The STORM Method: PhD-Level Research in Claude in 5 Prompts

The five prompts

Open a fresh Claude chat. Paste these in order, one at a time, waiting for each answer before you send the next. The order is the whole method — each prompt eats the output of the one before it. Swap [YOUR TOPIC] in the first one and then don't touch anything else.

1

Explain the topic

You're not after the answer here. You're loading the chat with shared vocabulary so the next four prompts have something to grip.

Explain [YOUR TOPIC] to me. Assume I'm smart but completely new to it. Cover: - What it actually is, in plain language - Why people are talking about it right now - The three things I'd need to understand before I could hold a real opinion on it No hedging, no "it depends" unless you then tell me what it depends on.

2

Five experts, five voices

This is where it stops being a search engine. One answer gives you the consensus. Five arguing experts give you the shape of the disagreement, which is where all the useful information lives.

Now answer that same question again as five distinct experts, each in their own voice, each with their own heading: 1. A PRACTITIONER who works with this every day and cares what actually happens on the ground 2. A SCEPTIC who thinks it's overhyped and wants to say why 3. An ECONOMIST who cares about cost, incentives, margins and who captures the value 4. A HISTORIAN who has seen this pattern before and will name the previous one 5. An ACADEMIC who cares about definitions, evidence quality and what we genuinely don't know yet Rules: - Do not blend them into one balanced voice. Let them disagree with each other openly. - Each expert gets their own section and their own argument, in their own register. - Each one must make at least one claim the others would push back on.

3

Find the cracks

The prompt almost everyone leaves out. You're asking Claude to audit its own five answers instead of politely summarising them.

Now interrogate those five answers. 1. Where do two or more of these perspectives DIRECTLY contradict each other? Quote the contradicting lines side by side. 2. Which perspective has the STRONGEST evidence behind it, and which has the WEAKEST? Say specifically why — name the evidence, don't just assert it. 3. What do all five agree on? Treat that agreement with suspicion and tell me whether it's actually well-supported or just conventional wisdom. 4. What did NONE of them address? This is the important one — take your time on it.

4

Synthesise into a brief

Now you collapse all of it into something you can actually act on. Note the last line — a finding you can't do anything with is trivia.

Now synthesise everything above into a research brief containing: 1. EXACT SUMMARY — one paragraph. Precise, not hedged. If there's genuine uncertainty, name it rather than writing around it. 2. KEY FINDINGS — a short list. Each one a claim I could defend out loud. 3. ONE HIDDEN CONNECTION — something none of the five experts made explicit, that only becomes visible when you put their answers next to each other. 4. ONE SPECIFIC THING I SHOULD DO with this evidence this week. Concrete. An actual action, not "consider exploring". Keep the whole brief under 600 words.

5

Make it mark its own homework

The one people skip because they've already got something that looks finished. Run it anyway — this is the prompt that catches the confident nonsense.

Now grade your own work. 1. CONFIDENCE — score every finding in that brief from 1 to 10 for how confident you actually are, and for anything under 8, tell me exactly what evidence would raise it. 2. WEAKEST CLAIM — name the single weakest claim in the brief and say what's wrong with it. 3. BIAS CHECK — which of the five perspectives dominated this synthesis? What did that domination cost us? Which voice got squeezed out? 4. Then REGENERATE the final brief with the weak claims either fixed, properly caveated, or removed entirely. Be harsh with yourself. A brief that survives this is worth more to me than one that flatters me.

One tweak that makes it much better

Turn web search on in Claude before you start and add "search the web and cite your sources with links" to prompt one. Without it, Claude answers from what it already knows — which is a lot, but it has an end date and it can't show you where anything came from.

Why five voices beats one

Ask Claude a question once and you get the middle of the road. It averages everything it knows into one confident, reasonable, faintly beige paragraph. Useful for facts. Useless for judgement.

The five-expert move works because a practitioner and an academic genuinely want different things from the same subject. Force them into separate sections and Claude can't average them — it has to commit to each position properly. Then prompt three asks where those committed positions collide, and the collisions are the actual research. That's the bit you'd normally pay a consultant for.

The practitioner

Tells you what breaks in real life. Cuts through anything that only works in a slide deck.

The sceptic

The most valuable of the five. Gives you every objection before a client, investor or your own gut raises it.

The economist

Follows the money. Answers "who's actually making a living from this" — which explains most of the hype.

The historian

Names the last time this exact thing happened and what it looked like five years later. Deflates a lot of "unprecedented".

The academic

Insists on definitions and evidence quality. Catches the places where everyone's using one word for four different things.

Swap them out

Researching something regulated? Trade the historian for a lawyer. Health-adjacent? Swap in a clinician. The five roles are a starting kit, not scripture.

Where the name comes from

Stanford's research lab built a system called STORM that writes Wikipedia-style articles by simulating conversations between different expert perspectives, then researching each one. The five prompts here borrow that idea — multiple perspectives first, synthesis second — and run it by hand inside Claude. See the section on catches for what that does and doesn't mean.

