With the skill installed, you provide a topic and the pipeline runs in five phases (cyan) — scope, lenses, map, synthesis, and verification — through to the verified HTML report. It's the mental map for Track 2.
Course path map
Detailed content
📄 storm-research skill anatomy
How the skill works under the hood: two files, frontmatter that triggers it, and safeguards that make the briefing reliable.
The skill consists of two files in a folder: SKILL.md (the pipeline in text) and report-template.html (the report template).
Understanding that "skill = two files" demystifies installation and shows exactly what to copy.
SKILL.md · report-template.html · skill folder.
The block at the top of the SKILL.md with name, description e argument-hint; a description is the trigger Claude reads to invoke the skill.
É a description that triggers STORM when you ask for research—understanding this helps avoid "why didn't it run?".
name · description · argument-hint · trigger.
The skill uses only native tools (Agent, Write, web search/fetch); no external scripts, APIs, or paid services.
Because it’s self-contained, just drop the folder into .claude/skills/ and it works on any machine.
Self-contained · portable · native tools.
report-template.html is the report’s visual template; the pipeline clones and fills it in without rewriting the CSS.
Keeping the template intact ensures clean, consistent reports—Montserrat, Roboto Mono, and blue highlight.
Template · clone · visual identity.
A run creates about 9 to 11 agents: 5 perspectives in Phase 1 and ~4 to 6 verifiers in Phase 4.
Knowing the expected cost avoids surprises and keeps you from inflating the panel beyond what’s recommended.
Cost · 5 perspectives · 1 verifier per cluster.
The skill's safety rules: real sources only, nothing made up, and anything that can't be verified is downgraded or cut.
These safeguards are what make the briefing reliable — they’re the method’s "promise kept."
Real research · downgrade · verification required.
⬇️ Install the skill
Download the files, drop them into the right folder, and confirm that Claude recognizes the skill — including common errors that trip up beginners.
The four files to download: SKILL.md e report-template.html, in Portuguese and English.
This is the starting point — without the files, there’s no skill to install. This module includes the download links.
Download · PT/EN · skill + template.
Both files go together in .claude/skills/storm-research/, in your project or on your home page.
The right folder is what lets Claude Code find and recognize the skill.
.claude/skills · skill folder · files together.
The same skill works with other agents (Codex and similar tools); only the folder changes— .codex/ or .agents/.
Shows that the skill is portable and that you aren’t tied to a single agent.
Portable · .codex · .agents.
A simple prompt for Claude to list the skills and confirm that storm-research appears.
Confirm the installation worked before you rely on it.
List skills · recognition · check.
Tell Claude who you are and what your business is, so the reports are tailored to your role.
Context makes the actionable insight speak directly to your reality, not to a generic person.
Context · target reader · tailored.
Typical failures: wrong folder, only the SKILL.md without the template, or file permissions.
Knowing the common errors resolves most “it didn’t work” issues in seconds.
Wrong folder · missing template · permissions.
⚙️ The 4-phase pipeline
From scope to delivery: the phases that turn a topic into a verified briefing — scope, five lenses, contradiction map, synthesis, and verification.
Claude reads the topic, summarizes it in one line, identifies the target reader, and derives a topic-slug.
A good scope targets the practical section to the right reader and names the final file.
Interpretation · target reader · topic slug.
Five general-purpose agents launched in a single message, each with the same frame and its own lens, in fewer than 400 words.
This is the multi-perspective heart of the method — and running in parallel is what makes the panel practical.
Parallel · general-purpose · shared frame.
Claude cross-checks the five briefs and identifies direct conflicts, strong versus weak evidence, universal agreement, and blind spots.
This is the raw material for the findings and the report’s “supports/challenges” chips.
Direct conflict · universal agreement · blind spot.
Clone report-template.html and fill in the summary, 5 findings, hidden connection, 6th lens, insight, boundary, and references.
This is where the research becomes a readable report, saved in storm-reports/.
Clone template · sections · storm-reports.
Self-evaluation (scores from 1–10, weakest link, bias, 6th lens) plus ~4 to 6 parallel verifiers checking each citation.
This is what separates STORM from an ordinary report; without it, the briefing isn’t reliable.
Adversarial · verification · banner.
The post-verification deliverable: the HTML v2, opened with xdg-open, plus a summary in the chat (scorecard, universal finding, boundary).
Knowing what to expect in the output helps you judge the quality of what you received.
v2 · scorecard · universal finding.
🧭 Your first report
In practice: choose a good topic, launch the pipeline, follow the lenses, open the HTML, and iterate until you have a better version.
A good topic has multiple points of view and checkable claims—not a simple factual query.
The right topic accounts for half the result; a shallow topic wastes the entire panel.
Contested · checkable · suitability.
A simple command—"Run a Storm Research on <seu tópico>. The reader is <seu papel>." — with the variable parts marked.
This is the actual trigger; knowing how to fill it in is what gets the pipeline running.
Prompt · variables · target reader.
During Phase 1, you see 5 subagents running in parallel; you can click to view each one's prompt.
Following shows that the research is real and helps you trust the result.
Subagents · parallel · transparency.
At the end of Phase 1, Claude posts 2 to 3 lines saying where the lenses converge and what the strongest divergence is.
This is an honest summary of the panel — not a dump of raw briefings into the chat.
Convergence · divergence · synthesis.
The report is in storm-reports/<slug>-briefing.html; open with xdg-open (Linux), open (macOS) or start "" (Windows).
Knowing how to open the file is the last step in reading your briefing.
xdg-open · path · HTML briefing.
You can request a new version—for example: "run the missing 6th lens and generate a V3".
Iterating refines the report and explores angles the first pass missed.
Iteration · 6th lens · V2/V3.