SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsstorm-research

storm-research

Writing & research

View repository

Updated 10-16 05:45

About this skill

STORM Research

What this skill does

Turns a topic into a multi-perspective, verified HTML briefing. It simulates five expert angles on the topic, maps their contradictions, synthesizes everything into a standalone HTML report, and applies a critical review and source verification. The pipeline is exhaustive and skips no phase: it is intentionally heavier than a quick web search.

This is an in-house adaptation of Stanford's STORM method (Synthesis of Topic Outlines through Retrieval and Multi-perspective Question Asking, NAACL 2024). Assumed differences to always keep in mind are described in the internal documentation: prioritization of primary sources, a systematic verification constraint, and a panel built by the author.

Honest broker positioning (non-negotiable)

This skill exists only if these three guardrails hold. Without them, it is not a Storm Recherche.

  1. Real research only. Every angle and every citation must trace back to a real, open, read source. No invented study, figure, or URL. If a figure is not verifiable, it must be explicitly marked as unconfirmed.
  2. The panel is built in-house. Always disclose it in the report. Agreement between angles is a strong hypothesis, not independent proof, and above all not a consensus of the field.
  3. Verification is mandatory. A report delivered without Phase 4 is not a Storm Recherche. The verification banner must be true (actual count of invented / corrected / downgraded).

Portability

This skill is self-contained. It depends only on Claude Code's built-in tools (the tool Agent with the agent general-purpose, Write, and the web search/fetch used in those agents) plus the files in the skill folder (HTML template, CSS). It requires no other non-standard external resources.

Language and style

Output in French, neutral formal address in the report (it is a deliverable). Sober, factual tone, no overselling. No em dashes, use commas. HTML report charter: navy #0D1B2A, orange #F47C3C, cream #F7F1E1 (keep the template CSS). Inter font. Respect the layout provided by the template.

Phase 0: frame the topic

  1. If $ARGUMENTS contains the topic, use it. Otherwise, ask what to research.
  2. State your interpretation of the topic in one line and move on. Only ask a clarifying question if the topic is genuinely ambiguous in a way that changes the research. By default, move on.
  3. Identify the reader's role so that the "action" section targets them. Infer it from the topic and the given context; if it is unclear, ask in one line, or default to "a product creator / strategic decision-maker".
  4. Derive a topic-slug in kebab-case from the topic, for the file name.
  5. Tell the reader that the pipeline is running (5 angles, then verification). One line is enough.

Phase 1: five expert angles (agents in parallel)

Launch five agents general-purpose in a single message so that they run in parallel. Each receives the SAME framing of the topic plus its own angle. Use these exact prompts, substituting the variables {SUJET} and {CADRE}.

  1. THE PRACTITIONER — Tu es LE PRATICIEN pour : {SUJET} ({CADRE}). Tu travailles ce sujet au quotidien. Fais de la vraie recherche web (priorité aux sources récentes, études de cas, retours d'opérateurs). Produis : a) un résumé opérationnel, b) 6–10 points concrets, c) 6 références réelles (URL + citation courte).
  2. THE ACADEMIC — Tu es L'ACADÉMIQUE pour : {SUJET} ({CADRE}). Tu te fies aux preuves évaluées par les pairs et aux tailles d'effet, pas aux anecdotes. Fais de la vraie recherche web (études revues par les pairs, méta-analyses). Produis : a) synthèse des preuves, b) lacunes méthodologiques, c) 6 références principales.
  3. THE SKEPTIC — Tu es LE SCEPTIQUE pour : {SUJET} ({CADRE}). Tu penses que la vue dominante est surestimée ou fausse. Construis le steelman le plus solide du scénario pessimiste. Fais de la vraie recherche web pour trouver contre-exemples, limites et biais. Produis : a) arguments contraires, b) scénarios où ça casse, c) 6 références.
  4. THE ECONOMIST — Tu es L'ÉCONOMISTE pour : {SUJET} ({CADRE}). Tu suis l'argent. Fais de la vraie recherche web sur les revenus, valorisations, taille de marché, flux de financement, unit economics et modèles de monétisation. Produis : a) chiffres de marché, b) acteurs clés, c) 6 références (rapports financiers, études de marché).
  5. THE HISTORIAN — Tu es L'HISTORIEN pour : {SUJET} ({CADRE}). Tu as déjà vu des cycles de disruption et tu cherches les motifs. Fais de la vraie recherche web pour trouver de vrais parallèles historiques et des enseignements. Produis : a) analogies pertinentes, b) cycles comparables, c) 6 références historiques ou d'archives.

Each agent returns a structured brief (title, 6–10 points, exact references with URLs). When the five briefs are received, post a 2–3 line note in the chat: general direction of convergence and the sharpest disagreement. The detailed briefs remain internal (they feed Phase 2).

