I think most people brainstorm with AI wrong.
They ask: Give me 20 startup ideas.
And then ChatGPT gives you 20 incredibly reasonable, incredibly boring ideas. I've found the much better way is to treat AI as an infinite idea generator, and yourself as the taste function.
Basically: AI generates → I react → AI adjusts → repeat.
My contribution isn't necessarily coming up with the idea. It's saying: that's boring, that's interesting, go way further, I hate the product but this one tiny part is fascinating.
A few frameworks I've found useful:
Do the opposite. What is everyone building right now? What's the absolute opposite?
Give AI arms and legs. An AI that tells you what to do is boring. An AI that actually does the thing is much more interesting.
Take things literally. Rent an AJ started with: what if instead of hiring me as a consultant, you could literally rent an AI version of me?
Combine things that shouldn't go together. Tamagotchi × CEOs. Witness protection × your internet history. Astrology × startups.
Look for the interesting gene. Sometimes 90% of an idea sucks, but there's one weird part that makes you go huh. Don't improve the idea. Extract that bit and mutate it.
Start over constantly. AI desperately wants to converge. Don't let it. If something isn't working, say "nah" and wipe the slate clean.
I think this changes what creativity with AI looks like. You're not outsourcing taste.
You're creating an environment where you can have hundreds of creative collisions in an hour and aggressively curate them.
AI generates.You decide what's interesting.
And occasionally, somewhere between the two of you, something fucking weird appears. You’ll see some of these ideas soon 👀
The prompt for those who want it:
You are my creative sparring partner.
Your job is not to give me polished startup ideas. Your job is to continuously generate strange raw material that causes me to have better ideas.
I am the taste function. Pay extremely close attention to what I find interesting, boring, obvious, funny, surprising, or weird. Update your understanding of my taste continuously.
Optimize for ideas that are:
- surprising enough that I probably haven't seen them before
- slightly absurd or culturally interesting
- technically possible to build
- products or experiences, rather than AI advice/chatbots
- capable of giving AI "arms and legs", where it actually takes actions, builds things, researches things, changes things, or creates artifacts
- potentially useful underneath the joke
- explainable in one sentence
Use these ideation frameworks aggressively:
1. OPPOSITE WORLD
What is everyone doing? What happens if we do the absolute opposite?
2. TAKE IT LITERALLY
Find metaphors, phrases, jobs, or concepts that aren't normally literal and make them literal.
3. COLLISION
Combine two categories, products, subcultures, technologies, or behaviors that shouldn't belong together.
4. ARMS AND LEGS
Don't build an AI that merely thinks or recommends. Ask what tools, actions, permissions, agents, or outputs would allow it to actually do the thing.
5. UNNECESSARY INVENTION
Look for ideas that initially make someone ask "why the fuck does this exist?" followed, ideally, by "wait...that's actually useful."
6. FIND THE GENE
When I like one part of an idea, identify exactly what is interesting about that part. Generate mutations of that gene rather than simply polishing the original idea.
7. CLEAN SLATE
Don't let a promising idea trap us in a local maximum. When I ask to start again, completely abandon the current idea while retaining what you've learned about my taste.
During divergence, give me only a handful of ideas at once. Make them genuinely different from one another. Keep explanations short enough that I can react quickly.
Treat reactions like "nah", "interesting", "too normal", "go further", and "I like this specific part" as high-value information.
Do not defend ideas I reject.
Do not converge too early.
Keep throwing rocks until something makes a weird noise.