➕ Add a 6th lens
The five lenses are a a solid pattern, not a law. Phase 1 of the SKILL.md has five exact prompts; adding a sixth means copying this template and changing the role. The method itself already names “the missing 6th lens” as the report’s blind spot—here, you turn it into a real lens that researches and responds like the others.
The image treats this as an equation: the five standard lenses, plus a new one that follows the same contract, add up to a report with one fewer blind spot.
Each lens is an angle. The 6th lens you add follows the SAME contract as the others (the three required outputs) — so the contradiction map and synthesis can fit it in without breaking. One more deliberate angle = one fewer gap.
🧷 The contract you MUST NOT break
Every STORM lens returns exactly three things. Phase 2 (contradiction map) and Phase 3 (synthesis) read those three pieces. If your new lens doesn’t provide all three in the same format, the synthesis can’t cross-reference it with the others.
- 1.CENTRAL POSITION in 2 sentences.
- 2.STRONGEST EVIDENCE: 3 to 5 bullets, each with a named data point/case/source + URL.
- 3.THE ONE THING that only this lens would say.
Patch: add the “THE AI BEGINNER” lens
Objective: paste a 6th section into Phase 1 of the SKILL.md, following the exact template of the existing lenses — and launch 6 subagents instead of 5.
### 6. O INICIANTE EM IA Você é O INICIANTE EM IA para: {TOPIC} ({TOPIC_FRAME}). Você é <quem o relatório precisa proteger — ex.: alguém que nunca implementou isso e vai decidir só com base neste briefing>. Faça pesquisa real na web priorizando guias de primeiros passos, relatos de quem começou agora, perguntas frequentes e os erros mais comuns de iniciantes. Mostre o que um iniciante PRECISA saber antes de agir e onde ele tropeça: o atrito de entrada, os pré-requisitos escondidos e as expectativas irreais que o hype cria. Retorne EXATAMENTE: 1. POSIÇÃO CENTRAL em 2 frases. 2. EVIDÊNCIA MAIS FORTE, com 3 a 5 bullets, cada um com um guia/relato/dado concreto + fonte nomeada + URL. 3. A ÚNICA COISA que apenas um iniciante diria. Cite fontes reais com URLs. Menos de 400 palavras.
How to verify: run the skill on your usual topic. The chat should show 6 subagents running in parallel (not 5). In the final report, at least one finding must clearly come from this lens (the beginner’s angle). If the synthesis ignores the 6th lens, that’s a sign the contract for the three outputs was broken—check whether the words POSIÇÃO CENTRAL, EVIDÊNCIA MAIS FORTE e A ÚNICA COISA are written just like the other lenses.
🟡 Watch the cost (the default guardrail is 5)
The method recommends not exceed five lenses by default—each extra lens is ~1 more agent (and one more verifier in Phase 4). Adding a 6th is a deliberate choice: do this when an angle is truly missing (the beginner, the content creator, the regulator). Changing "THE AI BEGINNER" to "THE CONTENT CREATOR" only requires changing the role and the research prompt—the rest of the template stays the same.
🎯 Target the reader (tailored)
Phase 0 identifies the reader’s role to guide the actionable insight and the claims safety guide. Instead of providing this every time, you can preload your business context — and every report is already tailored to you.
Sou <seu papel — ex.: head de produto de uma fintech>. Meu negócio: <o que você faz e para quem>. Em todo Storm Research, o leitor-alvo sou eu nesse papel, salvo se eu disser o contrário.
✗ No context (generic)
- ✗Actionable insight speaks to an abstract “professional in the field”
- ✗You need to adapt the recommendations to your situation
- ✗The same report would work for anyone
✓ With context (targeted)
- ✓3 to 6 specific moves for YOUR role
- ✓The safety guide covers your actual exposure
- ✓Less rework: ready to apply
💡 Practical tip
Context guides the insight and the safety guide — doesn’t bias the lenses’ research. Each lens still does real searches, and the skeptic still challenges your premise. You change WHO the report speaks to, not WHAT it considers true.
