Brainstorming alone tends to converge on the first three obvious ideas. Brainstorming with ChatGPT, done badly, does the same — except now you also believe a machine validated them. This guide is for the founders, writers, and operators who want ChatGPT as a sparring partner that pushes you past the obvious, not as a yes-man that polishes your first instinct.
The whole method is one loop: diverge wide, pick three, attack them, decide. Below it is mapped to the exact model picker and prompts that work as of June 2026.
TL;DR
- Ask for 10 wildly different angles, no overlap, and tell ChatGPT the obvious three you have already considered so it skips them.
- Use GPT-5.5 Instant for the diverge step (variety beats depth) and GPT-5.5 Thinking for the attack step (you want it to find real failure modes).
- The valuable angles are almost always 4 through 10. Cross off anything that matches your pre-written first three.
- Never let the model pick the “best” idea. It has no stake in your decision. Force a synthesis instead and decide yourself.
- A medium-stakes loop fits in 25-40 minutes. Save the chat or pin it to a Project — 20% of “obviously bad” ideas look great six months later.
Which model to use (June 2026)
Since the GPT-5.5 rollout on April 23, 2026, the picker on paid plans shows three options: Instant, Thinking, and Pro. They are not interchangeable for ideation:
| Picker option | Best for in brainstorming | Why |
|---|---|---|
| GPT-5.5 Instant | The diverge step — generating 10 raw angles | Fast, looser, more willing to throw out odd combinations |
| GPT-5.5 Thinking | The attack step — finding why each idea fails | Reasons step-by-step, surfaces specific failure modes |
| GPT-5.5 Pro ($200 tier only) | Hard, high-stakes synthesis with long context | Overkill for most brainstorms; save it for big bets |
On the Free plan ($0) you get GPT-5.5 on tight limits (roughly 10 messages per 5 hours as of June 2026), which is enough for one diverge-and-attack pass but not an extended session. Go ($8/mo) gives far more Instant headroom. Plus ($20/mo) is the sweet spot for regular brainstorming: about 160 GPT-5.5 messages per 3 hours plus a separate weekly Thinking allowance, so you can run the full loop several times a week without hitting a wall.
If you want the model to remember your decision criteria across sessions, put the brainstorm in a ChatGPT Project. Project custom instructions (“you are a contrarian sparring partner; never agree with my first idea”) apply to every chat in that Project, and you can save any strong output as a Project source to reference later.
Who this is for
Founders evaluating product directions, writers stuck on framings, PMs naming features, operators comparing trade-offs. Anyone making a decision where “what are the options I am not seeing?” is the bottleneck.
Skip this workflow when the decision is purely about facts — that is research, not brainstorming. If you keep saying “I don’t know enough,” go find data first (ChatGPT research workflow covers that), then come back and diverge.
Before you start
- Write the problem in one sentence. If you can’t, brainstorming is premature — you are still framing the problem.
- Note your own first three ideas before opening ChatGPT. Otherwise its suggestions anchor you and you stop generating your own.
- Decide what “good” looks like. “Something I have not thought of” is a fine criterion; “the best idea” is not.
The loop, step by step
- State the problem with a constraint. “I am [role] choosing [decision] given [constraint]. List 10 wildly different angles, no overlap.” Constraints (audience, budget, timeline) are what produce specific ideas — without them you get “ideas about marketing.”
- Read all 10 on Instant. Cross off the ones that overlap with your pre-written first three. The valuable ones are usually 4 through 10. Note which angles you reflexively dismiss; those are often worth a second look.
- Pick the 3 most interesting (or most uncomfortable). Ask the model to elaborate each into a 5-line pitch with one named risk.
- Switch to Thinking and attack each pitch. “What is the strongest reason this is wrong? Who specifically would this fail for?” Reject vague answers like “might not work” — push until the failure mode is concrete.
- Force a synthesis. “Combine the best part of pitch A with the risk-mitigation of pitch C into a new option D.”
- Stop when you have a top-2 you can defend in writing. Save the chat. The rejected ideas often resurface.
