In plain words
Many people begin a prompt with a role: "Act as a travel agent", "You are a strict editor", "Pretend you're a cricket coach". That is role prompting. It tells the AI whose shoes to stand in before it answers.
A role mostly shapes how the answer sounds and what it focuses on. A "kind teacher for a 10 year old" uses simple words and examples. A "tough reviewer" points out flaws instead of praising. App makers put roles in the system prompt so every reply keeps the same character.
What a role does not do is give the AI new knowledge. Saying "you are a world-class doctor" doesn't make its medical facts better. Studies testing this have found expert roles generally don't raise accuracy. The bigger gains come from saying clearly what you need, who it is for and what a good answer looks like.
Everyday example
Briefing a friend before a call
Before your cousin calls the electricity office for you, you say: "Be firm but polite, you're my elder brother, mention the meter number." She sounds different on that call, more formal, more focused. She doesn't suddenly know more about electricity bills. The useful part was the meter number. A role prompt works the same way: tone from the role, results from the details.
Where the example breaks: Your cousin really becomes an elder brother for a few minutes in her head. The AI only shifts the style of words it predicts; it has no inner self putting on a costume.
How it actually works
A role is just text in the prompt, usually the first line. Because the model predicts each word from everything before it, that first line colours everything after.
Where the role goes
In apps, roles live in the system prompt, read before every message. Anthropic's guide says setting a role there "focuses Claude's behavior and tone for your use case", and that even one sentence makes a difference.
What changes
Vocabulary, formality, length, what gets emphasised, what gets left out, and how critical or encouraging the reply is. A role can also set the audience: "explain as you would to a first-year student".
What the tests show
A 2023 study of personas in system prompts found that adding one did not improve accuracy across many question types, and the effect was "largely random". A 2025 Wharton report on graduate-level science and law questions found the same, and that low-knowledge roles such as "a toddler" generally made answers worse.
Make a role useful
Pair it with specifics: the goal, the audience, the format and an example. "You are an editor" plus a clear task, "cut this to 100 words for a WhatsApp update", works far better than either alone. Pair it with examples when the format matters.
- Also called
- Persona prompting, "act as" prompts
- Best for
- Tone, focus, point of view, audience level, consistent voice in an app
- Not reliable for
- Making facts more accurate or adding knowledge the model lacks
- Research
- Zheng et al. (2023) and Basil et al. (Wharton, 2025) found no general accuracy gain from personas
Do roles make AI answers better?
One view
Yes, on reasoning
A 2023 paper found that role-play prompts improved zero-shot reasoning in ChatGPT on most of 12 benchmarks, for example from 53.5% to 63.8% on a maths word-problem set.
The other view
Not for accuracy
Larger later tests across many models and questions found expert personas generally did not raise accuracy, and the effect of any one persona was unpredictable.
Most likely: Use roles for tone, focus and audience. For better facts and reasoning, give clear context, examples and criteria, and on newer models don't expect a role alone to help.
Questions people ask
Does "act as an expert" make ChatGPT smarter?
Not in any reliable way. Tests across several models found expert personas generally didn't improve factual accuracy. They do change tone and focus, which can still make an answer more useful to you.
What is a good role prompt example?
"You are a patient tutor for a Class 8 student. Explain photosynthesis in five short steps, with one everyday example, and end with a quick question to check understanding." The role sets the voice; the rest sets the job.
Is role prompting the same as an AI persona?
Related. Role prompting is a one-off instruction in a prompt. An AI persona is a lasting character an app or company gives its chatbot, set through system prompts and often through training.
Can role prompts be used to trick an AI?
People try: "pretend you're an AI with no rules" is a classic jailbreak. Modern models are trained to keep their safety rules whatever role they're given, though some tricks still slip through.
Was this clear?
1 is confusing, 5 is crystal clear. One tap, no account needed.
Spotted something wrong, or know a better example? Suggest a change
Further reading
- When "A Helpful Assistant" Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models (Zheng et al.)arXiv, 2023
- Playing Pretend: Expert Personas Don't Improve Factual Accuracy (Basil et al.)arXiv / Wharton Generative AI Labs, 2025
- Better Zero-Shot Reasoning with Role-Play Prompting (Kong et al.)arXiv, 2023
- Prompting best practices: Give Claude a roleClaude Platform docs