Brainstorm and ideate

Most prompts you'll write are convergent — you know roughly what you want and you're trying to get to it efficiently. Brainstorming is the opposite. The point is divergence — generating more raw material than you'd come up with on your own, so the best of the lot is better than your default.

The model is exceptionally good at this. It has no ego about generating ideas that turn out to be bad — which is exactly the thing that limits human brainstorming. The bottleneck isn't quality; it's quantity. Get the quantity, then judge.

Five patterns. Pick the one that fits the problem.

1. Quantity first

The default mode. Use it when you're starting cold.

"Generate 20 angles for [X]. Don't filter. Don't grade. Don't try to be impressive. Just 20."

"Now critique each in one sentence, then surface the three you'd take into a meeting."

Two messages. The first one is deliberately bad — most of the 20 will be obvious or wrong. That's the point. You can't get to the top three without the bottom seventeen, and the model is much faster than you at producing the bottom seventeen.

When the model resists ("I'm not sure these are all strong…") push past it: "That's fine. Just generate them." Quality comes from the second message, not the first.

2. Reverse brainstorming

Use it when you've already brainstormed and the output is too predictable.

"How would we make this campaign fail? What would guarantee that the new positioning lands flat? How would we design the worst possible distributor scorecard?"

This unlocks ideas your forward brainstorm missed because they sounded too negative or contrarian. After the reverse list, flip each one: "For each item in the list above, write the inverse — the thing we should do to avoid that failure."

You'll usually get sharper ideas from this two-step than from a direct forward brainstorm. It's a known divergent-thinking technique that translates almost perfectly to LLM use.

3. Constraint-based ideation

Open-ended prompts produce open-ended (read: vague) ideas. Constraints unlock creativity.

"How would we run a Q1 awareness campaign with a $0 paid media budget?"

"How would we build a distributor scorecard if we could only show three columns?"

"How would we onboard a new distributor in one week instead of six?"

Pick a constraint that's uncomfortably tight. The point isn't that you'll necessarily live within the constraint — the point is that working within it forces the model (and you) past obvious answers. The constrained version often contains the seed of the version you actually use.

Try three constraints, not one. Time. Money. Headcount. Each surfaces different ideas.

4. Multi-perspective ideation

Use it when the work has multiple stakeholders.

"For the new pricing model: what's the first objection from the CFO? What's the first objection from the VP Sales? What does a frontline rep notice that the executive deck misses? What would a competitor watching us hope we do?"

Different audiences see different things. Asking the model to be each audience surfaces concerns you'd otherwise discover only after launch.

Don't combine: give each perspective its own prompt or its own turn. Mixed prompts produce blended-and-blunted answers; separate prompts produce sharp ones.

5. Stage-gated personas

This is the brainstorm-specific application of the critic pattern from the previous page.

Stage 1 — divergent persona.

"You're a relentlessly creative marketing strategist who has never been told no. Generate 15 angles for [X]. Be bold, be weird, don't self-censor."

Stage 2 — convergent persona, same conversation.

"Now switch. You're a hard-nosed CFO reading the list above. Kill the ones that wouldn't survive a budget review. For the three that survive, write one sentence each on why they make business sense."

The personas matter less than the switch. Brainstorming and editing are different jobs; asking the model to do both at once produces mush. Asking it to do them in sequence produces a usable shortlist.

When NOT to use any of this

• Time-pressured executions. If you need one good draft fast, brainstorm mode is overhead. Use a normal prompt.
• Well-defined deliverables. A monthly scorecard doesn't need brainstorming. Run the skill.
• Anything political or interpersonal. The model is generating plausible-sounding ideas, not reading your team. Don't outsource judgment about people to ideation prompts.

Common mistakes

• Asking for "good" ideas. "Good" filters out everything weird. Ask for ideas. Sort for quality in a second pass.
• Stopping at 5 when you asked for 20. If the model gives you 5 and quits, push back: "That's 5. Give me 15 more." The first 5 are almost always the obvious ones; the next 15 contain the new thinking.
• Combining the divergent and convergent steps in one prompt. "Brainstorm 20 ideas and tell me which is best" produces a soft list and a defensive judgment. Split the steps.
• Filtering too early. Once an idea is in the list, it's free to live there until you do the critique pass. Killing ideas as they appear shrinks the surface area.

Where to go next

• Make it a critic — the critique pass that converts a quantity-first list into a usable shortlist.
• Show, don't tell — when you've found a good idea, paste an example of how you want it written up.
• The four-part prompt — the structure that holds your stage-gated personas.