What it is
A role prompt opens with “Act as” or “You are”: you are a senior marketer, act as an HR advisor, you are a world-class copywriter. It is one of the oldest prompting habits, and it does something real. A role nudges the vocabulary, the tone and what the model pays attention to. Told it is a contracts lawyer, a model reads a contract looking for liabilities rather than summarizing it.
What a role does not do is add knowledge. The model knows what it knows whether or not it is told it is an expert. A study of 162 personas across 2,410 factual questions and four popular model families found that adding a persona to the system prompt did not improve accuracy compared with no persona, and that the effect of any one persona was largely random (“When ‘A Helpful Assistant’ Is Not Really Helpful”, Findings of EMNLP 2024). The models tested are older than today’s, so read it as a caution rather than a law. Still, “world-class” is not a fact the model can use.
The useful parts of a role can be said more directly. Instead of “act as a lawyer”, say who will read the answer, such as a bakery owner who is not a lawyer, and name the expertise to apply: the clauses a small-business contracts lawyer would check first. That keeps the lens and drops the costume.
When it helps
Role-play is where a role earns its place. Practicing a renewal call, a tough interview or a difficult conversation needs the model to stay in character as the other person, and that is exactly what a role does.
A short role line also helps in a standing assistant, such as a system prompt, a Custom GPT or a Claude Project, where it sets a consistent voice across many conversations. Anthropic, OpenAI and Google all describe a role or identity line there, as the last section shows, and Anthropic says even a single sentence makes a difference.
And a role can be a quick way to pick a lens, when you do not know which checks an expert would make. “Review this as a payroll auditor would” is better than nothing. Better still, ask the model what a payroll auditor would check, then put that list in the prompt.
When it does not
Flattery adds nothing. “World-class”, “genius” and “with 30 years of experience” cost tokens and do not change what the model knows.
A role is not a substitute for facts. “You are our head of sales” does not tell the model your prices, your customers or last quarter’s numbers. Give it the facts.
A role can pull against the reader. Ask for a lawyer’s review and you may get a lawyer’s language, which is the wrong register for someone who is not one. Name the reader and the language follows them instead.
And a role does not make the answer professional advice. A model told it is a doctor, a lawyer or an accountant is still a model, so treat its answer as a starting point for the conversation with the professional, not a replacement for it.
Before and after
Three role prompts from different kinds of work: two where the role is better replaced, and one where it is the point. The scores are the Prompt Grader’s free estimate with Written for set to Any chat model. A score counts what the prompt states; it does not grade the answer.
Act as an expert lawyer. Review this contract. """ [contract] """
Supplier contract: """ [contract] """ Review the contract above for me, the owner of a 12-person bakery, so I can decide what to push back on before I sign. I am not a lawyer, so explain each point in plain English. Read it with the knowledge of a commercial contracts lawyer who works for small businesses: look at automatic renewal, price increases, minimum orders, termination and liability. Leave out standard clauses that carry no real risk. Give a numbered list of up to six clauses to push back on, most important first, each with the clause number, the risk in one sentence, and what to ask for instead. If a clause depends on state law, say so rather than guessing.
The role prompt gets a formal clause-by-clause commentary written for another lawyer. The second keeps the lawyer’s lens, names the five things to check, and writes for the owner, so the answer is a short list of what to ask the supplier to change.
You are a world-class marketing expert. Write a LinkedIn post about our new feature.
Write a LinkedIn post announcing our new feature: automatic overtime alerts for restaurant managers, which warn them before a shift pushes someone into overtime. Readers are restaurant owners and general managers who scroll LinkedIn between services. They care about labor cost and do not care about how the feature is built. Write it the way an experienced B2B marketer would: open with the problem in the reader's words, give one concrete example with a number, and end with a single call to action to book a demo. Format: one short paragraph, then the call to action on its own line. It should read like a person talking to peers. Under 120 words, no hashtags, no emojis. If you need a figure I have not given, leave a [NUMBER] placeholder.
“World-class marketing expert” gets an upbeat post full of hashtags that could announce any feature. The second gets a post written to restaurant managers about labor cost, with the marketer’s moves spelled out and a [NUMBER] placeholder instead of an invented statistic.
You are a customer. Let's role-play.
I am a new account manager preparing for a renewal call next week, and I want to practice so I can handle pushback calmly. Play the customer: the operations director of a 200-person logistics company whose contract renews next month. They think our price rose 12% without new value, and they have a quote from a cheaper competitor. Stay in character. Your objections should be realistic for a busy, skeptical director. Raise one objection at a time and wait for my reply. Do not coach me during the role-play. If I ask about something this brief does not cover, invent a plausible detail and stay in character. When I type "stop", step out of the role and give me feedback as three bullets: what I handled well, what to change, and one phrase to use next time.
The bare role-play gets a friendly, vague customer who gives in quickly. The second gets a specific director with a reason to leave and a competing quote, who pushes back one point at a time, and feedback you can act on when you stop.
The line to add
In place of “act as an expert”, say who reads the answer and which expertise to apply:
This is for {{reader}}, who {{what they know and what they will do with the answer}}.
Apply the knowledge of {{kind of expert}}: check {{the things that expert would check first}}.
For practice, keep the role and give it a situation, a rule for turns and a way out:
Play {{who}}, who {{their situation and what they want}}. Stay in character, raise one point at a time, and wait for my reply. When I type "stop", step out of the role and give me feedback in three bullets.
In ChatGPT, Claude, Gemini and Copilot
The vendors treat a role as a short setting for behavior, not a source of knowledge. Anthropic says a role in the system prompt focuses Claude’s behavior and tone, and that even a single sentence makes a difference (Give Claude a role). OpenAI’s guide suggests a developer message that opens with an identity section describing the assistant’s purpose, communication style and high-level goals (Prompt engineering). Google advises putting role definitions, along with behavioral constraints and format rules, in the system instruction or at the very start of the prompt (Prompt design strategies).
Microsoft’s guide to Copilot prompts breaks a prompt into goal, context, source and expectations (Get started writing prompts). The reader belongs in context and the expertise in expectations, which is the same advice as above in Microsoft’s terms.
In all four, a role line in a chat message works the same way it does in a system prompt: it shifts the voice. The reader and the checks are what change the substance.
Key takeaways
- A role changes voice and lens; it does not add knowledge or accuracy.
- Drop flattery such as “world-class”; it gives the model nothing to use.
- Say who the reader is, and name the checks an expert would make.
- Keep roles for role-play and for a one-sentence voice in a standing assistant.
- A role does not turn an answer into professional advice.
Questions
Does telling ChatGPT to act as an expert make it more accurate?
Not on its own. A study of 162 personas and 2,410 factual questions found that adding a persona to the system prompt did not improve accuracy compared with no persona. A role changes the tone and what the model focuses on. To change the substance, give it the facts, say who the reader is, and list what an expert would check.
What is role prompting?
Role prompting means opening a prompt with a persona, such as “You are a financial analyst” or “Act as a recruiter”. It sets the voice and the lens of the answer. Vendors suggest a short role line in a system prompt for a consistent voice, and it is essential for role-play, such as practicing a sales call.
Should a system prompt start with “You are”?
It can. Anthropic, OpenAI and Google all describe a short role or identity line at the start of a system prompt. Keep it to one sentence, and spend the rest of the prompt on what matters more: who the users are, what the assistant should do, the rules it follows and what to do when it is unsure. The system prompt guide covers the rest.
Prompts to use with this
- Answer a customer question from your help docs and policyYou need an accurate, friendly reply that only uses what your documentation actually says.