How to Use AI for Structured Brainstorming

TL;DR: “Give me some ideas for X” is the worst prompt in common use: it returns the five most predictable answers to any question, formatted confidently. Used with structure, the same model becomes a legitimately good ideation partner, one that never gets tired, never gets defensive, and will generate idea #37 with the same energy as idea #1. The structure is a three-phase loop: diverge (volume, constraints, perspectives), converge (score against criteria you define), critique (make the model attack the survivors). This page gives you the session template.

Why default output is bland, and why that’s fixable

A large language model predicts likely continuations, and the most likely answer to “marketing ideas for a product launch” is the average of every listicle ever written on the topic. That’s not a defect you’re stuck with; it’s a default you can override. Three levers move a model off its center:

• Volume. The obvious ideas come out first. Ask for 30 and the model is forced past them, the interesting material reliably lives in ideas 15-30.
• Constraints. “Ideas that require zero budget,” “ideas that would only work for us,” “ideas our biggest competitor couldn’t copy.” Constraints prune the generic space and force lateral moves. (Counterintuitively, a higher temperature setting matters less than a good constraint, randomness adds noise; constraints add direction.)
• Perspectives. “Answer as a skeptical CFO. Now as a support rep who reads every complaint. Now as our angriest churned customer.” Each persona samples a different region of the model’s knowledge.

Setup

1. Write the problem statement first, alone. One paragraph: what you’re trying to achieve, for whom, by when, and what’s already been tried or ruled out. A vague problem statement is the number-one cause of vague output, the model amplifies whatever precision you give it.
2. Load context. Paste what the model can’t know: your product, audience, past attempts, real constraints (budget, team size, brand rules). Two paragraphs of context beats twenty exchanges of correction.
3. Run the divergence round. Ask for 25-40 ideas in one shot, each as a single sentence, no elaboration yet. Elaboration slows generation and pads weak ideas until they look like strong ones.
4. Run at least one banned-list round. “All of those are now banned. 20 more, none of which share a mechanism with the first set.” This is the single highest-value move in AI ideation; round two is where the defaults run out.
5. Converge with your criteria. Define scoring dimensions that reflect your reality, e.g., impact, effort, time-to-first-result, dependency on other teams, and have the model score every surviving idea in a table. You adjust the scores; the model’s job is producing the grid, not the verdict.
6. Critique the shortlist. For the top 3-5: “Argue against each. What has to be true for it to work? What’s the most likely failure? Who inside a typical company kills this idea and why?” A model is far better at attacking a specific idea than at generating a brilliant one, use that.

Example prompt

The divergence prompt worth saving:

“Brainstorm with me in rounds. Problem: [one paragraph, goal, audience, deadline, what’s been tried]. Context: [your product/team/constraints, 2-3 paragraphs]. Round 1: 30 ideas, one sentence each, numbered. Cover at least three ideas in each category: obvious/safe, unconventional-but-doable, and deliberately extreme (impractical is fine, extremes reveal directions). Hard constraints: [e.g., no paid ads, must ship inside a quarter, cannot touch pricing]. Don’t evaluate, don’t elaborate, don’t group them yet. Quantity and range over polish.”

Then the follow-ups, in order: “Ideas 1-30 are banned. 20 more with different mechanisms.” → “Score all survivors 1-5 on [your criteria] in a table; show your reasoning in one clause per score.” → “For the top 4: strongest argument against each, and what evidence would change your mind.” Asking the model to show its reasoning per score is basic chain-of-thought practice, you’re not after the number, you’re after the clause you’ll disagree with.

Four more patterns worth keeping

The divergence prompt above is the workhorse. These four pull different levers, and each is a one-line addition to a session:

• Inversion. “List 15 ways to guarantee this project fails.” Then flip each failure into a preventive or opposite move. Models are unusually good at failure enumeration because criticism is heavily represented in what they’ve read, and the flipped list often contains ideas the direct question never surfaces.
• Analogy transfer. “How do emergency rooms handle triage? How do airlines handle overbooking? How do game studios handle beta feedback? For each: what’s the mechanism, and what would the equivalent be for [our problem]?” Pick industries with a structurally similar problem, not a similar surface.
• Persona panel. Instead of one persona at a time, run a panel in a single prompt: “Five people are in the room: a CFO who hates recurring costs, a support rep who reads every complaint, a new customer on day 3, a competitor’s product manager, and a regulator. Each proposes 4 ideas and objects to one idea from someone else.” The objections are frequently more useful than the proposals.
• Constraint ladder. Re-run the same question at escalating constraint levels: unlimited budget, then $1,000, then zero budget and one afternoon. The zero-budget round is where the lazy defaults die, and the unlimited round occasionally surfaces something worth funding properly.