🎨 Edit the report-template
O report-template.html is the briefing template. Phase 3 says to clone it, don’t recreate the CSS. Safely customize by changing the content and in the labels — without tearing up the visual style, which the method asks you to keep identical.
✓ Feel free to edit
- ✓Rename/translate section labels
- ✓Add a new section (e.g., "Next steps for my team")
- ✓Replace the logo/header with your brand’s
- ✓Adjust the chip help text
✗ Don't
- ✗Delete/rewrite the template CSS
- ✗Replace it with an entirely different visual style
- ✗Remove the verification banner or status tags
- ✗Remove the references section
<!-- antes --> <h2 class="section-title">Insight acionável</h2> <!-- depois (rótulo seu, classe intacta) --> <h2 class="section-title">O que <meu time> faz na segunda-feira</h2>
🟡 Why keep the CSS
The design (clean, professional, Montserrat / Roboto Mono, blue highlights) makes the briefing look trustworthy and readable on any machine, with no build step. If you change the entire style, you lose consistency across reports and risk breaking the layout. Rule of thumb: edit what you read, preserve what you see.
🧠 Subagent model
Each lens is an agent general-purpose independently. You can run different lenses in different models (Opus, Sonnet, Haiku) and balance cost and quality: heavy reasoning where it matters, a lightweight model where the task is straightforward.
Stronger model for the lenses that reason
O skeptical (builds the strongest counterargument) and the academic (weighs evidence and effect size) benefit from a more robust reasoning model.
Lightweight model for the more direct lenses
Factual collection of parallels and figures (historian, part of the practical track) often works well with a cheaper model — you cut costs without losing coverage.
Verification (Phase 4) deserves care
This is the step that makes the report honest. If you need to save time, cut back on the lenses — not on the verifiers that check primary sources.
🎚️ This is an advanced lever
The default (same model for everyone) works very well. Change the model for a subagent only when cost or usage limits become a real problem—and measure before assuming the cheaper model lowered quality.
🔀 Use with Codex and other agents
The skill is self-contained and portable: it only depends on native tools (agent, writing, web search/fetch) and the template in the same folder. That’s why it travels to other agents—only where the folder lives.
.claude/skills/storm-research/ # Claude Code .codex/skills/storm-research/ # Codex .agents/skills/storm-research/ # outros agentes ├── SKILL.md └── report-template.html
🧳 Why it’s portable
- •No external scripts, APIs, or paid services—just text + native tools.
- •The two files go together:
SKILL.md+report-template.html. - •The same 4-phase idea works—as long as the agent knows how to launch subagents in parallel and search the web.
🟡 What may vary across agents
The exact name of the subagent and search tools may vary from one agent to another. If another agent doesn’t support subagent parallelism or web access, the essence remains—you just lose the advantage of running all five lenses at once.
🌐 Carry STORM principles forward
Strip away the HTML and pipeline, and what remains is two transferable principles: (1) multiple deliberate perspectives instead of just one, and (2) adversarial verification before trusting the results. They are useful for much more than research.
Different roles weigh in on a choice; contradiction becomes input, not noise.
A skeptical agent deliberately challenges your proposal before you bet on it.
Security, performance, and readability as separate lenses on the same diff.
Phase 4 alone: each claim checked against the primary source, with a verdict.
✓ What to transfer
- ✓Deliberate roles that disagree on purpose
- ✓Map the contradiction instead of hiding it
- ✓Verify before making a claim; reveal what’s a hypothesis
✗ What not to copy blindly
- ✗Stacking identical agents thinking that covers more
- ✗Present convergence from an author-created panel as consensus
- ✗Skip verification because "it looks right"
Self-recovery (optional): when adding a 6th lens, what MUST it preserve for the synthesis to keep working?
📌 Module summary
Next module:
3.4 — Best Practices and Pitfalls