What “You are a senior product manager” changes in a prompt

Many agent prompts open the same way: “You are a senior product manager with 15 years of experience.” It reads like handing the model a résumé. Two research teams checked whether the résumé changes the answers.

On questions with one right answer, it did not make the answers more accurate. It does change how they sound. Below: what was measured, what the line is still good for, and what to write under it.

What the studies measured

• 162 roles, 2,410 questions. Five researchers tested 162 roles on 2,410 factual questions across four families of models. Their paper (Findings of EMNLP 2024) concludes that “adding personas in system prompts does not improve model performance across a range of questions” compared with no persona. Choosing the best persona for each question automatically did not help either: the results matched choosing at random.
• Six models, two graduate-level benchmarks. In December 2025 Wharton’s Generative AI Labs ran expert personas on GPQA Diamond and MMLU-Pro. Their result: “persona prompts generally did not improve accuracy relative to a no-persona baseline.” Personas with little knowledge made answers worse. One model on one benchmark gained a little.

Both studies asked questions that have an answer key. Neither tested judgment calls, where two competent people disagree. So the finding is narrow: a persona is not a source of knowledge. For judgment there is a separate argument, and it does not need a benchmark.

What a persona line still does

It sets the register of the answer, and that is worth keeping:

• Vocabulary. “Product manager” gets you roadmaps and trade-offs instead of a generic essay.
• Audience. “Explain it as you would to an engineer” changes what is spelled out and what is assumed.
• Scope. A role tells the model which parts of a broad question to leave alone.

Treat it as a style setting. Keep it to one line.

Why a title cannot carry a decision

Put the same request to two senior product managers: the largest customer wants a custom feature. One builds it, because the account pays for the quarter. The other declines, because the roadmap is the product. Both are senior. The title does not tell you which one you hired.

A model given only the title answers for the whole profession. You get the advice most people in that job would give. If your call is the usual one, that is fine. If it differs, nothing in the prompt says how.

What to write under it

Here is an illustrative prompt that relies on the title:

And one that says how this particular person decides:

The second prompt has four parts the first lacks:

• A priority. Product focus comes before one account’s revenue.
• A threshold. Three customers, not one.
• An exception. An account about to leave changes the answer.
• A question to ask first. The model is told which facts it needs.

None of the four can be guessed from a job title. Each came from a decision somebody made and can explain.

A two-minute test for your own prompt

• Ask a real question from your work with the persona line, then without it. If the two answers differ only in tone, the line carries no decisions.
• Ask a question where you know you depart from common practice. Does the answer follow you, or the textbook?
• Leave out a fact you would need. Does the model ask for it, or answer anyway?

When the list of decisions grows

A handful of priorities and thresholds fits in a prompt. Past that, the question becomes where to keep them. System prompts and past decisions covers what each one carries, and the decision log guide covers how to write the decisions down.

If you would like your agent to start from someone else’s decisions, the Decision Models on avva can be tried with a question before you connect anything.