How To Build A Multi-Agent AI STORM Research Team

SStorm Research Tutorial

Signals in the Noise · Tutorial

How To Build A Multi-Agent AI STORM Research Team

Run Deep Research With a Panel of AI Experts: The STORM Skill in Claude, Step by Step.

Alright. Let's freaking go.

This guide was made for you or anyone on your team, even if you guys have never opened Claude Code before.

By the end, you and your employees will have a research tool running on your own computers that studies a topic with a panel of AI experts, checks its own facts against the original sources, and hands back a clean, beautiful looking report that opens in a web browser.

There won't be any work in a terminal. And there won't be any code to mess with. Everything happens inside the Claude desktop app. I can't promise this will always be the case with my tutorials, because to get into some of the more powerful AI superhero skills, we'll need to get into Claude Code or Codex and start doing work with some commands, but not today.

Today, we're going to keep the training wheels on. If you're already an advanced AI user, sorry. If you're new to all of this stuff, you're welcome.

What the heck is "STORM"?

STORM is a research method that came out of Stanford University, from people much smarter than me.

The name stands for Synthesis of Topic Outlines through Retrieval and Multi-perspective Question Asking. Right. That's a mouthful. They also didn't include the "Q" and "A" in the acronym. I guess STORMQA is less sexy?

Stanford's researchers built it to generate comprehensive, well-organized, Wikipedia-style articles from scratch.

The thinking they did to design this method is worth taking a moment to grok, because it explains why these reports come out so much stronger than a normal AI answer, even when you click on the "deep research" button in an AI chat window.

Most AI research works like one person sitting down to study a subject. They read, they grab the main topics from several sources, they summarize, they give you a write-up.

Stanford's point was that a single viewpoint leaves gaps. Depth comes from several people with different backgrounds asking different questions about the same topic and then comparing notes.

So the original system simulates several expert perspectives, has them research the topic against trusted sources, and pulls the findings into a structured article with citations.

(If you want to see the source, Stanford publishes the project at storm-project.stanford.edu/research/storm.)

How this upgraded STORM research model works

Okay look. Just because I don't have a PhD from Stanford doesn't mean I can't improve their research approach. The core idea is good, and it generates good research results, but I had a feeling I could make it better, so I did.

The method for creating a STORM multi-agent AI research team in this guide takes that core idea, a panel of experts researching in parallel, and builds it into a longer pipeline with a few additions the academic version didn't include, because a) we're not in academia, and b) our goal isn't to create Wikipedia-like reports, which is what the Stanford version was designed to do. I've adapted the research approach for a broader context, and one that works better for business related research.

Here's what one run of the STORM…I reeeeaaallly want to call this updated version the STORM SWARM. Can I call it that? It's literally a swarm of AI agents that go out and research things for you in parallel, and then a team of other agents swoops in and fact checks all the other agent's work.

Right. Anyway, here's what running the STORM SWARM does:

First, you pick your panel. The skill looks at your topic, proposes a set of expert roles that fit it, and stops to let you confirm or change them before anything starts.

Three roles always run: a Practitioner who knows the day-to-day reality, an Academic who cares about rigorous evidence, and a Skeptic whose job is to build the strongest case against the popular view.

Depending on the topic, you can add an Economist who follows the money, a Historian who looks for past parallels, an End User who speaks for the people on the receiving end, a Regulator who flags legal exposure, or a Forecaster who reasons about what happens next.

Then the panel goes to work. Each expert is a separate AI agent that runs at the same time as the others. Each does its own web research, marks where its evidence came from, and names the strongest fact that cuts against its own position. None of them wait in line to do their research. They all launch and do their research together at the same time, like a happy swarm of AI agents.

Next, it maps the contradictions. The skill reads the briefs from all the expert personas you selected and finds where the experts disagree, what they all confirm, and what none of them thought to raise. That last gap becomes a fresh "blind-spot expert" who runs one more pass to cover what the panel missed.

Then it writes the report and attacks its own work. Yeah this sounds like an autoimmune disease, but it's not. It's part of the magic that makes the research reports this method generates WAY better than what you'll get out of a vanilla Claude or ChatGPT deep research run.

An independent reviewer agent reads the research report draft cold and tries to break it like Igor tried to break Rocky. (+10 points if you saw that movie.)

Separate agents re-check every citation against its original source, and every practical recommendation against official documentation, correcting anything that turns out wrong or out of date.

Each finding then gets a confidence score tied to whether it survived that check. The report even ends with a guide to which claims are safe to state as fact and which ones are resting on squishy evidence.