A worked example

Here's the version I ran, so you can see what "good" looks like before you spend your own ten minutes on it. Topic: agentic AI — the thing everyone's shouting about where AI does multi-step work on your behalf rather than just answering you.

1

Prompt one gave me a definition

Fine. Competent. Nothing I couldn't have got from any article. This is scaffolding, not the payoff — resist the urge to stop here.

2

Prompt two split the room

The practitioner talked about agents failing silently three steps into a chain. The economist pointed out the cost per task rises with every retry. The historian brought up business process automation in the nineties. Five separate arguments, not one summary.

3

Prompt three found the real question

The contradiction: the practitioner said the bottleneck is reliability, the economist said the bottleneck is cost. Both can't be the main constraint. That tension is the actual thing worth thinking about — and no single-shot answer would ever have surfaced it.

4

Prompt four turned it into a decision

The "one specific thing to do" came back as: pick one task you already do the same way every week, and automate only that, with a human checking the output — rather than buying an agent platform. Actionable on a Tuesday.

5

Prompt five knocked a claim out

Self-graded, one finding came back at 5/10 with "this rests on vendor-published figures I can't verify". That claim got cut from the regenerated brief. Ten seconds of work, one fewer thing I'd have said out loud and regretted.

Good topics for this

It shines on anything contested — where smart people disagree and you need to form a view rather than look up a fact. "Should we hire or contract for this role", "is this platform worth migrating to", "what's actually happening with AI regulation here". It's overkill for "what's the GST threshold" — just look that up.

Point it at your business

Same machinery, different experts. Instead of five academics arguing about a topic, you get five people arguing about your offer — and you'd much rather have that argument in a chat window than in your DMs after launch day.

Run this as prompt two instead of the standard five experts, then carry on with prompts three, four and five exactly as written.

Here's an offer I'm considering launching: WHAT IT IS: [YOUR OFFER] WHO IT'S FOR: [YOUR IDEAL CUSTOMER] PRICE: [YOUR PRICE] HOW IT'S DELIVERED: [FORMAT, LENGTH, WHAT THEY GET] WHY I THINK IT'LL WORK: [YOUR ACTUAL REASONING, INCLUDING THE SHAKY BITS] Now respond as five distinct people, each with their own heading and their own voice. Do not blend them: 1. MY IDEAL CUSTOMER — reading the sales page cold. What do they think, what confuses them, what makes them close the tab? 2. THE SCEPTIC — someone in my market who would never buy this. Why not? Be specific and a bit unkind. 3. THE ECONOMIST — is the price right? What's my real margin once delivery time is counted? What happens at 5 sales versus 50? 4. THE OPERATOR — someone who has launched this exact shape of offer before. What broke for them? What did they wish they'd built first? 5. THE POSITIONING STRATEGIST — who else is already doing something close to this, and what would make mine the obvious choice rather than the cheaper one? Each one must say at least one thing I won't enjoy hearing.

Feed it real material

Paste in your actual sales page copy, your real pricing, your genuine reasoning — including the part you're not sure about. If you give it a tidied-up version of your idea you'll get a tidied-up critique, and that's a waste of a good sceptic.

Based on everything above, give me a go / no-go / change-it recommendation. - If GO: the three things I must not compromise on. - If CHANGE IT: the single smallest change that fixes the biggest objection, and what it costs me to make it. - If NO-GO: what would need to be true for this to become a yes. Then tell me the one assumption this whole offer rests on, and the cheapest way I could test that assumption this week without building anything.

The honest catches

This method genuinely makes Claude's research better. It does not make Claude infallible, and there are a few things I'd rather you heard from me than found out in front of a client.

This isn't literally Stanford's STORM

Stanford's lab built a real research system called STORM — it's open source and you can use it at storm.genie.stanford.edu or read the code at github.com/stanford-oval/storm. What's on this page is the prompt method inspired by that idea, run by hand in Claude. Claude is not running Stanford's software. Call it "a multi-perspective research prompt" if you're ever describing it to someone precise.

Self-grading is marking your own homework

Prompt five is the most useful prompt in the pack and also the least trustworthy number in it. When Claude says "8/10 confidence", that's a considered vibe, not a measurement. It's brilliant at spotting which claim is weakest — treat the relative ranking as real and the absolute score as decoration.

The five experts are characters, not sources

Claude is playing five roles convincingly. That means it can produce a very persuasive sceptic citing a study that doesn't exist. Before you repeat any number, date, study name or quote out loud — check it. The structure of the argument is the deliverable here. The specific facts still need a human.

Run it in one chat, not five

Every prompt depends on the answers above it. Open a new chat halfway through and prompt three has nothing to interrogate. If a chat gets long and starts drifting, ask Claude to restate the five expert positions in full before you carry on — cheap insurance.

It's slower than asking once

Ten minutes and five round trips versus twenty seconds. Worth it when you need to form a view you'll defend. Complete overkill when you need a fact. Use it on the decisions, not the lookups.

Get a research brief from five experts

Make Claude research like a whole panel, not one keen assistant

What's inside