Storm research

STORM Research

A four-phase research method that produces grounded, multi-perspective analysis instead of a single flat answer. Inspired by Stanford OVAL Lab's STORM paper (Shao et al., NAACL 2024), adapted into a repeatable prompt pipeline.

The core idea: don't ask one question from one angle — interrogate a topic through several distinct experts, surface where they disagree, then synthesize and self-review. This reliably produces more organized, better-calibrated output than a one-shot answer.

When to use

Trigger this skill when the user wants genuine depth on a non-trivial topic: research requests, "analyze this from every angle," decision memos, literature-review-style summaries, contested or high-stakes questions, or any explicit mention of STORM / multi-perspective research. Skip it for simple factual lookups or quick answers — the overhead isn't worth it there.

Two modes (pick based on the topic and any user flag)

• --grounded (default for current/empirical/contested topics): Run a web_search for each perspective and cite sources inline. Use this whenever facts may have changed, numbers matter, or claims are disputable.
• --reasoning (fast, no web): Reason from existing knowledge only. Use for conceptual, philosophical, or timeless topics, or when the user explicitly wants speed / no browsing.

If the user gives no flag, infer: anything touching current events, markets, technology versions, statistics, or "what's the latest" → grounded. Pure concepts, frameworks, or thought experiments → reasoning is fine. When unsure, default to grounded.

Phase control

The user controls how the four phases run. Match their earlier instruction if they gave one; otherwise ask once at the start:

• Auto (one pass): Run all four phases and return the full report in a single response.
• Step-by-step: Run one phase, show the output, and pause for the user to say "continue" before the next. Good for steering mid-research.

If the user hasn't indicated a preference and the topic is large, default to step-by-step so they can course-correct. For quick topics, auto is fine.

---

The Four Phases

Run these in order. Full prompt templates and the exact expert-lens definitions live in references/phases.md — read that file before executing so you use the precise wording and structure.

Phase 1 — Multi-Perspective Scan

Analyze the topic through five distinct expert lenses, each with its own priorities, blind spots, and vocabulary:

1. Practitioner — what actually works in the field, hands-on tradeoffs
2. Researcher / Academic — evidence base, theory, what's proven vs. hypothesized
3. Skeptic / Critic — failure modes, weak assumptions, what could be wrong
4. Economist / Incentives — costs, benefits, who gains, second-order effects
5. Historian / Context — how we got here, prior attempts, what's genuinely new

In --grounded mode, search per lens and cite. Keep each lens genuinely distinct — do not let them collapse into the same voice.

Phase 2 — Contradiction Map

Cross-examine the five perspectives:

• Where do they directly disagree, and on what?
• Which claims have the strongest evidence? Which are weakest / most speculative?
• What did every perspective miss (blind spots common to all)?
• Which disagreements are empirical (resolvable with data) vs. values-based (won't resolve)?

Phase 3 — Synthesis

Combine everything into one coherent research report:

• A clear thesis / bottom line up top
• Ranked findings — most to least well-supported, with confidence noted
• Actionable insights or recommendations
• Open questions that remain genuinely unresolved

Phase 4 — Peer Review

Have Claude critically grade its own Phase 3 report:

• Overall confidence (0–100%) with justification
• Weakest claim in the report and why
• Bias check — what viewpoint is over/under-weighted?
• Missing angles — perspectives or evidence not covered
• One concrete revision that would most improve it

---

Output conventions

• Label each phase clearly (## Phase 1 — Multi-Perspective Scan, etc.).
• In grounded mode, cite sources for factual claims; keep the five lenses visibly separate.
• Keep the report readable prose, not a wall of bullets — reserve ranked lists for the findings in Phase 3 and the review in Phase 4.
• End Phase 4 with a one-line honest verdict on how much to trust the report.

For the full prompt wording and lens details, see references/phases.md.