video youtu.be Watch ↗

STORM: A Fixed Panel of Adversarial Research Lenses, Packaged as a Skill

TL;DR

A 12-minute walkthrough (Nate Herk / AI Automation) of a free Claude skill that repackages Stanford's STORM research method into a fixed pipeline: scope the topic → run five role-typed expert lenses in parallel (practitioner, academic, skeptic, economist, historian) → map where they contradict → synthesize → adversarially peer-review its own output and verify every citation against its primary source → emit a consistent, self-contained HTML briefing. The load-bearing idea is not the skill but its shape: a single prompt is "Claude as a search box — one angle in, one confident answer out; it can't see what it can't see," so you manufacture structured disagreement across a small fixed panel and force a verification pass, rather than trusting one pass. This is the Adversarial Planning Council pattern turned from a build/kill gate toward verified research synthesis — a third independent instance of "parallel personas whose real value is the cross-review round." The source's contrast case is Claude's native Deep Research, which fanned out 100+ dynamic agents into a thin brain-dump markdown; STORM's ~12-agent fixed roster was, per the source, faster, "100% cheaper," rate-limit-safe, consistent in output shape, and judged better on six dimensions by an external Codex model — the demand side of Bounded Fan-Out over an unbounded swarm. A secondary teaching beat draws the subagent vs agent-team distinction (Agent Communication Topology): STORM uses hub-and-spoke subagents that can't talk to each other, and gets its cross-examination from an explicit hub-run contradiction+review stage instead of a mesh debate. The honest caveat is on the artifact's own face: the briefing labels its panel "author-constructed… all five lenses share one framing, so where they agree, treat it as a strong hypothesis, not independent proof."

Concepts introduced

Bounded Fan-Out — new. A fixed, small roster of agents vs an unbounded, dynamically-sized swarm as two operating points; the source's STORM (~12) vs Deep Research (100+) contrast is the worked case for "bounded wins on cost, latency, rate-limit exposure, and output consistency — without necessarily costing quality."
Agent Communication Topology — new. Subagents (hub-and-spoke; workers can't talk to each other) vs agent teams/councils (mesh; agents talk to each other and debate to consensus) as a distinct design axis from how many agents or which tier.
Adversarial Planning Council — extended. STORM is a third independent instance of the parallel-personas-plus-cross-review primitive, applied to research synthesis rather than a Go/Reshape/Kill gate; the source's own artifact sharpens the concept's open question about whether persona diversity substitutes for model diversity.
Role-Typed Agent Roster — the five lenses are a role-typed roster with the council intent (personas engineered to disagree), not the division-of-labour intent.
Evidence-Gated Completion — the citation-verification pass (sources confirmed / corrected / demoted, "1 fabricated, dropped") is completion gated on showable, source-checked evidence.
Model-Tier Routing — running the lens subagents on Opus 4.8 with the explicit option to drop them to Haiku/Sonnet is per-subagent tier routing.
LLM-as-Judge — the Codex cross-model verdict and the pipeline's own adversarial peer-review are judges in the loop; routing the verdict to a different vendor is the independence discipline.
Recency-Grounded Research — the sibling research pattern: both replace a fluent single-pass answer with a ranked, source-anchored, written-to-file artifact.
Golden Templates — the report-template.html is an output-format exemplar the skill adapts for consistency ("adapt, don't invent" at the report layer).
Skill Invocation Trigger — the skill fires from a natural-language description match, no slash command: the model-invoked side of the trade.

Held, not dropped (themes touched, not yet warranting their own page):

Reliability-ranked findings with per-claim provenance — each key finding scored 1–10 and tagged with which lenses supported vs challenged it. A concrete output contract; spin out if a second source converges on scored-provenance-per-claim as a reusable format.
Self-surfacing coverage-gap / missing-lens audit — the briefing names the assumption it rests on and the lens that is absent, turning "what did we not look at?" into a first-class output field.
Personal tailoring of a research skill — feeding the skill your business/goals so every briefing is scored against your situation (the personal-dedup discipline applied to research output).
Author-constructed panel ≠ independent evidence — the shared-framing caveat the artifact prints on itself; a sharpening of the persona-vs-model-diversity question held under Adversarial Planning Council.

