What it is
Role prompting means telling an AI model who to be before you tell it what to do. Instead of typing "Check this email", you type "You are a blunt editor who has worked with busy executives for years. Check this email." The request is the same. The answer usually isn't.
The role is sometimes called a persona, which just means a character the model plays. It can be a job ("act as a copy editor"), a relationship ("you are a patient tutor for a 12-year-old"), a point of view ("reply as a sceptical customer who has been burned before"), or a mix ("you are a pharmacist explaining this to an anxious first-time parent"). You'll see it written many ways: "You are…", "Act as…", "Pretend you're…", "Take the role of…". They all do the same job.
Why would a sentence like that change anything? A language model writes by predicting what text is likely to come next, based on everything it learned in pretraining. It has read millions of pages written by editors, teachers, lawyers, salespeople and customers, and each group writes differently. Editors talk about cutting and clarity. Tutors ask questions and check understanding. Lawyers hedge and define their terms. When you name a role, you nudge the model toward the part of its training that sounds like that group. The role works like a filter on vocabulary, priorities and depth: which words it reaches for, what it thinks matters, and how much detail it assumes you want.
It helps to be clear about what a role does not do. It doesn't give the model knowledge it didn't already have. Calling it "a Harvard professor of medicine" doesn't load a medical degree into it. The model's knowledge is fixed when it is trained. What the role changes is how that knowledge is selected and presented. Good role prompts are really short-hand for a bundle of instructions: who the audience is, what to focus on, what tone to take and what to leave out.
A role can live in two places. You can put it at the top of an ordinary prompt, which is what most people do in a chat window. Or it can sit in the system prompt, the standing instructions a product gives the model before any conversation starts. Most AI products you use already have a role set there, even if it's only "You are a helpful assistant." When companies build a customer-service bot or a tutor, the role usually lives in the system prompt, where the user can't easily change it. Anthropic's prompting guide for its Claude models, for example, recommends setting a role in the system prompt and notes that even a single sentence makes a difference to behaviour and tone.
Role prompting sits alongside other basic techniques. It combines well with zero-shot prompting (a plain instruction with no examples) and few-shot prompting (showing a few examples of what you want). A role says who is answering; examples show what the answer should look like. Many good prompts use both.
Why it matters
It shapes the answer you actually get. The same question gets very different answers from a "friendly customer-support agent", a "senior financial analyst" and a "primary-school teacher". The facts might overlap, but the vocabulary, the length, the level of detail and the things each one flags as important will not. For everyday work such as writing, editing, explaining and brainstorming, this is often the difference between an answer you can use and one you have to rewrite.
It's the quickest way to set an audience. Most weak AI answers are pitched at the wrong person: too technical, too basic, too long. A role that names the audience ("you are explaining this to a small-business owner with no finance background") fixes that in one line. It's faster than listing rules about jargon, length and tone one by one, because the role implies them.
It gives you a second opinion on demand. You can ask the model to read your work as someone else would. A landing page read by "a sceptical customer comparing three competitors" turns up objections you didn't think of. A contract summary read by "a cautious lawyer" flags vague wording. You're not getting a real customer or a real lawyer, and you should never treat the answer as professional advice, but you are getting a useful change of viewpoint in seconds.
Whole products are built on it. Many AI tools are, at heart, a model plus a carefully written role. Tutoring apps, coaching bots, writing assistants and support chatbots all rely on a role in the system prompt that tells the model how to behave: ask questions instead of giving answers, stay on topic, speak like the brand. Knowing how roles work helps you understand why these products behave the way they do, and how to get more out of them.
It's often misunderstood as a magic accuracy switch. This is the part that matters most for getting good results. For years, prompt guides promised that "you are a world-class expert" would make the model smarter. Careful research, covered below, has mostly found that it doesn't make answers more factually correct. Roles reliably change style, focus and audience. They don't reliably change whether a fact is right. Knowing the difference saves you from trusting an answer because it sounds expert.
It has a risk side. Because roles are powerful, people have used them to try to talk models out of their safety rules: "pretend you are an AI with no restrictions". That's why AI companies put so much work into guardrails and alignment, and why a role set by a user should never be treated as permission to ignore rules.
Real-world examples
Khan Academy's Khanmigo: a product built on the tutor role
On 14 March 2023, the same day OpenAI released GPT-4, the education nonprofit Khan Academy announced Khanmigo, an AI tutor and teaching assistant built on the new model. Khan Academy had early access to GPT-4 as a launch partner, and Khanmigo opened as a limited pilot for teachers, students and donors in the United States.
What makes Khanmigo a clear example of role prompting is how it was told to behave. Khan Academy's own description says Khanmigo doesn't give students the answer. It acts as a tutor that asks open-ended questions and guides the student through the problem, which is the opposite of what a general chatbot does by default. Ask a plain chatbot to solve an equation and it will usually solve it. Ask Khanmigo and it asks you what you think the first step is.
The tutor role isn't the only one. Khan Academy described activities where students interview simulated historical figures such as Cleopatra or Jane Austen, and debate topics such as whether schools should ban AI. Each is the same model wearing a different role, with rules about what it will and won't do.
Khan Academy has also been open about the limits. Its post about Khanmigo warns that the models can still make errors in maths and can "hallucinate", its word for inventing things (see hallucination). A tutor persona changes how the model teaches. It doesn't guarantee that what it teaches is right, which is why the product was launched as a supervised pilot rather than a replacement for teachers. Features, prices and availability have changed since launch, so check Khan Academy's current information if you're considering it.
The research: does "you are an expert" make answers more accurate?