What you get back is a single, good looking, easy to read report that you read in a web browser. A briefing with a ranked set of findings, the disagreements laid out in the open, confidence scores, and a record of what was verified and what was corrected.

Why this beats a normal deep research run

The deep research features in ChatGPT, Claude, and Gemini are good tools. They read many sources and cite them. What they tend to produce is one smooth narrative, and a smooth narrative can hide the places where some evidence fights against other evidence.

If several studies on your topic contradict each other, a standard report often blends them into one confident-sounding paragraph. The STORM SWARM keeps the disagreement visible, because separate experts argued separate corners and the report shows where they clashed.

The other difference is the checking. A normal research run cites its sources, but it rarely goes back to verify each claim against the original, and it doesn't appoint a reviewer to argue against its own conclusions. This skill does both. When a figure is wrong, out of date, or pulled from a weak source, the verification pass catches it and the report says so.

One big plus of using this approach: it pretty much stops an AI model from hallucinating and presenting its findings as if it knows exactly what it's talking about, and as if it's 100% correct with no errors.

The frontier models are great, but they still get things wrong and can hallucinate. The problem with this when it comes to trusting their research results is that you'll most often run research reports on topics where you won't be able to spot when an AI model got something wrong and is presenting it as if it's right. This multi-agent plus fact checking team of other agents that confirm or deny every claim and then correct the mistakes greatly reduces the "the AI hallucinated, got things wrong, and pretended it was 100% correct." It doesn't stop it completely, but it minimizes it.

An important caveat

All the AI experts on the panel are the same AI model playing different roles. When they agree, that is a strong signal rather than proof, because they share the same underlying blind spots. The report is built to make that explicit, rather than hide it, though. Treat it as very well-researched groundwork that shows its work, not as a final verdict.

Last thing before we get to the building part: for a quick fact you could look up in an AI search, this is overkill. This is the kind of tool you want to pull out of your toolbox when you have a more complicated topic that people are likely to have multiple perspectives or opinions on. That's where this STORM SWARM method shines.

What you'll need to build this puppy

A paid Claude plan. The Claude Code tab that you'll need to access requires at least a Pro subscription. Max, Team, and Enterprise work too. The free plan will stop you at the door. One heads-up on usage: a single STORM run spawns roughly ten to sixteen AI agents and does a lot of web research, so it uses more of your plan's monthly allowance than an ordinary chat. Pro handles it. If you expect to run these often, Max gives you more room before you reach a limit.

A Mac or a Windows PC. Both work. Windows has one small extra install, Git, covered in Step 3.

The two skill files. Download both of these files by clicking on these links to download them: SKILL.md file and the storm-report-template.html file. Those files are what make this whole method possible. Keep them together like peanut butter and jelly. You’ll need them in a few minutes.

About twenty minutes for the first setup, then five to ten minutes for each research run after that.

Awareness that the user interface will probably change, and the things I'm pointing at below may move to different places in Claude Code, because software companies always change their UI after I make a tutorial. It's like a law of the universe.

The steps

Confirm you're on a paid Claude plan

The STORM skill runs inside Claude Code, and Claude Code needs a paid plan: Pro at the minimum, or Max, Team, or Enterprise. To check, sign in at claude.ai and open your account settings. If it shows the free plan, upgrade to Pro before you go on. If you already pay for Claude, you're set. You'll also find out in Step 4: if the app asks you to upgrade the moment you open the Code tab, that means you're still on the free plan, but it's not free as in free beer. It's free, as in "yeah bro you can't do things in Claude Code without paying."

Download and install the Claude desktop app

Go to the official download page at code.claude.com/docs/en/desktop-quickstart. Near the top you'll see buttons for Download for macOS and Download for Windows. Click the one for your computer.

On a Mac, open the downloaded file and drag the Claude icon into your Applications folder, then open it.

On Windows, run the downloaded installer and follow the prompts.

The desktop app already includes Claude Code, so there's nothing else to install for it to work. You do not need Node.js or any other developer tools because for this set up, we're not going to use a terminal, which means you won't have to deal with any commands or code for now.

Windows only: install Git

Skip this step if you're on a Mac. Macs come with Git already, so there's nothing to do here.

On Windows, Claude Code uses a small free tool called Git to work with local files. Installing it takes about two minutes.

Go to git-scm.com and click the Windows download. Run the installer. You can accept every default by clicking Next through the screens and then Install. When it finishes, Git is ready and you won't have to think about it again.

Open the app and go to the Code tab

Open the Claude desktop app and sign in. Across the top, in the upper left, you'll see two tabs: Chat/Cowork, and Code. Click Code.

The Code tab button in the top right of the Claude desktop app — Click the </> Code button in the top right.

