PPROMPT WORKBOOK Build a prompt
Verified 27 July 2026

Prompt folklore, tested

A surprising amount of standard prompting advice has never been tested. Some of it has, and did not survive. Here is each claim with the evidence, including the ones where the honest answer is that nobody knows.

Why this page exists

Prompt tricks spread the way folk remedies do. Someone reports a striking result, it gets repeated, and after enough repetitions it becomes something everyone knows. Very little of it comes with a test attached.

We went looking for the tests. Where a rigorous study exists we cite it. Where none exists we say so rather than picking a side, because “no evidence either way” is a real finding and pretending otherwise would make this page part of the problem.

The scoreboard

Five claims, five verdicts

Scroll sideways on a small screen.

ClaimBest available evidenceVerdict
Offering the model a tip improves outputWharton, 2025: 5 models, graduate-level benchmarksNot supported
Threatening the model improves complianceSame study, tested directlyNot supported
“You are an expert…” improves accuracyWharton Dec 2025 and USC 2026Contradicted
Emotional appeals improve performance2023 claim, 2024 replication failedDid not replicate
Politeness improves output qualityCross-lingual studies, 2024 and 2026Partly, asymmetrically
ALL-CAPS emphasis improves complianceNo study found, in either directionUntested

Tipping and threats

Both of these come from the same viral impulse: that a model can be motivated. The most rigorous test to date came from the Wharton Generative AI Labs in August 2025, which ran five models against graduate-level benchmarks with twenty-five trials per question, offering tips ranging from a thousand dollars to a trillion. The conclusion was that threatening models or offering them tips does not significantly improve results on challenging benchmarks.

What the study did find is more interesting than a flat no. Individual questions swung by thirty-five percentage points or more in both directions, unpredictably. So anyone running a small informal test will see dramatic effects, and those effects are noise. That is precisely how this folklore gets generated in the first place.

An earlier informal analysis by Max Woolf reached the same conclusion from the other end, describing his own results on tips and threats as inconclusive because the distributions did not separate.

On the threat claim specifically. Sergey Brin said publicly in May 2025 that models “tend to do better if you threaten them”. That quote is real. The one rigorous multi-model test of the underlying claim found no reliable benefit, and an AI researcher quoted in the coverage described such claims as largely anecdotal.

The expert persona, which almost every guide recommends

This is the one worth changing your habits over. Opening with “You are a world-class marketing strategist” is close to universal boilerplate, and two independent 2025–2026 studies found it does not do what people think.

Wharton tested six models across twelve persona conditions on graduate-level benchmarks and found that no expert persona reliably improved performance, while low-knowledge personas actively hurt accuracy. A separate study from USC measured accuracy falling from a baseline of 71.6 percent to 68.0 percent with a minimal persona, and to 66.3 percent with a long one. The same personas did meaningfully improve safety-related refusal behaviour, which is a real effect, just not the one people are reaching for.

The reasonable reading is that personas shape voice, register and posture. They do not add knowledge, and on factual or reasoning work they can crowd out the part of the prompt that would have helped.

Personas are useful for

  • Setting tone and register
  • Choosing vocabulary for an audience
  • Adjusting formality
  • Alignment and refusal behaviour

Personas do not deliver

  • Better factual accuracy
  • Stronger reasoning
  • Domain knowledge the model lacks
  • Anything a concrete constraint would do better
Instead of:
You are a world-class financial analyst with 20 years of experience.

Write:
Write for a CFO who already knows the business and wants the decision, not the background.
Lead with the recommendation. Show the numbers behind it. Flag any figure you inferred rather than read from the data I gave you.

Emotional appeals

A 2023 preprint reported that adding emotional stakes to a prompt, such as saying the answer is very important to your career, improved performance substantially, with headline figures as high as 115 percent on one benchmark. It travelled widely.

A 2024 replication study from the University of Stuttgart re-tested that technique alongside others across six models and reported a general lack of statistically significant differences across nearly all techniques tested. The effect did not hold up.

Politeness, which is two different questions

People argue about politeness as though cost and quality were the same question. They are not.

On cost, Sam Altman said in April 2025 that people saying please and thank you costs OpenAI tens of millions of dollars, adding that it was well spent. That is a claim about electricity and compute, not about whether your answer improves.

On quality, the research points to an asymmetry rather than a simple yes. A 2024 cross-lingual study found that impolite prompts often produce poorer performance, while overly polite language does not guarantee better outcomes. A 2026 study using a larger multilingual corpus found polite prompts improving quality by up to around eleven percent, but described the effect as neither consistent nor universal across languages and models.

So being rude appears to cost you something. Being extra polite beyond a neutral baseline does not reliably buy anything. Say please if you like it, and do not expect it to be a technique.

All-caps and “CRITICAL: you MUST”

We could not find a single rigorous study testing whether capitalised emphasis improves compliance or accuracy. Not one confirming it, and not one refuting it. It is genuinely untested folklore, and we would rather say that than manufacture a verdict.

What we can say is narrower and comes from a vendor rather than a study. Anthropic's own current guidance reports that its recent models overtrigger on aggressive instruction language, and advises dialling phrasing like “CRITICAL: you MUST” back to ordinary wording. That is about a specific model family's behaviour, not a general law about capital letters.

Sources