Ask five people to have AI brainstorm ideas for their newsletter and you’ll get back roughly the same list. A weekly roundup. Interviews with people in the niche. Repurposed threads. A beginner’s guide to whatever the topic is.
The model is doing exactly what it was built to do. It predicts the most likely next word given everything it has read, and the most likely answer is the most common one. The most common one is the thing everybody already thought of.
Asking for creativity doesn’t fix it. You get the most predictable version of creative, usually a metaphor about jazz.
What works is making the average answer impossible to give. Eight prompts below, plus seven constraints you can bolt onto anything.
Why it got narrower, not wider
Training on human preference ratings makes models more useful and less varied at once. Kirk and colleagues measured this at ICLR 2024 and found it cuts output diversity on every measure they tested.
Models got better at giving you a good answer and worse at giving you an unusual one. Nothing in the chat interface turns that back up.
Prompt 1: Obvious answers first
What it does: Pulls the twenty most predictable answers out of the model so you can ban them before you start.
When to use it: At the front of any ideation session, before you’ve asked your real question.
The Prompt:
I’m going to ask you about [TOPIC OR PROBLEM]. Before you answer, list the 20 responses you’d most likely give to anyone who asked you this cold, with no extra context about them.
Rank them from most predictable to least. Include the ones you’d be slightly embarrassed to lead with. Don’t answer my actual question yet.
I’m going to ban this entire list before we begin.
How to use it:
Name the real topic, not a sanitised version of it.
Read the list for recognition. You’ve seen most of these before, probably from your own last brainstorm.
Keep the list. You need it in the next prompt.
Example input: I’m going to ask you about growing a B2B newsletter past 5,000 subscribers. Before you answer, list the 20 responses you’d most likely give to anyone who asked you this cold...
What you’ll get: Exactly the ideas you’d have got anyway, which is the point. The value is that they arrive labelled as off-limits instead of arriving as suggestions.
Advanced note: Twenty is deliberate. At ten the model hands you the answers it thinks are good and holds nothing back. Past fifteen it starts scraping the bottom of the obvious pile, and that’s where the ban list earns its keep.
What you’re missing
Seven more prompts, and the swap in thinking that makes them work.
Prompt 2, Ban list rerun: Feeds the twenty back and forces answers that don’t share a mechanism with any of them.
Prompt 3, Constraint stack: Piles four hard limits on at once so the standard answer stops being possible.
Prompt 4, Rare answer probe: Pulls from the uncommon end of what the model would normally produce, with an honest note on why the rarity estimates aren’t real data.
Prompt 5, Cross-domain transplant: Rebuilds your problem using the operating rules of an unrelated field, then names the part of the mapping that breaks.
Prompt 6, Named method over job title: Kills “act as a marketing expert”, which is the single biggest cause of average output, and replaces it with something that has rules.
Prompt 7, Prior art check: Tells you which of your favourite ideas already exist and how obviously.
Prompt 8, Specificity ratchet: Three rewrites, each one stripping out anything that would still work with someone else’s name on it.
Plus seven constraints you can paste onto the end of any prompt to push it off the average, with a worked example of each.
Prompt 1 shows you the problem. The rest of it is the fix.