Prompt Brainstorm

Brainstorm prompts: turning an LLM into a divergent thinking machine

A brainstorm prompt asks a Large Language Model — ChatGPT, Claude, Gemini, Mistral — to behave as a divergent tool rather than a convergent one. Divergent thinking generates many candidate ideas; convergent thinking narrows them down. Most prompts default to convergent (“what is the right answer?”), but ideation needs the opposite. The trick is forcing the model out of its safe, average completion and into territory where the next token is genuinely uncertain — raising temperature to 0.9-1.1 instead of the analytical 0.2, and adding constraints that shrink the solution space and force novelty.

Classic ideation frameworks you can paste into a prompt

• SCAMPER (Bob Eberle) — Substitute, Combine, Adapt, Modify, Put to other uses, Eliminate, Reverse. Seven angles on an existing product.
• Six Thinking Hats (Edward de Bono) — white facts, red emotions, black caution, yellow optimism, green creative, blue process.
• Mind Mapping (Tony Buzan) — central node, radial branches, sub-branches; ideal for visual tools like Miro or Whimsical.
• Lotus Blossom — 8 sub-themes around a core, then 8 ideas around each sub-theme (64 ideas total).
• Random Word Stimulation — force a connection between an unrelated noun and the problem to jump out of the rut.
• Brain Writing 6-3-5 — 6 people, 3 ideas, 5 minutes; adapt to “ChatGPT, simulate 6 personas in parallel”.

A prompt template that works

Generate 20 creative ideas for {topic}.
Apply the {framework} framework.
For each idea, include:
 1) the idea (max 12 words)
 2) why it works (1 sentence)
 3) one potential challenge (1 sentence)
Constraints: weird ideas only, budget < $1k, no SaaS clichés.
Format: Markdown table, columns Idea | Why | Risk.

The constraint line is the secret. Asking for “ideas” with no boundaries returns generic, safe completions — the model is trained to please. Forcing “weird only”, “under $1k” or “impossible ideas to seed thinking” pushes it out of the average. Constraints fuel creativity: this is true for humans and even more so for LLMs.

Multi-agent brainstorm and lateral thinking

Single-pass brainstorms repeat themselves. The fix is to chain prompts and simulate multiple roles. A favourite recipe: “Simulate 5 personas debating {topic}: an investor, a designer, an engineer, a target customer and a contrarian. Each one proposes 3 ideas and challenges the others.” You get 15 ideas already filtered by conflict, which is much closer to what Pixar’s Braintrust, Disney’s Imagineering or IDEO design-thinking sessions do in real life. De Bono’s Lateral Thinking with forced random connections (“link elevator with insurance”) is another reliable jolt out of obvious territory.

Tools that combine AI with visual ideation

• Miro AI and FigJam AI — generate sticky notes from a prompt and cluster them.
• Mural AI Assist — theme summarisation and idea expansion.
• Whimsical AI — mind maps and flowcharts generated from text.
• Notion AI and Mem.ai — idea capture wired into your knowledge base.

FAQ

Can an LLM replace a brainstorming meeting? No. Humans contribute embodied context, politics and emotion that an LLM cannot. Use the model as a pre-meeting accelerator — generate 30 ideas, bring the best 5 to the room.

Which framework is best? It depends on the task. SCAMPER for product variations, Six Hats for decisions, Lotus Blossom for strategy maps, Random Word for breaking creative blocks.

Why does the LLM keep repeating itself? Because temperature is too low or the constraint set is too thin. Raise temperature to 0.9+, add explicit anti-patterns (“no SaaS clichés”) and chain multiple prompts that each forbid the previous list.

When should I stop generating? When you slip from ideation into evaluation. Keep the phases separate — first diverge (quantity, no judgement), then converge (quality, ruthless filter).