Phase 2: mapping the contradictions

Working only from the five briefs, determine (live, without agents):

  1. Direct conflicts — where two or more angles assert the opposite. Name the precise claims that oppose each other, not just the themes.
  2. Strongest vs weakest evidence — which angle is best supported (hierarchy: causal peer-reviewed > official data > financial report > commissioned survey > anecdote/analogy > preprint) and which is weakest, with justification.
  3. The deciding question — the single empirical question that would settle the biggest contradiction.
  4. Universal agreement — what every angle confirms, even the opponents. This is probably true.
  5. The blind spot — what NO angle has addressed. This becomes the "missing 6th angle" and feeds the Frontier Question.

This map is the raw material for the report's teachings (supported/contested), the hidden connection, the 6th angle sidebar and the frontier question.

Phase 3: synthesize the HTML report

  1. Read report-template.html in this skill folder. Clone it, do not rebuild the CSS.
  2. Fill in each section. Mapping from the phases:
    • 60-second summary — decision-maker level, nuanced but clear. Start with the established fact, then the contested interpretation.
    • 5 key teachings, ranked by reliability — the most important things now known, most reliable first. Each teaching carries a confidence score 1–10 (set in Phase 4) and a list of 1–3 primary references.
    • Hidden connection — the non-obvious link from Phase 2 that appears only by cross-referencing the five angles.
    • Hypothesis / missing 6th angle — the blind spot from Phase 2, presented as the angle that could change the conclusions.
    • Action — 3 to 6 precise moves for the reader's role identified in Phase 0. Precise, operational, dated if possible.
    • Guide "what you can claim" — three columns: Safe / Caveat / Avoid, filled in after the Phase 4 verification.
    • Boundary question — the one question that would change everything.
    • References — each citation with its verification status (set in Phase 4) and a direct link to the primary source.
  3. Write the final file to storm-reports/{topic-slug}-briefing.html (relative to the current working directory; create the folder if needed).

Phase 4: adversarial peer review + verification (do not skip)

This is what separates a Storm Recherche from an ordinary report. Do this phase before delivering.

4a. Self-review (live). Rate each of the 5 takeaways 1–10 for reliability and justify. Identify the weakest link and what would verify it. Do a bias check (which angle dominated the evidence? what assumptions guided it?).

4b. Verify each citation (agents in parallel). Launch agents general-purpose in a single message, one per distinct citation cluster (group related claims; ~4–6 agents). Prompt template: Vérifie indépendamment cette citation contre sa source PRIMAIRE. Sois sceptique, ne te fie pas aux résumés de blogs secondaires. CLAIM : {claim + chiffre cité + source nommée}. Trouve la vraie source primaire, vérifie chiffre/date/contexte, et rends un verdict (VÉRIFIÉ / CORRIGÉ / NON TROUVÉ) avec preuve (URL + extrait). Each agent returns a list of structured verdicts for its cluster.

4c. Apply the corrections. Edit the report:

  • Correct any false figure, title, date or characterization.
  • Lower confidence scores where the evidence turned out to be thin; downgrade preprints and disputed claims into the "Contested signal" box.
  • Honestly re-attribute single or commissioned poll stats.
  • Fill in the verification banner (X inventée(s), Y corrigée(s), Z rétrogradée(s)) and the per-citation status tags.
  • Fill in the "what you can claim" guide from the verdicts.

Delivery is only allowed once all verification and correction steps are completed.

Output

  1. Final deliverable: storm-reports/{topic-slug}-briefing.html (v2, after verification).
  2. Open the file for the reader with the platform's default opener: macOS open <chemin>, Linux xdg-open <chemin>, Windows start "" <chemin>. If the OS is unclear, just give the path.
  3. In the chat, provide: the file path, the verification count (N/N vérifiées, X inventée(s), Y corrigée(s), Z rétrogradée(s)), the single universal result, the boundary question, and the list of proposed actions.

Guardrails

  • Real research only. See honest broker positioning. Never an invented source.
  • The panel is built in-house. Always disclose the panel composition in the report. Convergence is a hypothesis, not a consensus of the field.
  • Verification mandatory. No Phase 4, no Storm Recherche. The banner must reflect the reality of the verifications performed.
  • Reliability = quality of evidence, not confidence. Rate according to the hierarchy: peer-evaluated causal > official/financial data > single commissioned poll > analogy > preprint.
  • Aim for the reader, not for a person by default. The action and the guide address the role identified in Phase 0. Stay generic if no role is given.
  • Cost. One execution launches ~9 to 11 agents per run. This is normal. Do not exceed five angles or one verifier per citation cluster.
  • Design. Navy/orange/cream palette, Inter. Keep the template CSS as is, do not replace it with another visual style.