Storm

STORM — multi-perspective, retrieval-grounded research

This skill reproduces the method behind Stanford OVAL's STORM (Synthesis of Topic Outlines through Retrieval and Multi-perspective Question Asking) inside Claude's own agent loop. STORM's core insight is that the hard part of research is asking good questions, and you get good questions by (1) attacking a topic from several genuine, discovered perspectives and (2) grounding every answer in real sources. references/method.md has the full mapping to the paper — read it if a situation isn't covered here.

The one rule that makes this STORM and not a vibe

Every expert answer and every claim in the final report must be grounded in a real source retrieved with websearch / webfetch, and that source must be recorded. The moment an "expert" answers from memory, you've rebuilt exactly the thing STORM was designed to beat: fluent, confident, unsourced text. If you can't find a source for a claim, cut the claim or go find one.

Inputs

• topic (required) — what to research.
• role / audience (optional) — who the briefing is for; sharpens the

"actionable insight" at the end. Infer it if it's not given and obvious; only ask if it's quick and would change the output.

• depth (optional) — how many perspectives, default 5, sensible range 3–6.

Workflow

All intermediate files live in storm-runs//. Scaffold the run with the bundled helper (path is relative to this skill's folder; use its full path if your working directory is elsewhere):

python3 scripts/storm_run.py init ""

It prints the run paths (run_dir, raw_dir, …) as JSON. Use those paths for everything below.

1 — Survey & discover perspectives

Run 2–3 broad searches to get the lay of the land (overviews, the major debates, who the stakeholders are). From what the sources actually show, derive depth distinct perspectives. Do not hardcode a stock cast — the paper's move is to let real coverage reveal the angles. A perspective is a lens with a stake: e.g. for "Ozempic for weight loss" you might land on a prescribing endocrinologist, a health-insurance actuary, a fat-acceptance researcher, a supply-chain analyst, and a long-term-safety pharmacologist — not a generic "practitioner / skeptic / economist". Write them to perspectives.md: each with a one-line lens and why it sees something the others miss.

2 — Grounded interviews (the core)

For each perspective, run a short simulated interview (3–5 turns): the perspective asks a pointed question, and a researcher answers only from websearch / webfetch results, logging sources as it goes. Each follow-up question should react to what the previous answer surfaced — that reactive threading is where depth comes from.

Hand each interview to a subagent so they run in parallel (see Fan-out below). Give each subagent the brief in references/interviewer.md with the topic, the single perspective, and the run paths filled in. Each subagent writes:

• prose findings → storm-runs//raw/.notes.md
• one source per line → storm-runs//raw/.sources.tsv,

formatted URLTitle

Separate files per perspective means no write contention during the fan-out.

3 — Merge sources, then map contradictions

Number every source once, deterministically:

python3 scripts/storm_run.py merge storm-runs/

This writes the deduped, numbered sources.md and prints a url → [n] map — use those numbers as your citation keys from here on.

Then read all the raw/*.notes.md and write contradictions.md: where do two or more perspectives directly clash (quote the clashing claims)? What does every perspective agree on (likely true — even opponents confirm it)? What did none of them address (the blind spot — often the most valuable finding)?

4 — Outline

From the notes plus the contradiction map, write a hierarchical outline.md. Organize by theme, not by perspective — the perspectives are inputs, not sections.

5 — Draft, grounded

Write report.md section by section following the outline and the structure in references/report-template.md. Every non-obvious claim carries an inline [n] citation keyed to sources.md. A sentence with no [n] should be either self-evidently general or cut. Only use citation numbers that actually exist in sources.md — never invent one.

6 — Peer review (verification, not flattery)

Re-read report.md against sources.md and do an honest self-review: a confidence score (1–10) for each key finding with a reason, the single weakest claim and what would verify it, any perspective that ended up over-weighted, and any 6th angle that would change the conclusions. Fix what you can immediately; surface the rest in the Reliability notes block at the top of the report. The template shows the exact format. This step exists because STORM's own authors flag that the method can suffer source bias and fact-misassociation — the review is a check against the recorded sources, not a confidence-boosting summary.

Fan-out (parallel interviews)

If subagents are available (e.g. Claude Code's Task tool), launch one interview subagent per perspective in a single batch, each with the references/interviewer.md brief. That parallelism is what makes a 5-perspective run fast.

If subagents are not available (e.g. a plain chat surface), run the interviews yourself, one perspective at a time, in the same session — the file layout and the rest of the pipeline are identical.

Final output

Give the user report.md (with citations and the Reliability notes block) plus a 3–4 line plain-language summary of how solid the findings are and where the gaps remain. Mention the intermediate files (perspectives.md, contradictions.md, sources.md, raw/) are there if they want to audit the trail. Don't paste a long report into chat — point them to the file and summarize.

Reference files

• references/method.md — how each step maps to the STORM / Co-STORM papers and

why the design choices matter.

• references/interviewer.md — the brief to hand each interview subagent.
• references/report-template.md — the exact report structure, citation style,

and Reliability-notes format.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: kunjan-p
• Source: kunjan-p/storm-skill
• License: MIT

Install and usage instructions live in the source repository linked above.