15 AI Prompts for Brainstorming That Actually Work

I typed "brainstorm 10 ideas for X" into a chat box more times than I'd like to admit before I noticed I was getting the same answers every time. Not the same words, exactly. The same flavor. Safe, tidy, the kind of ideas I'd have scribbled on a napkin in 30 seconds without any help. For a while I figured that was just what AI brainstorming was.

It isn't. The problem was me. "List 10 ideas" is an instruction to list ten things, and that's exactly what you get — a list. The fix is to stop asking for a list and start forcing the model into a different mode: argue, invert, fail on purpose, think like someone who isn't you. Below are 15 prompts I actually use, grouped by what I want out of the session.

When you're starting cold

1. The 100-ideas dump

Give me 100 ideas for [topic]. Don't filter for quality. Include obvious ones, weird ones, bad ones, and one-word fragments. I want quantity, not curation.

The good stuff doesn't show up until somewhere around idea #40. Every "brainstorm 10" list dies at idea #10 — not because the model ran dry, but because that's where you told it to quit.

2. The constrained version

Give me 20 ideas for [topic] — but constraint: each idea must [fit a specific limit, e.g., "cost under $50", "be doable in one weekend", "involve only existing tools we already have"].

This one surprised me. I assumed fewer rules meant more creativity. It's the opposite. Take away every constraint and you get generic. Give it a tight box and it has to get clever to fit.

3. The "what would [person] do" prompt

Brainstorm ideas for [topic] from the perspective of [specific person: a 12-year-old / a frustrated customer / Marie Kondo / a hedge fund manager / someone who hates this product].

A persona drags the model out of its default house style. The "someone who hates this product" version is weirdly my favorite.

When you have ideas but they're stuck

4. The contrarian

Here are my current ideas: [list]. Argue against every single one. Be specific about why each would fail.

Left to its own devices, AI is a yes-man. It'll tell you your mediocre idea has real potential. Make it disagree and you find out where the body's buried.

5. The remix

Here are 5 ideas: [list]. Combine them into 10 new ideas where each is a fusion of at least 2 of the originals.

Half the time the best idea on the page is just two boring ideas stapled together. This prompt does the stapling for you.

6. The "opposite day"

Here's my idea: [idea]. Now give me the opposite. Now give me what's halfway between them.

You stop staring at one point and start seeing the whole spectrum it sits on. The halfway version is usually the one I end up using.

When you need lateral angles

7. The analogy generator

What's a non-obvious analogy for [problem]? Pick something from a totally different domain — biology, sports, war, cooking, music — and tell me what we can learn from how they solve similar problems.

Cross-domain is where the genuinely novel stuff comes from. The ideas that feel like they came out of nowhere almost always came from somewhere else entirely.

8. The "five whys" excavator

Here's the problem I'm trying to solve: [problem]. Ask "why" 5 times in sequence, going deeper each time, to find the root cause. Then brainstorm solutions for the root cause, not the surface problem.

Most brainstorming swings at the symptom. This makes you keep digging until you hit the thing that's actually causing the symptom, and then it solves that.

9. The constraint inversion

The constraint on my problem is [constraint]. What if that constraint were the opposite? Brainstorm ideas that only work if [opposite constraint] were true.

Half the constraints I treat as fixed turn out to be ones I just made up. This is how I find out which is which.

When you need to evaluate

10. The pre-mortem

Imagine it's 12 months from now and [idea/project] has completely failed. Write the post-mortem. What went wrong? List the top 7 reasons in order of likelihood.

There's something about writing the failure in past tense that surfaces risks you'd never catch while you're still optimistically planning in present tense. I don't fully understand why it works. It just does.

11. The customer skeptic

Pretend you're a customer who just heard about [idea]. List every objection, doubt, and reason you wouldn't buy it. Be specific and harsh.

Same trick as the contrarian, pointed at the buyer instead of the idea. The "be specific and harsh" line matters — drop it and you get polite, useless skepticism.

12. The "what would kill this"

If I launch [idea], what's the single thing most likely to kill it in the first 90 days? Then the next 90 days. Then the next year.

"What are the risks" gets you a vague list. Putting risks on a clock gets you something you can actually plan around.

When you need to communicate the result

13. The elevator pitch generator

Here's my idea: [explanation]. Write 5 different one-sentence pitches — one for an investor, one for a customer, one for a teenager, one for a journalist, one as a tweet.

If you can't boil it down for the teenager, you don't have the idea yet. This is a clarity test disguised as a writing task.

14. The skeptic translator

Here's my idea: [explanation]. Now rewrite the explanation as if you were a skeptic explaining what's wrong with it.

You understand your idea about twice as well once you've watched someone articulate the case against it in plain language.

15. The teach-it-back

Explain [idea] back to me as if I'd never heard it before, in under 100 words. If anything is unclear, ask me to clarify before writing.

The "ask me to clarify" line is doing the real work here. When the model stops to ask a question, it's usually pointing straight at a gap in my thinking, not its own.

How to actually run a session

A single prompt is rarely the whole game. A few things I've learned the hard way:

Don't stop at one. These chain. Start with the dump (#1), grab the most interesting thing it spits out, run it through the contrarian (#4), then send the survivors to the pre-mortem (#10). Layering modes beats any one prompt on its own, every time.

Push back when it goes flat. When you get a limp list, just say "those are all generic, give me 10 weirder ones." The second pass is where it stops playing it safe. I almost never keep anything from a first pass.

Save the chat before you close the tab. Nine of ten ideas will be junk — fine. The tenth is the reason you did this, and the number of good ideas I've lost by not bookmarking the conversation is genuinely embarrassing.

And mix modes inside one session. Run the contrarian right after the dump. The whiplash between "give me everything" and "now tear it all apart" surfaces patterns you'd miss if you stayed in one gear.

Try it

You can run any of these in Smillee AI without signing up. Copy a prompt, drop in your topic, see what comes back. If the first answer's weak, push back — that's not the brainstorm failing, that's where the brainstorm starts.

— Maya