storm

name	storm
description	Execute a deep research and writing pipeline using the STORM methodology, leveraging web tools and extensive human-in-the-loop feedback to produce grounded, comprehensive articles.

STORM Deep Research Skill

This skill runs the STORM (Synthesis of Topic Outlines through Retrieval and Multi-perspective Question Answering) workflow. It turns a topic into a deeply researched, cited long-form article.

Tools to Use

Use the agent's native tools to run the pipeline:

• tavily_search: Find broad source coverage. Use max_results aggressively and prefer search_depth: 'advanced' when coverage matters.
• obscura_web_scrape: Read full pages when snippets are not enough.
• context7_search / context7_docs: Use for software libraries and framework docs.
• Local file tools (write, edit, read, bash): Keep all STORM artifacts in the local .storm directory.

Artifact Layout

Always keep state in .storm/<topic-slug>/ at the repository root.

Use this structure:

.storm/<topic-slug>/
  topic.md
  outline/
    draft.md
    refined.md
  perspectives/
    index.md
    01-general/
      perspective.md
      conversation.md
      research/
        turn-01.md
        turn-02.md
    02-<perspective-slug>/
      perspective.md
      conversation.md
      research/
        turn-01.md
        turn-02.md
  references/
    index.json
    sources/
      source-0001.json
      source-0002.json
  sections/
    01-<section-slug>.md
    02-<section-slug>.md
  final_article.md

File Rules

• Each perspective must live in its own directory under .storm/<topic-slug>/perspectives/.
• Each perspective directory must contain its own perspective.md and conversation.md.
• Research must not be kept in one shared file. Save each research turn in its own file under that perspective's research/ directory.
• Store source material separately under .storm/<topic-slug>/references/sources/, with one file per source.
• Keep .storm/<topic-slug>/references/index.json as the lookup table for citation ids, URLs, titles, and file paths.

Architecture & Workflow

The workflow has two phases: Pre-writing and Writing. Stop at every human checkpoint.

Phase 1: Pre-writing

1. Setup

   • Create a URL-safe topic slug.
   • Create .storm/<topic-slug>/ if missing.
   • Initialize topic.md, outline/, perspectives/, references/, references/sources/, and sections/.
   • Write the original topic and any user constraints to .storm/<topic-slug>/topic.md.

2. Perspective Discovery

   • Identify 5-7 distinct editorial perspectives.
   • Always include one general perspective.
   • Save the full list to .storm/<topic-slug>/perspectives/index.md.
   • Create one directory per perspective, using a numbered slug such as 01-general, 02-market-view, 03-technical-view.
   • In each perspective directory, create perspective.md with:
     • perspective name
     • short description
     • key questions to explore
   • Human-in-the-Loop Checkpoint 1: Ask the user to review, add, remove, or rename perspectives before proceeding.

3. Simulated Conversation & Deep Research

   • Run a multi-turn Writer/Expert conversation for each perspective.
   • Keep each perspective isolated in its own directory.
   • Save the running Q&A transcript to that perspective's conversation.md.
   • For each research turn, create a separate file in research/turn-XX.md.
   • Each turn-XX.md file should include:
     • the writer question
     • the search queries used
     • the sources reviewed
     • the grounded findings
     • open questions for the next turn
   • When a source is used, save it as its own JSON file under .storm/<topic-slug>/references/sources/.
   • Update .storm/<topic-slug>/references/index.json with the source id, URL, title, perspective, research turn, and source file path.
   • Reuse existing source files when the same URL appears again. Do not create duplicates.
   • Human-in-the-Loop Checkpoint 2: After the conversations finish, ask the user whether the source depth is enough or whether any perspective needs more research.

Phase 2: Writing

1. Section-by-Section Generation

   • Write each section from .storm/<topic-slug>/outline/refined.md.
   • During writing, use only the closed local reference set in .storm/<topic-slug>/references/.
   • Do not run new web searches during writing.
   • Insert inline citation markers such as [1], [2] that resolve through references/index.json.
   • Save each section as its own file under .storm/<topic-slug>/sections/.
   • Human-in-the-Loop Checkpoint 4: Ask whether the user wants to review each section or the assembled draft.

2. Polishing & Final Assembly

   • Assemble the final article from the section files.
   • Add a lead summary.
   • Deduplicate repeated content.
   • Generate the final citation index from references/index.json.
   • Append a References section at the end of the article. Render it as a Markdown table with columns: #, Title, URL, and Perspective. Include every source that appears in references/index.json, ordered by citation number.
   • Save the article to .storm/<topic-slug>/final_article.md.
   • Human-in-the-Loop Checkpoint 5: Request final review.

Execution Rules

• Be exhaustive: Search broadly during research.
• Stay grounded: During writing, use only the saved local references.
• Always cite sources: Every URL visited or used during research must appear in references/index.json and in the final article's References table. No source may be dropped silently.
• Keep artifacts split: Do not collapse perspectives into one shared conversation file. Do not collapse research turns into one shared research file.
• Keep filenames stable: Use numbered directories and numbered turn files so reruns stay predictable.
• Stop at checkpoints: Never skip a human approval step.
• Minimum length: The final article (final_article.md) must contain at least 50,000 words. Before saving, run wc -w final_article.md and keep expanding sections until the count is met. Every section must be substantive — no padding, no repetition. Reach the target through depth: more analysis, more examples, more evidence, more context.