How to do an in-depth brainstorm on a problem with AI

Which tool to choose

They're not equivalent for this task.

Choose ChatGPT if you want ideas in rapid fire, divergent and out of the box. It's stronger at open brainstorming and imaginative content; Claude is more precise and orderly, which produces cleaner but less "explosive" results. For the divergent phase — when you're after quantity and unexpected angles — ChatGPT is the better choice, and the free plan holds up the start: web browsing, file analysis, uploaded and generated images.

Choose Claude if the problem is delicate and you want rigorous reasoning, without ideas invented out of whole cloth. The two tools are close, but Claude works better as a sparring partner when you need intellectual honesty more than brilliance.

The operational steps below apply identically on both.

How to do it

The secret is to turn the AI from an "answer engine" into a facilitator (someone who guides and structures a session without giving the answers themselves). You do it in three moves: you set the method, you diverge, then you converge.

1. Create a dedicated space, not a throwaway chat. On ChatGPT use the Projects feature: you find it in the sidebar, where you create new workspaces. A Project groups multiple chats, accepts reference files and fixed instructions, so the session accumulates context instead of restarting from scratch. It's available on all plans, free and paid.

2. Frame the problem in a brutally honest way. Before asking for ideas, write to the AI: the real problem, what you've already tried, the constraints (budget, time, people), and what "solved" would mean. Without this, you'll get banalities.

3. Impose a structured method. Free-wheeling brainstorming produces disconnected ideas; a framework forces systematic exploration. Two methods that work well with AI: Edward de Bono's Six Hats (you examine the problem from six angles — facts, emotions, risks, benefits, creativity, process) and SCAMPER (generates ideas via Substitute, Combine, Adapt, Modify, Put to other use, Eliminate, and Reverse attributes of an existing concept).

The operational syntax to paste for the divergent phase:

Act as a facilitator of an in-depth brainstorm. Don't give me the predictable
answers.

The problem: [describe the problem here in 3-4 lines, with constraints and what
you've already tried]

Proceed like this:
1. First reframe the problem for me in 3 different ways, because often the real
   problem isn't the one I think it is.
2. Then apply the Six Hats framework and for each hat (White-facts,
   Red-emotions, Black-risks, Yellow-benefits, Green-new ideas, Blue-process)
   generate at least 3 concrete ideas for me, not generic ones.
3. At the end, list the 5 most promising ideas and for each tell me why
   it could fail.

One idea per line, no preamble. If you lack information to be
precise, stop and ask me before proceeding.

Feedback: if the AI starts firing off generic ideas like "advertise more" or "improve communication", the prompt didn't take hold. Stop it and write: "too generic, I want ideas specific to my case, restart from point 1".

4. Push on divergence. The first list is almost always the surface level. Push again:

Give me 10 more, but these must be ideas that the majority would discard
as absurd.

This is where the real ideas emerge.

5. Convert into convergence. When you have enough material, switch modes:

Now play devil's advocate. Tear apart my 3 favorite ideas [list them] and
tell me under what conditions they would actually work instead.

Concrete example

Marta runs a small independent bookshop and has a problem: on Tuesday afternoons the shop is deserted and she loses money on the rent. She opens a Project on ChatGPT, names it "Empty Tuesday afternoon", and pastes the prompt above, describing the constraint ("zero advertising budget, I'm alone in the shop, residential neighborhood with many elderly people and families").

On the first round the AI reframes the problem: it's not "people are missing", but "I haven't given people a reason to come specifically on Tuesday". The Six Hats produce ideas like reading groups for pensioners on Tuesdays, school-order pickup, a "suspended literary coffee". Marta discards the first two and asks for 10 "absurd" ideas: out comes a handwriting club for screen-free children, which the neighborhood — full of parents worried about too much phone time — welcomes. In the convergent phase the AI flags the risk (it needs a leader), and she solves it by involving a retired teacher who's a customer. After three Tuesdays the workshop brings in 12 regular families. Problem solved, zero cost.

When it does NOT work (and how to fix it)

If the AI churns out only obvious and generic ideas

It's the most common failure and almost always stems from poor framing: the AI fills the gaps with clichés. Fix: increase the specific constraints (the more fences you give, the more creative it becomes) and explicitly ask for "ideas that 90% of people would discard". Always add "applied to MY case, not in general".

If all the ideas resemble each other

It's not an impression. Research published in Nature Human Behaviour shows that ChatGPT can raise the quality of the individual idea, but reduces the diversity of ideas within a group: in one experiment 94% of the ideas of those using ChatGPT shared overlapping concepts, while those generated by humans were entirely distinct. Fix: generate the starting directions yourself and use the AI to dig deep into each one. Then launch the same question on a second tool (Claude if you used ChatGPT) and compare: discard the overlaps, keep the differences.

If you get stuck on the free message limit mid-session

On the free plan the caps come quickly in long sessions, and they change often: today free ChatGPT users have a handful of messages on the full model before sliding to a lighter version, while paid plans hold up many more per window. Fix: instead of many short requests, group into dense messages (ask for "20 ideas + critique" in one shot). If it dumps you onto the reduced model, resume after the window or continue on Claude so as not to lose the rhythm.

If you lose the thread between one session and another

Single chats don't communicate with each other. Fix: the Project solves it. If you brainstorm in one chat and later open another in the same Project asking for a summary, ChatGPT recovers the details of the previous session, something impossible between separate chats.

A tip from someone who really uses it

The trap is stopping at the first list because it already seems good. The best ideas come on the third or fourth follow-up, when you've exhausted the obvious and force the AI off the beaten path. Keep a fixed rule: before choosing, always have the AI tell you why each idea could fail. The ideas that survive the critique are the ones to bet on. The AI reshuffles the known: you provide the original spark, it multiplies it.

Frequently asked questions

Which prompt do I use if I don't want to learn the Six Hats or SCAMPER?

Paste this:

Act as a brainstorming facilitator. I give you a problem, you ask me one question
at a time to understand it deeply before proposing any idea. When you have
enough information, generate 15 ideas for me from the safest to the boldest.

The facilitator's logic matters more than the name of the method.

Does the free plan hold up or do I have to pay?

For an occasional brainstorm the free version holds up: it's fine for quick questions, drafts, and one-off tasks. If you do long and frequent sessions, the message limit will interrupt you at the best moment, and there it makes sense to consider a paid plan so as not to lose momentum.

Is it worth using a specialized app for brainstorming instead of ChatGPT?

For someone starting out, no: a well-instructed general-purpose tool covers the need. Dedicated apps add value only if you work in a distributed team or have to visually organize large amounts of ideas.

Doesn't AI make everyone lazier and more mediocre with ideas, instead of helping us?

It's the serious criticism, and it has a kernel of truth: used badly, AI levels downward and everyone produces the same ideas — the data on group diversity seen above confirms it. But the problem is the method, not the tool. If you use it to expand and stress-test your intuitions instead of replacing them, you become more creative, not less. The difference is made by whoever holds the helm.