Turning on Claude's Super Creative Mode
The secret prompt that pros love.
The best answer Claude has is not the one it hands you first.
A recent LLM research paper explains why, and it is the origin of the secret prompt that fixes this. The pros use this prompt daily to get amazing results.
The Why Behind This & the Fix
During training, human raters graded millions of answers, and raters reliably favor familiar text. The model learned the lesson: the safe, expected answer wins.
A paper called Verbalized Sampling, accepted at ICML 2026 (one of the field's top research conferences), put a number on it. Asking a model for a set of responses instead of one raised the variety of its creative writing by 1.6 to 2.1 times and lifted human ratings of the work by 25.7 percent.
Same model. Same knowledge. The only thing that changed was the size of the order.
👉 Takeaway 1: The more responses you ask for in one prompt, the better the batch.
The researchers from this same study got their biggest variety gains by asking the model to attach a probability to each response and to sample from the unlikely end of its own range.
This is the secret prompt that the researchers used:
List each option with the probability you would normally give it, and pull most of the set from the low-probability end.
That one line tells Claude to skip its usual hits and bring back the answers it normally keeps to itself.
👉 Takeaway 2: Use the low-probability add-on to force variety and creativity.
For example, I ran this add-on asking Claude to give me 20 alternative titles for this post. I kept the prompt as simple as possible (Give me 20 better titles). This was to show you that even a basic request can deliver a great variety.
Here’s the prompt:
Give me 20 better titles. List each option with the probability you would normally give it, and pull most of the set from the low-probability end.
This is what I got back for the output. I ended up using the sixth one on the list for my title.
The usual range:
Make Claude Give You Ten Answers, Not One — 85%
Better Claude Answers Start With a Bigger Ask — 70%
Claude’s Best Answer Is Never Its First — 60%
The low probability end:
Force Claude Off Its Default Answer — 50%
The Ask That Doubles Claude’s Creativity — 30%
Turning on Claude’s Super Creative Mode — 28%
Claude Has Better Answers. You Have to Order Them. — 25%
Your Claude Prompt Is Missing One Number — 20%
Claude Gave You Its Safest Answer. Demand the Rest. — 18%
Ask Claude for More Answers, Get Better Ones — 15%
Claude Is Holding Back. Make It Show Everything. — 12%
You Are Asking Claude for Too Little — 10%
Treat Claude’s First Answer as an Opening Offer — 9%
Claude Answers Better When It Answers Ten Times — 8%
Make Claude Compete Against Itself — 7%
Get Claude’s B-Sides, Not Just the Hits — 6%
Claude’s First Answer Is an Audition, Not a Hire — 5%
Order Options From Claude Like a Casting Director — 4%
Make Claude Empty Its Pockets — 3%
Ten Claudes Walk In. You Hire One. — 2%
Ten is a default, not a law
No study has found a magic number of options, so treat anyone who insists on one with suspicion, including me.
The ICML paper’s own example order was five. Other research papers have asked for fifty ideas, or thirty versions of a single sentence.
💡The honest rule is this: The right count is the largest amount you will actually read with attention.
In practice that means three for quick jobs, ten for anything you will publish or send, and thirty or more when you are hunting for one diamond in a brainstorm. Ten earns its place as the default because it is enough spread to escape the average and few enough to read in one screen.
Four moves to pick a winner
The set of ideas arrive. Now the skill is selection.
#1 Skim for the flinch.
Read all ten fast, once. The winner is usually the one that made you pause on the first pass, not the one that survives a spreadsheet of pros and cons. Trust the flinch, then verify it.
#2 Splice the best parts.
Combine the first line of #3 with the close of #7 and show me the result.
You are not required to take any option whole. The strongest drafts are usually seams.
#3 Make Claude defend its picks.
Name your top 3 and argue for each in one sentence. Then name the sleeper: the option I am most likely to skim past that deserves a second look.
The sleeper request pays off more often than it should.
#4 Slate the winner. When #7 wins, order a second slate inside it:
Take #7 and give me 10 variations of it. Same idea, different execution.
The first slate finds the right idea. The second finds the right words for it.