/skills/

Option B — Project skill (one repository)

./install.sh project     # copies to ./.claude/skills/storm-research

or copy storm-research/ into your repo's .claude/skills/ and commit it for your team.

Option C — As a Claude Code plugin (advanced)

Put storm-research/ under skills/ in a plugin repo with a .claude-plugin/plugin.json (name, version, description) and a marketplace.json, then /plugin marketplace add <your-repo> and /plugin install storm-research.

Usage examples

/storm-research
storm-research로 'imitation learning 데이터셋 구축' 장문으로 조사해줘
storm-research로 '<topic>' 분석해줘 (generic 팩, 로컬 포함)
storm-research prompts-only   # raw prompt scaffolding only (skips dynamic discovery)

Local corpus (optional, opt-in)

In Physical AI mode you can let the skill also read a local corpus. Only when you enable it, it looks (relative to your working directory) for folders named arxive/, benchmark/, robots/ and project memory. These are relative, opt-in leads — there are no hardcoded paths, and the skill works fine with external sources only.

Package layout

storm-research-skill/
├── README.md           (this file)
├── README.ko.md        (한국어)
├── CHANGELOG.md
├── install.sh
├── LICENSE
└── storm-research/      (the skill — drop into .claude/skills/)
    ├── SKILL.md
    └── references/
        ├── storm-pipeline.md        (STORM + Co-STORM operational mechanics)
        ├── personas-physical-ai.md  (Physical AI seed perspectives)
        ├── personas-generic.md      (generic seed perspectives)
        ├── grounding-physical-ai.md (sources, tools, evidence rules)
        ├── output-template.md       (article + References + Debates + Reliability appendix)
        └── prompt-templates.md      (prompts-only mode)

Credits

Inspired by Stanford OVAL Lab's STORM (Shao et al., NAACL 2024) and Co-STORM (Jiang et al., EMNLP 2024), and by the viral "four-prompt" reinterpretation of STORM. This skill is an independent adaptation, not affiliated with or endorsed by Stanford. License: MIT.