Storm

storm

Research a topic the way Stanford's STORM does its pre-writing: discover the perspectives that matter, ask grounded questions from each, and synthesize the answers into a cited briefing. The wedge is breadth of perspectives — what different lenses surface about a topic — not breadth of sources on a single question.

This is an agent-native port of the STORM mechanism, not Nav Toor's 4-prompt chat workflow. The agent already holds retrieval, so grounded is the default — there is no un-grounded mode.

---

Phase 0 — Boundary gate

Activate storm when the request wants sourced external knowledge across perspectives. Route away otherwise:

Request shape	Skill
"Research X", "what's known about Y", "survey the evidence on Z", "grounded briefing on…"	storm
An upstream skill (e.g. team-composer) needs sourced knowledge it can't reason to	storm (hand-off)
Single-source-of-truth fact-check, one authoritative report, "verify this claim"	deep-research
Opinion, critique, planning, "assemble a team", "review my plan"	team-composer
General agent loop / build work	the harness / cerby

The deep-research line is the load-bearing one: it owns source-breadth (fan-out, fetch, adversarially verify into one report). storm owns perspective-breadth. If the request wants the ground truth on one question, it's deep-research. If it wants what several lenses reveal about a topic, it's storm.

---

Phase 1 — Perspectives in

Two paths:

• Seeded (hand-off). A caller supplied perspectives (e.g. team-composer passes its active roles). Use them verbatim as the perspective set — they are already domain-relevant lenses. Do not re-discover.
• Discovered (standalone). No seed. Discover perspectives by surveying adjacent material on the topic (STORM's move): a quick retrieval pass to find the recurring stakeholder/expert/angle framings, then name 3–4 distinct ones. A fixed generalist set is one possible seed, never the only option.

Bound it: max_perspectives (default 4). Stated up front so the user can correct it.

---

Phase 2 — Grounded Q&A per perspective

For each perspective, generate questions that lens would ask, and answer them via retrieval. Bound: max_turns_per_perspective (default 3) — at most max_perspectives × max_turns retrieval cycles total.

• Every answer carries its source(s). Keep a source→claim map from here forward — synthesis depends on it (see references/honesty-rubric.md).
• Retrieval is capability-gated: use whatever retrieval the runtime holds (web search/fetch, a *_search MCP, drive search). Do not assume a specific tool name.
• If a question's retrieval comes up empty, record the answer as [unverified / reasoning-only] — do not pad it with confident introspection.

Mechanics (question generation, dedup, the contradiction-map pass) live in references/loop-mechanics.md.

---

Phase 3 — Curate → outline

Cluster findings across perspectives, dedupe sources, and draft a structure. This is where overlap collapses and the shape of the briefing emerges. Carry the source→claim map through unchanged.

---

Phase 4 — Cited synthesis

Emit the output. Every factual line is either source-attributed or tagged [unverified / reasoning-only] — no exceptions. The honesty rule is structural: because the source→claim map is carried through Phases 2–3, a line with no mapped source cannot be presented as verified.

Output formats:

• briefing (default) — decision-oriented: a one-paragraph summary, findings ranked by reliability (each with supporting and challenging sources), the actionable insight, and open questions. Each claim cited.
• report (--report) — a longer sourced write-up (closer to STORM's article form) for when breadth matters more than turnaround.

---

Phase 5 — Optional rigor passes (flagged on; --no-rigor to skip)

Cheap, on by default, off with --no-rigor. Both are specified in references/honesty-rubric.md:

• Contradiction map — where perspectives clash, what they all agree on, and what none addressed (blind spots).
• Peer review / confidence — a confidence score per finding tied to source-backing (not fluency), the weakest link, a bias check, and the missing perspective. This pass exists primarily to catch STORM's two self-flagged failure modes — source-bias transfer and over-association of unrelated facts — so it is not cosmetic.

---

Inputs / outputs contract

Standalone:

/storm <topic> [--report] [--no-rigor] [--max-perspectives N] [--max-turns N]

No seed perspectives → Phase 1 discovers them.

team-composer hand-off:

• In: the research question + team-composer's active roles, passed as the seed perspectives.
• Out: a cited briefing returned to the panel, which folds it back into its conclusion.

Invariant: no claim ships without a source or an explicit [unverified / reasoning-only] tag. With retrieval unavailable, the entire output is tagged — the honesty invariant holds by construction, not by reminder.

---

Capability-gated retrieval & graceful degradation

Gate on "is a retrieval tool available in this runtime?" — not on which agent or vendor is running (a brittle check). If none is available, do not refuse silently and do not fabricate: produce the perspective structure and reasoning, label every line [unverified / reasoning-only], and state plainly that retrieval was unavailable so the output is ungrounded.

---

What this skill is NOT

• Not single-source-of-truth research. One authoritative report from source fan-out is deep-research. storm is multi-perspective.
• Not a panel. Opinion, critique, planning, team assembly → team-composer.
• Not Co-STORM. No human-in-the-loop turn-by-turn collaboration (separate, heavier paper).
• Not a conversation-simulation engine. The bounded Phase-2 Q&A is the whole loop; no agent-vs-agent dialogue beyond it.
• Not a Wikipedia-article generator by default. Briefing is the default; the long report is opt-in.

---

Cross-skill integration

Skill	Relationship
deep-research (if installed)	The distinct neighbor. deep-research owns single-topic, source-of-truth research (fan-out → fetch → adversarial verify → one cited report). storm owns multi-perspective research. Route "verify this claim" / "give me the authoritative answer on X" there; route "what do different lenses reveal about X" here.
team-composer	The primary caller. When a panel hits a grounding need ("verify the current state of X", "survey what's known about Y", "we're guessing here"), team-composer hands off to storm with its active roles as the seed perspectives; storm returns a cited briefing the panel folds into its conclusion. The team-composer-side routing — a storm row in its Cross-Skill table + a grounding-need trigger, plus the contradiction-map and peer-review lenses storm seeded — is wired as of team-composer v5.4.0.
sub-agent-coordinator	If per-perspective Q&A is heavy enough to parallelize, the coordinator owns the spawning/briefing — storm does not duplicate that logic.

---

Slash invocation

/agent-skills:storm <topic>
/agent-skills:storm <topic> --report
/agent-skills:storm <topic> --no-rigor --max-perspectives 3

Equivalent to the natural-language trigger, just unambiguous.

---

Status

v0.1 — new skill. Ports the STORM mechanism (perspective discovery → grounded per-perspective Q&A → cited synthesis), bounded by two knobs (max_perspectives, max_turns_per_perspective). The team-composer back-port (the grounding-need hand-off row + the contradiction-map lens in its Discuss phase + peer-review/confidence in its Audit phase) shipped in team-composer v5.4.0.

When to use this skill

Load this skill when:

• The user wants grounded research across perspectives on a topic ("research X", "what's known about Y", "survey the evidence on Z", "build me a grounded briefing on…").
• An upstream skill needs sourced external knowledge it can't reason its way to.

Do NOT load this skill when:

• The user wants the authoritative answer on a single question or to verify one claim → deep-research.
• The user wants opinion, critique, planning, or a team assembled → team-composer.
• There's no external-knowledge need — let the normal skill stack handle it.

Tags: research, retrieval-grounded, cited-briefing, perspectives, storm, hand-off-target