Key claims

A single prompt gives one angle and inherits that angle's blind spots; multiple opposed perspectives each find holes the others miss. principle — durable and the video's core claim; it is the same manufactured-disagreement logic as Adversarial Planning Council, stated for research rather than idea-critique. "It can't see what it can't see."
Get cross-perspective critique by adding an explicit contradiction-map + peer-review stage over parallel outputs, rather than requiring the agents to talk to each other. best practice — context: research/synthesis where you want the blind-spot payoff on hub-and-spoke subagents (cheaper) without paying for a mesh agent-team debate. See Agent Communication Topology.
A bounded fixed roster of agents can beat an unbounded dynamic swarm on cost, speed, rate-limit exposure, and output consistency for a well-scoped research task. best practice — context: a task whose relevant angles can be enumerated up front (a fixed set of lenses covers it); where the needed subtasks can't be predicted, a dynamic orchestrator-workers swarm is the right tool. See Bounded Fan-Out.
Gate research completion on source-checked evidence: verify every citation against its primary source and mark each confirmed / corrected / demoted / dropped. best practice — context: briefings that feed a decision; the receipt ("16 sources checked, 1 fabricated dropped, 4 corrected, 2 demoted") is what makes V2 trustworthy where V1 wasn't. See Evidence-Gated Completion.
Route the "which is better?" verdict to a different model/vendor than the one that produced the work. best practice — context: reducing same-model self-bias; the source used a Codex model as the external judge. See LLM-as-Judge.
An author-constructed panel of personas on one model with a shared framing produces correlated, not independent, judgments — agreement across lenses is a strong hypothesis, not independent proof. principle — durable; stated by the artifact itself, and the honest limit on persona-diversity research. It is why the verification pass, not the panel's agreement, carries the trust.
A four-prompt chain (angles → contradiction map → synthesis → peer review) can be packaged into one invocable skill with a fixed output template for consistent, hands-off reuse. observation — how the pipeline was productized; the report template is a Golden Templates instance.

Check-worthy source claims (attributed, not adjudicated — a later grounding pass can verify):

Stanford's STORM has been shown in peer-reviewed testing to produce articles "25% more organized than the next best method." observation — a specific benchmark figure; groundable.
"Stanford's own researchers flagged that STORM does not self-critique — source bias and fact misassociation sneak in," which the peer-review prompt is designed to counter. observation — attributed limitation of the underlying method; groundable against the STORM papers.
Claude Code "natively has a feature called deep research" that spins a dynamic workflow of hundreds of agents (the source cites 103), and large simultaneous fan-outs got hit by API rate limits. observation — a named product capability and behaviour; groundable, point-in-time.
In this run STORM used ~12 agents and was "faster and 100% cheaper" than the 100+-agent Deep Research run, and a Codex model judged the HTML briefing better on evidence quality, source diversity, thesis, actionability, risk control, and video/content. observation — the source's own single-run comparison, not a controlled benchmark; the "100% cheaper" figure is a claim, not a measurement.
The lens subagents ran on Opus 4.8 (with Haiku/Sonnet named as cheaper options); a "Claude Fable 5 is currently unavailable" notice was on screen. observation — named models, point-in-time UI detail, groundable.

Why this builds on the graph

The dominant stance is builds_on: STORM extends Adversarial Planning Council out of its original build/kill-gate frame into verified research synthesis, and it does so as a third independent instance of the same primitive already recorded there (a personal "roast" ritual and the public llm-council-skill are the first two; STORM is the third — parallel personas whose payoff is the cross-review round, not the parallel answers). Two secondary relations matter. It corroborates that concept's caution about persona-vs-model diversity, and does so from an unusually honest angle — the artifact prints on itself that its author-constructed, shared-framing panel yields correlated rather than independent judgments, which is why the trust rides on the verification pass, not on the lenses agreeing. And it supplies the demand-side case for two axes the graph named abstractly: Bounded Fan-Out (a fixed roster beating an unbounded swarm on cost/consistency/rate-limits, and — per the source's own external judge — not on quality either) and Agent Communication Topology (hub-and-spoke subagents plus an explicit review stage substituting for a mesh debate). There is no contradiction with existing pages; STORM is convergent evidence plus two new distinctions, not a tension to resolve.

Illustrated walkthrough