This is the tab that can work with files on your computer and run the skill. The other two can't run STORM, so make sure you click on Code.

If clicking Code shows an upgrade prompt, you're on the free plan, so go back to Step 1. If it asks you to finish signing in, do that and reopen the app.

Make a folder for your research and open it

The skill needs a home folder on your computer. Make a new, empty folder somewhere easy to find, like your Desktop, and give it a clear name such as STORM Research Reports.

Back in the Code tab, click the Local button at the bottom near the chat box. This runs everything on your own machine using your own files. Click Select folder and then Open Folder right next door, and pick the folder you just made. (In the screenshot below, I've already got a directory called "Pithy" open.) Claude Code is now pointed at that folder. Yay! Progress!

The Local button and the Open folder menu in the Claude Code tab — 1. The Local button. 2. Open folder to pick the folder you made.

Add the two skill files to your folder

Find the two files on this page here. Their names are: SKILL.md and storm-report-template.html.

Drag them both into the message box in the Code tab. In the screenshot above, I'm referring to the box with the text "Describe a task or ask a question."

You'll see them attach to your message.

If you'd rather, you can drop them into your STORM Research folder directly, using Finder on a Mac or File Explorer on Windows. Either way works.

Set up the skill and confirm it's installed

In the message box, type this and send it:

Please install these two files as a Claude Code skill. Create a folder at .claude/skills/storm-research and put SKILL.md and storm-report-template.html inside it.

Claude will create the folder and move the files in. If it asks permission to make the changes, approve them.

To confirm it worked, type a single forward slash ( / ) in the message box. A list of available commands and skills appears. You should see storm-research in that list. If it isn't there yet, start a new session from the sidebar and check again.

Start your first research run (LFG!)

Now the fun part.

We'll research a real business question, one where a normal AI answer tends to come out too rosy: whether a small business should move to a four-day workweek.

It's a good test because the evidence genuinely conflicts, and you'll watch the panel surface that conflict instead of smoothing it over.

In the message box, type:

Run the storm-research skill on this topic: Is switching to a four-day workweek a good move for a small business?

Then hit return. Full send.

You can also type /, choose storm-research from the list, and add your topic after it.

Choose your panel of experts

Before it does any research, the skill stops and shows you a numbered list of expert roles, with a short note on why each one fits this topic. It waits for you to choose. This is the checklist. The checklist is your friend.

For the four-day workweek question, a strong panel is the three core experts plus the Economist and the End User. The Economist weighs the cost and productivity math. The End User speaks for the employees who'd actually live the schedule. To pick those, reply with their numbers, for example:

1-3, 4, 6

You can also reply recommended to take the roles the skill marked as recommended, or all to run every expert. Send your choice and the research begins.

Watch the panel work

Once you choose, you'll see signs that a real panel is running, not a single chatbot answering.

A tasks panel shows several agents starting at once, each with a label like STORM P1: Skeptic lens, four-day workweek. Those are your experts, researching in parallel. Congratulations are in order. I mean, you just magically spawned an entire freaking team of expert research agents.

In the main conversation, short status lines mark each stage: the panel being dispatched, then the contradiction map, then the extra blind-spot expert, then the writing, then the verification agents checking sources. You'll see counts like Phase 1 complete: 5 of 5 lenses returned and, near the end, a tally of how many citations were checked and corrected.

This runs for a few minutes. The labels and status lines tell you which stage it's in the whole time. When you see the verification tally and a message that the report is ready, it's finished.

Open your finished report in a browser

Because you chose Local back in Step 5, the report saved straight to your computer, so there's nothing to download. Claude puts it inside your STORM Research folder, in a subfolder called storm-reports, with a filename that ends in -briefing.html.

Open your STORM Research folder in Finder (Mac) or File Explorer (Windows), open the storm-reports folder, and double-click the file ending in -briefing.html. It will open in a new tab in your default web browser, formatted and ready to read. Very slick. Very good looking.

That's your report: the findings ranked by how well they held up under checking, the places the experts disagreed shown in the open, confidence scores on each finding, and a record of what was verified.

From here, change the topic to anything you're weighing and run it again. The setup only happens once. Every future run is just Step 8 through Step 11.

If you want to send a report to someone else, email them the html file as an attachment and tell them to download that html file just like they'd download a pdf file. Then, once they download it, tell them to find it in their Finder/File Explorer and double click on it, which will open the report up in a browser tab.

And lickety split just like that you've got yourself an agentic AI research team of light and wonder that will bring you a much deeper, more nuanced, balanced view of topics that have a lot of possible perspectives and opinions.

Go run some reports, my friend!

Signals In The Noise