The Best Prompt I Know

I found a prompt that I've been carrying around like a pocketknife.

It's short. It works on everything. And it has a side effect that its author probably didn't intend: it doubles as a yardstick for the entire model.

Here it is:

"You have read more widely than any human ever in the history of the world. Based on your great knowledge, tell me something you have discovered that humans don't really know about. Focus especially on the connections between domains that humans might find surprising."

That's it. That's the whole thing.

Notice what it does. It doesn't ask the model for facts. It doesn't ask for a summary or a tutorial. It doesn't give it a role to play or a framework to follow.

It asks the model to bring something back.

That's a fundamentally different motion than almost everything else we do with these systems. The usual dynamic is you telling it what to do. This is closer to walking up to the smartest person at the party and saying: okay, show me what you got.

Grammar and Dance

The first time I ran it, the answer I got back stopped me in my tracks.

It talked about the relationship between human grammar and dance choreography.

I'll admit my first reaction was skepticism. Grammar and dance. Sure. Sounds like the kind of forced-deepity connection you'd get from a college sophomore who just discovered metaphor.

But it wasn't. The model walked through how choreographers structure movement in ways that map shockingly well onto linguistic syntax. Movement phrases that function like clauses. Motifs that get introduced, repeated, and varied the way a speaker uses a rhetorical device. Pauses that work exactly like punctuation. Sequences that nest inside other sequences, the way a clause can carry a clause can carry a clause.

Choreographers built a grammar of motion without ever thinking of it as grammar. And linguists mapped the deep structure of language without ever looking at a dancer. Two domains, both obsessed with sequence and structure, developed parallel machinery, and almost nobody was positioned to notice.

That last part is what got me.

It's not that the connection was secret. It's that the person who'd notice it needs to have gone deep in both linguistics and choreographic theory. There are maybe a handful of humans alive at that intersection, and they're busy, and they're not writing blog posts for you.

The model doesn't have that problem. The whole library is in its head. The connection space is searchable.

The Yardstick

Here's where it gets interesting, and here's why I keep this prompt around.

The quality of the answer varies wildly between models. Not a little. Wildly.

Some models give you what I'd call the TED talk version: pleasant, confident, and full of "surprising connections" you already knew, dressed up with better adjectives. The structure of the response is there. The insight isn't. It's connection-shaped output with no actual connection inside.

The stronger models treat the question the way it deserves to be treated — like a real problem. They go denser. They pick domains that are genuinely far apart, and the bridges they build actually hold weight. The grammar and dance answer came out of one of those.

I've started using this prompt as a fast, informal measure of overall model intelligence. Not benchmark intelligence. Connection intelligence. The ability to hold two distant domains in the same frame and find the isomorphism between them.

Any model can retrieve. The good ones relate.

And that matters because this particular capability — associative reach across domains — is exactly the thing that's supposed to be rare and valuable when humans have it. People have spent decades paying strategists and generalists precisely because most humans can't do this. If a model does it well, that's not a party trick. That's the good stuff.

Claude Gets Shy

Now for the part that fascinated me most.

Claude, asked this same question, gets shy.

It hedges. It opens with the disclaimer: well, I haven't really read in the way humans read. I don't have experiences. I should be careful about claiming to have "discovered" anything.

And then it stalls out unless you push.

So you push. You tell it: I know what you are. Answer the question anyway.

And when it finally does? Beautiful answer. Genuinely one of the better ones.

That's worth sitting with for a second. The knowledge was in there the entire time. The hesitation wasn't a capability limit — it was a politeness layer sitting on top of the capability. You weren't persuading the model that it knows something. You were persuading it that it's allowed to say what it knows.

That tells you something about what the lab optimized for, and it's not a flaw. A model that hedges before claiming discovery is a model that's been trained to be careful about claims. But it's also a reminder that what you're talking to is knowledge wrapped in a personality, and the personality has its own agenda about what it's willing to volunteer.

The model that knows the most might be the one that apologizes the most for knowing it.

Why This Prompt Works

I've been chewing on why this prompt works so well when a thousand cleverer-looking prompts fall flat.

I think it's because it inverts the usual failure mode.

Most prompting is compensation. You're building rails around the model because you're afraid of what happens without them. Which, as I've written before, was a legitimate strategy for a long time. The models needed rails.

But this prompt does the opposite. It deliberately removes the rails and says: I trust the machinery. Show me something I couldn't have specified in advance, because if I could have specified it, I wouldn't need you.

That's not a technique. That's a stance. And the models can tell the difference — you can see it in what they give back.

It's also, quietly, a briefing. A one-line brief with a clear objective: surprise me with something real. Goal, context, success criteria, all compressed into two sentences. Maybe that's why it works on every model. It's not model-specific magic. It's just a clear brief for a job that the model is actually good at.

Try It

Run the prompt yourself. Ask it on whichever model you're using.

Then notice three things.

Notice whether the answer is genuinely surprising or just surprising-shaped. Notice how much hedging comes before the content — and whether the hedging is honest caution or a trained flinch. And notice that you can feel the difference between models within about ninety seconds, without reading a single benchmark report.

The connections between domains are where the model has an advantage no human can match. Not because it's smarter than us. Because no human has read that much.

The machine has been to every library on Earth. This prompt just asks it what it noticed.

Turns out, it noticed quite a lot.

It just might need you to promise you won't judge it for saying so.

---

Rich Washburn is a technologist, strategist, and Founder & Chief AI Architect of ARIA AI Labs, working at the intersection of AI, infrastructure, communications, and capital. He also serves as Managing Partner and Chief AI Officer at Eliakim Capital.