Storm research skill

STORM Research

Use Stanford STORM for source-grounded, long-form topic research. Treat its output as a research draft: verify consequential claims and citations before presenting it as final.

Choose the workflow

• Use STORM for an autonomous report with citations.
• Use Co-STORM when the user wants to steer an interactive research conversation. Read references/storm.md before configuring Co-STORM.
• Use ordinary web research instead when the request is small, time-sensitive, or does not justify API cost and a long pipeline.

Run STORM

1. Confirm that the user has authorized the expected model/search API usage. Never print, commit, or copy API keys into output files.

2. Check prerequisites without exposing values:

   test -n "$OPENAI_API_KEY" && echo "OPENAI_API_KEY set"
   

3. Select a retriever. Prefer duckduckgo for a keyless trial; prefer tavily, brave, bing, you, or serper when the matching key is available and reliability matters.

4. Run the bundled launcher from the skill directory:

   scripts/run-storm "TOPIC" --retriever duckduckgo --output-dir ./storm-results
   

   Pass additional official example flags after the launcher options, such as --max-thread-num 2 or --search-top-k 5.

5. Inspect the generated topic directory. Prefer storm_gen_article_polished.txt; otherwise use storm_gen_article.txt. Preserve url_to_info.json and the research logs for traceability.

6. Validate the report: spot-check citations, remove unsupported claims, distinguish inference from sourced fact, and note gaps or stale sources.

7. Deliver the requested format and link the report and evidence files. Mention model/retriever limitations when material.

Resume or reduce cost

STORM stages can reuse prior files. Run only the missing stages by passing official flags. The launcher runs all four stages by default; add any stage flag explicitly to control the stages:

scripts/run-storm "TOPIC" --retriever duckduckgo --output-dir ./storm-results \
  --do-generate-article --do-polish-article

Reduce --max-perspective, --max-conv-turn, --search-top-k, or --max-thread-num for cheaper/faster runs. Do not silently lower depth when the user explicitly requested comprehensive research.

Handle failures

• Missing model key: ask the user to configure a supported LiteLLM/OpenAI provider; do not request that they paste a secret into chat.
• Rate limits: retry with --max-thread-num 1 and lower search depth.
• Retrieval failures: switch retrievers and disclose the change.
• Weak or mismatched citations: inspect url_to_info.json, verify important sources directly, and revise the final report.
• Package/API drift: read references/storm.md, then compare against the official repository before changing the launcher.

Install for agents

Run scripts/install-skill to symlink this skill into both ~/.codex/skills/storm-research and ~/.claude/skills/storm-research. Use --copy instead when symlinks are unsuitable.