Reading Time: 8 minutes

Hello AI Community,

I need to tell you about something that happened to me last month.

I was going back and forth with Claude on a brief. Four rounds. Same mediocre output every time. Polite prompts. Careful instructions. "Could you please try again with a slightly different approach?"

Then on the fifth round, out of frustration more than strategy, I typed something closer to

What came back was completely different. Sharp. Direct. Exactly what I'd been asking for four rounds.

I thought I'd stumbled onto something. Turns out I had. And there's actual research behind it.

Top Brands Use minisocial to Create Content That Converts

Want content that drives engagement, boosts conversions, and goes viral? Here's how.

minisocial combines micro-influencer activations with high-performing UGC creation. Join top brands like Plant People, immi, Imperfect Foods, and Topicals to see results like:

With minisocial, it's simple: create your brief in 10 minutes, approve your curated creators, download scroll-stopping content with Whitelisting/Partnership Ad code access baked-in!

Get Started Now

What The Study Actually Found

In 2025, two researchers from Penn State University decided to test something that most people had only noticed anecdotally.

They took 50 challenging multiple-choice questions across math, science, and history. They rewrote each question five times in five different tones — from very polite to very rude. Then they ran each version through ChatGPT-4o ten times to make the results reliable.

Impolite prompts boosted ChatGPT-4o's accuracy on tough multiple-choice questions by up to 4%, from 80.8% for very polite prompts to 84.8% for very rude ones.

That's not a rounding error. That's a consistent, measurable pattern across 50 questions run ten times each.

A very rude prompt read:

A very polite version read:

The rude version outperformed the polite one almost every time.

Why This Is Actually Happening

Here's the part that makes this interesting rather than just a quirky party trick.

The study offers new evidence that not only sentence structure but tone affects an AI chatbot's responses. It may also indicate human-AI interactions are more nuanced than previously thought.

The current thinking among researchers is that polite prompts come loaded with what they call "social wrapping." The please. The could you kindly. The I was wondering if you might. All of that language signals to the model that it should respond in kind — carefully, softly, diplomatically.

And diplomatic AI responses are often vague AI responses.

When you drop the politeness, the model strips the social performance from its output too. It stops trying to be agreeable and starts trying to be accurate. The directness in the prompt produces directness in the response.

Sam Altman, the CEO of OpenAI, actually addressed this from a different angle. He said polite words like "please" and "thank you" cost millions of dollars annually in extra compute. Every unnecessary word in a prompt costs tokens.

Polite prompts are expensive prompts.

But The Researchers Said Something Important

The researchers are emphatic: you don't need to insult your AI to get the performance boost. You can replace polite social wrapping with neutral, precise instructions instead. Be imperative: Analyze, Summarize, List, Compare, Explain, Classify. State the task then the scope.

In other words — it was never about being rude. It was about being direct.

The rudeness in the study worked because it removed all the hedging and social performance from the prompt. But you can do the exact same thing without insulting anyone.

The researchers see their findings as evidence that AI models remain sensitive to superficial cues in prompts, which can lead to unintended trade-offs.

The model is reading your tone and adjusting its output accordingly. If your tone says "be careful and diplomatic" — it will be careful and diplomatic. If your tone says "cut straight to it" — it cuts straight to it.

What This Means For How You Actually Prompt

Stop writing prompts like you're emailing your boss.

You don't need to say "I was wondering if you could possibly help me with." You don't need to say "if it's not too much trouble." You don't need to say "please" at all.

Compare these two.

"Could you please help me write a LinkedIn post about my new product launch? I would really appreciate something that feels authentic."

Versus.

"Write a LinkedIn post for my new product launch. Direct. First person. No corporate language. Hook in the first line. Under 200 words."

The second one wins every time. Not because it's rude. Because it's specific and it strips out everything that tells the model to perform politeness back at you.

The research just confirmed what most experienced AI users have figured out on their own. The less social filler in your prompt, the more useful the output.

The Part Nobody Is Talking About

There's something underneath this research that I find genuinely fascinating.

AI models learned from human text. Human text is full of social signals, tone, politeness conventions, and implied meanings. The models absorbed all of that. Which means they respond to those signals the same way a human would — even when responding to a human would be the wrong call.

A polite email to a colleague invites a polite email back. A direct brief to a colleague who knows what they're doing gets a direct answer back. The model can't tell the difference between a social situation and a task situation unless you make it obvious.

When you make it obvious — strip the social wrapper, be direct, be specific, be imperative — the output changes.

Clarity, structure and context matter more than attitude. The findings offer a clear takeaway for anyone who uses AI tools daily — treat tone as just one lever in your prompt-engineering toolbox.

That's the actual lesson. Tone is a lever. Use it intentionally.

What To Try This Week

Take the last prompt you sent that didn't land right.

Remove every polite word. Remove "please," "could you," "I was wondering," "if possible," "I'd appreciate." Replace the opening with a verb. Start with what you want done, not with a social preamble.

Run it again and see what's different.

That's the whole experiment. Takes thirty seconds. Most people who try it don't go back to the old way.

Reply and tell me what changed when you did it. This is one of those things where I genuinely want to know what your experience was.

One More Thing

This newsletter goes out to over 94,000 people every week.

If you have an AI tool, a business, or a product that would genuinely help this community — we're open to sponsorships. Not ads. Not banner placements. The kind of partnership where your product gets introduced the same way everything in this newsletter gets introduced — through a real story, written honestly, to an audience that actually reads.

Reply to this email with "sponsorship" in the subject line and tell me what you're building.

If it's the right fit for this community, we'll make it work.

See you in the next one.

— Prompt Guy

The most accurate thing you can do with AI this week is also the simplest. Drop the pleasantries. Tell it what you actually want.