Three frameworks to widen the diverge step
When 9 of 10 angles feel obvious, your problem statement was too narrow. These named techniques force genuinely different axes — paste one into the diverge prompt:
- Cross-axis prompting. Tell ChatGPT each angle must sit on a different axis — audience, distribution, format, business model, integration — and that no two may share an axis. This alone kills most of the near-duplicate “ideas” that pad a normal brainstorm.
- Inversion (contrarian). “Give me 10 ways to guarantee this fails,” then flip each into what to avoid. Failure modes are often easier to generate than wins, and the flips are non-obvious.
- Six Thinking Hats. Ask the model to evaluate one shortlisted idea from six stances in turn — facts, feelings, risks, benefits, creative alternatives, and process. It’s a fast way to pressure-test a single direction from angles you wouldn’t naturally take.
Example diverge prompt
I am a solo founder choosing the wedge for my new ops-tools
product. Constraint: I have 8 weeks of runway to validate.
Give me 10 wildly different wedges — each on a different axis
(audience, distribution, format, business model, integration).
No overlap. Skip the obvious 3 (Notion-style template, Slack
bot, Chrome extension). For each: one sentence pitch, target
user, and the first thing you would build in week 1.Example attack prompt (switch to Thinking)
Here are my top 3 wedges from the list: [paste 3].
For each, give me:
1. The single strongest reason it fails in the first 8 weeks.
2. The specific user segment it would NOT work for, and why.
3. One cheaper experiment that would expose that risk in week 1.
Be adversarial. Do not hedge. If a wedge is weak, say so plainly.Quality check
Before you act on the output, confirm:
- Did the model give you angles you would not have generated alone? If most feel obvious, your problem statement was too narrow — loosen one constraint and re-run.
- Are the attacks substantive (specific user, specific failure mode) or vague (“might not work”)?
- Could you defend your final choice to a skeptical investor or editor without re-running the chat?
Common mistakes
- Stopping at the first 3 ideas — those are the ones you would have generated solo anyway.
- Not pushing back. The model will agree with whatever you escalate. Disagree on purpose to find the weak spots.
- Running the whole thing on one model. Instant for variety, Thinking for the attack — mixing them up dulls both steps.
- Letting ChatGPT pick which angle is best. That is your job — the model has no skin in your decision.
- Using brainstorming when the bottleneck is actually research.
- Throwing away the rejected ideas. Save the chat or Project; 20% of “obviously bad” ideas look great six months later.
How to reuse this as a system
- Spin up a ChatGPT Project named “Brainstorm” with custom instructions that set the contrarian tone once, so every new chat starts there.
- Keep a
brainstorm-decisions.mdlog outside ChatGPT: problem, 10 angles, chosen path, what would change your mind. Re-read quarterly. - For team sessions, pre-run this solo, bring the 10 angles to the meeting, and skip the warm-up phase.
FAQ
- Should I tell the model what I have already considered?: Yes — list your dismissed-three so it skips them and generates genuinely new ones.
- Reasoning model or fast model?: Instant for the diverge step (variety matters more than depth), Thinking for the attack step. On the $200 Pro tier you can reach for GPT-5.5 Pro on truly high-stakes synthesis, but it is overkill for routine brainstorms.
- Can I do this on the Free plan?: Yes for a single pass. Free runs GPT-5.5 on tight limits (about 10 messages per 5 hours as of June 2026); for repeated sessions, Plus at $20/mo gives roughly 160 GPT-5.5 messages per 3 hours.
- What if all 10 angles feel weak?: Your constraint is probably wrong. Loosen one constraint and re-run, or attack the problem statement itself.
- Is brainstorming with AI worse than with humans?: Different. AI is faster and gives more variety; humans bring lived constraints and emotional stakes. Use both, not one.
Related
- ChatGPT writing assistant
- ChatGPT for Research — Outline, Source, Synthesize
- ChatGPT Resume Workflow — From Old to ATS-Ready
- ChatGPT Voice Mode — When It Actually Helps
- ChatGPT Canvas — Long-Form Editing Without Losing the Thread
- ChatGPT for Meeting Notes — Transcript to Action Items
- ChatGPT for Translation — Better Than Just Pasting and Hoping
- ChatGPT Vision — Explaining Screenshots, Charts, and UI