storm-research

STORM Research — Multi-Perspective Deep Research Orchestrator

Run the full STORM pipeline on a topic. One prompt asks one question and returns the majority view — the surface. STORM asks the same topic from five independent expert lenses, maps where they fight, synthesizes a briefing no single expert could write, then red-teams its own output. In Stanford's peer-reviewed testing, multi-perspective articles were ~25% more organized and ~10% broader than single-pass research. This skill operationalizes that with dedicated sub-agents so the heavy reading happens off the main context window.

The Pipeline

Topic
 └─ Phase 1  FAN-OUT (parallel): practitioner · skeptic · economist · historian · academic
 └─ Phase 2  CONTRADICTION MAP: contradiction-mapper reads all 5 → finds clashes/agreement/gaps
 └─ Phase 3  SYNTHESIS: synthesizer reads 5 + map → cited research briefing
 └─ Phase 4  PEER REVIEW: peer-reviewer audits the briefing → reliability grade + fixes
 └─ Present consolidated result to the user

Workflow

Step 0 — Scope the Topic

1. Get the topic. If the user supplied one (storm research <topic>), use it. If it's vague or sprawling (e.g. "AI", "the economy"), ask one sharpening question to narrow it — a tight topic produces a far better briefing than a broad one.
2. Capture the audience and angle if offered (deciding / investing / writing / presenting / learning). It tunes the synthesizer's recommendations. Don't interrogate — one optional ask.
3. Derive a slug (lowercase, hyphenated, canonical: e.g. lab-grown-meat-viability).
4. Set the workspace: .storm/{slug}/. Create it: mkdir -p .storm/{slug}.

Step 0.5 — Tune the Panel (optional, powerful)

The five default personas fit most topics. For some topics, a swap sharpens the analysis — e.g. add a Clinician for a medical topic, a Regulator for a policy topic, an End User for a product topic. If a swap clearly helps, mention it to the user and spawn the extra persona with the Task tool using the same output contract as the built-in agents (worldview + owned bias + evidence-gathering + structured file with sources). Keep the panel at 5–6; more dilutes. Default to the standard five if unsure.