Skip to content
Prompt Corpus
Technique21 August 20262 min read

"You are an expert" probably makes it worse

The most repeated prompting advice in existence has the weakest evidence behind it, and on knowledge tasks the research says it actively costs you accuracy.

Almost every prompting guide opens with some version of "give the model a role." It is the first thing most people learn and one of the few prompting techniques with a direct empirical test.

The test did not go well for it.

What the research found

Zheng et al. (2023) evaluated persona system prompts across a range of models and questions. The paper's title states the result: When "A Helpful Assistant" Is Not Really Helpful. Adding a persona did not improve factual question answering, and which persona happened to work best was not predictable from the task.

Later work sharpened it into a split worth memorising:

  • On open-ended generation — writing, roleplay, extraction, explanation — expert personas tend to help.
  • On knowledge benchmarks, expert personas damage accuracy. And the more elaborate the persona, the worse the damage.

Why it backfires

The mechanism is not mysterious once you say it out loud.

"You are a world-class expert" selects for text that sounds like a world-class expert. Expert-sounding text is confident. It does not say "I'm not sure about this part."

So on exactly the tasks where you needed calibrated uncertainty, you have shifted the distribution toward confident assertion. You did not make it know more. You made it hedge less.

Worth noting how restrained Anthropic's own documentation is here: a role "focuses Claude's behavior and tone." Behaviour and tone. Not correctness.

What to do instead

Roles are genuinely useful — for the thing they actually do, which is priority ordering, not expertise.

Write the stance, not the credential:

  • Not "You are a senior security engineer with 20 years of experience"
  • But "Review this the way a security auditor would: lead with anything exploitable by untrusted input, treat style as out of scope, and say so if you are unsure whether something is exploitable."

The second buys an ordering rule, a scope boundary, and explicit permission to be uncertain — which is precisely what the confident persona would have suppressed.

And the strongest use of roles has nothing to do with expertise at all. It is audience simulation:

You are the reviewer who recommended rejecting this. Write your review.

That works because it changes what the model looks for, not what it claims to know. Three different hostile readers will find three different problems, and none of them can respond with praise.

The rule

Skip the role entirely on retrieval and classification. It costs tokens and, on the evidence, costs accuracy.

Use it for register, scope and audience — where it does real work.

Never let it imply certainty. If accuracy matters, pair any role with an explicit instruction to flag uncertainty.

And treat output produced under an expert persona as more in need of checking, not less. The persona removed the hedging that would have warned you.

Full lesson: role and persona prompting.

All notes