Anchoring: the quiet session-killer

Anchoring cuts both ways in AI brainstorming, and both directions are avoidable:

• You anchor the model. If you share your favorite idea in the setup (“we’re thinking of a referral program, what else?”), most of the output will be referral-program variants. State the problem, not your candidate solution. If you want the model to riff on your idea, do that in a separate conversation after the open round, never before it.
• The model anchors you. The first list arrives fluent and confident, and it’s tempting to treat it as the option space. It isn’t, it’s the center of the distribution. This is why the banned-list round is mandatory, not optional: it’s your structural defense against accepting the model’s defaults as the menu.
• Fresh conversations break residue. Long brainstorming threads drift toward whatever got positive reactions early. For a genuinely new angle, start a new conversation with the same problem statement rather than pushing a tired thread past its 40th message.

Where this pays off most

Anywhere the blank page is the bottleneck: campaign angles and content pipelines in marketing, objection-handling and outreach angles in sales, feature and experiment backlogs in product, process-improvement candidates in operations. It also quietly fixes the worst property of group brainstorms, anchoring on the first loud idea, because the group now starts from a pre-generated field of 30 instead of an empty board and one extrovert.

Where it produces mush, and you should notice

Some problems make even a well-structured session return polished nothing. Recognize these and either change the input or skip the tool:

• Problems the model has no context for. “Ideas to improve morale on our team” without describing the team, the friction, and what’s been tried returns HR-blog boilerplate. If you can’t write a specific problem statement, the session is premature, go gather context first.
• Decisions disguised as ideation. If the real blocker is choosing between two known options, more options make it worse. That’s an analysis job (tradeoff table, pre-mortem on each), not a brainstorm.
• Taste calls. Naming, taglines, and visual direction generate fine as volume exercises, but the model’s ranking of them is close to worthless. Its sense of “good name” is the average of every naming article ever written. Use it for the long list; keep the shortlisting entirely human.
• Domains where you’re the expert. If you’d recognize all 30 ideas instantly, the model is below your floor on that topic. Flip its role: have it attack your ideas instead of generating its own.

The tell in every case is the same: output that would fit any company in your industry. When you see that, the fix is context and constraints, and if you’ve already supplied both, the fix is closing the tab.

Pitfalls

• Stopping at round one. The first batch is the model’s defaults, the same list your competitor gets from the same prompt. The banned-list round is where differentiation starts. If you only do one thing from this page, do that.
• Letting the model pick the winner. Scoring grids, yes; final calls, no. The model optimizes for what sounds good, and it doesn’t know that idea #12 died in committee last year or that #7 is your CEO’s pet peeve.
• Evaluating during divergence. If you critique idea #4 mid-generation, the model reads the room and turns conservative for ideas #5-30. Keep the phases separate, generate everything, then judge everything.
• Treating model claims as research. Brainstorming output is hypothesis, not fact. If an idea rests on a factual claim (“competitors don’t offer X”), verify it, a plausible hallucination makes an exciting, doomed shortlist entry.
• Leading with your pet idea. Mention your candidate solution in the setup and the whole session becomes variations on it. State the problem cold; test your own idea in a separate pass.
• Skipping the problem statement. Ten minutes of prompt iteration can’t rescue a problem you haven’t defined. Garbage in is amplified, not filtered.

Session template

Copy this into a note and run it top to bottom (30-45 minutes solo):

1. Problem statement, goal, audience, deadline, ruled-out options (write it yourself)
2. Context dump, product, constraints, past attempts (paste)
3. Round 1-30 ideas, one sentence each, three categories
4. Round 2, ban all, 20 more, new mechanisms
5. Round 3 (optional), 3 personas × 5 ideas each
6. Converge, score survivors against your 3-4 criteria in a table
7. Critique, top 4: argument against, failure mode, evidence that would change the call
8. Human step, pick, name an owner, define the first concrete action

FAQ

Doesn’t AI just produce generic ideas? Unprompted, yes, the defaults are the training-data average. Volume targets, hard constraints, personas, and banned-list rounds are the mechanics that force it past average. Generic output is almost always a one-round, no-constraint session.

Should AI brainstorming replace team brainstorming sessions? It replaces the blank-whiteboard part, not the session. Generate before the meeting; spend the room’s time on reacting, combining, and committing, the parts that need context and ownership.

How many ideas should I ask for? 20-40 per round, one sentence each. The obvious material fills the first third; you’re paying for the back half.

Can I trust AI to evaluate which ideas are best? As a scorer against your criteria and a devil’s advocate, yes, it’s excellent at structured critique. As the decider, no: it lacks your budget, politics, and history, and it flatters idea-shaped ideas.

Get one practical AI implementation brief per week, join the free newsletter.