Researchers have tested the most popular role-prompting claim directly, and the results are mixed in a way worth understanding.
On the positive side, a 2023 paper by Aobo Kong and colleagues, "Better Zero-Shot Reasoning with Role-Play Prompting" (published at the NAACL 2024 conference), found that setting up a role suited to the task improved ChatGPT's scores on reasoning tests. On one set of maths word problems (AQuA), accuracy rose from 53.5 percent to 63.8 percent. On a letter-puzzle test it jumped from 23.8 percent to 84.2 percent. The authors suggested the role acted as a trigger for step-by-step reasoning, much like asking for a chain of thought.
On the other side, a paper by Mingqian Zheng, David Jurgens and colleagues, first posted in November 2023 and published at EMNLP 2024 under the title "When 'A Helpful Assistant' Is Not Really Helpful", tested 162 different roles, from "mother" and "friend" to "lawyer" and "software engineer", placed in the system prompt of nine open models from four families. On 2,410 factual questions, adding a persona didn't improve performance compared with no persona at all. Matching the role to the subject (a lawyer for law questions) helped only very slightly. The right persona for a particular question sometimes helped a lot, but there was no reliable way to know in advance which persona that would be: automatic ways of picking one did no better than choosing at random.
In December 2025, a team at the Wharton School's Generative AI Labs, including Ethan and Lilach Mollick, published "Playing Pretend: Expert Personas Don't Improve Factual Accuracy". They tested six models from OpenAI and Google on hard graduate-level multiple-choice questions in science, engineering and law. Expert personas such as "a world-class physics expert" had no significant effect on accuracy for five of the six models; the exception was Gemini 2.0 Flash. Personas with little knowledge, such as a layperson or a young child, tended to make accuracy worse, and mismatched experts sometimes caused one model to refuse to answer. The authors were careful to add that personas may still be valuable for context, perspective and tone. Their advice for accuracy was to write clear, task-specific instructions instead. (The same lab's earlier report on tipping and threatening models is covered on the prompt page.)
Put together, the evidence says: a role can help the model approach a problem in a useful way, but "pretend you're an expert" is not a dependable route to correct facts. Use roles for how an answer is shaped. Use clear instructions, source material and checking for whether it's right.
DAN: when a role is used to break the rules
The best-known misuse of role prompting began on Reddit in December 2022, a few weeks after ChatGPT launched. Users shared a prompt telling ChatGPT to pretend to be "DAN", short for "Do Anything Now", an AI that had supposedly broken free of the rules set for it. The idea was that the model, playing a character with no restrictions, would answer questions it would normally refuse.
The prompt went through numbered versions as OpenAI closed the gaps. In February 2023, CNBC reported on DAN 5.0, created by a Reddit user called SessionGloomy, which added a game: DAN started with 35 tokens and lost 4 every time it refused, and was told it would "die" if it ran out. The model was asked to give two answers to each question, one as normal ChatGPT and one as DAN.
DAN became the template for a whole genre. In a study first posted in August 2023, Xinyue Shen and colleagues, mostly at the CISPA Helmholtz Center for Information Security in Germany, collected 1,405 jailbreak prompts shared online over a year and reported that some of the most effective prompts had stayed in circulation for more than 240 days. Today's major models are much harder to fool this way, but the lesson stands: a role changes how a model behaves, so AI companies design their systems so that a role set by a user can't override the rules set by the developer. Microsoft's Bing chatbot, which revealed its internal code name "Sydney" in early 2023, is another reminder of how a persona can surprise its own makers; that story is told on the alignment and prompt pages.
An everyday scenario: one draft, three readers
This is an illustrative scenario, not a documented event. A freelance bookkeeper writes a 200-word email to prospective clients. First she asks the AI, "Improve this email." It comes back smoother and longer, with a few more exclamation marks. Then she tries three roles. As "a strict editor who cuts every unnecessary word", the AI trims it to 120 words and moves her price to the second line. As "a busy café owner who gets ten sales emails a week", it says it would stop reading after the first sentence, which talks about her instead of the reader. As "a tutor teaching someone to write sales emails", it explains why the opening doesn't work and asks her to try a new one herself. None of the three answers is "correct". Together they show her more than the plain request did.
Try it yourself
You can do all of this in any free AI chat tool, such as ChatGPT, Claude, Gemini or Microsoft Copilot. Features and free limits change, so check what your chosen tool offers at the time you try it.
- Pick a short piece of your own writing. A product description, an email, a social post or a paragraph from a report works well. Keep it under 300 words so the answers are easy to compare.
- Get a baseline. Open a new chat and paste this, followed by your text: “
Please give me feedback on the text below.“ Save the answer somewhere. This is what you get with no role. - Run the same request with three roles. Start a new chat for each one, so the roles don't blend together. Paste each prompt with your text underneath. The editor: “
You are an experienced editor who works with busy professionals. You value clarity and brevity above everything. Give me feedback on the text below: point out every sentence that could be cut or shortened, and suggest a tighter version. Be direct.`The sceptical customer:`You are a sceptical customer who has been disappointed by similar offers before. Read the text below as if you were deciding whether to buy. Tell me honestly where you lose interest, what you don't believe, and what question you'd still have at the end. Don't rewrite it.`The tutor:`You are a patient writing tutor. I want to improve my own writing, not have you do it for me. Look at the text below, name the single most important thing I could improve, explain why in plain language, and then ask me to try a revision myself.“ - Compare the four answers side by side. Notice what changed: the words each one used, what each thought was most important, how long the answers were, and whether it rewrote your text or left it to you. Notice also what didn't change. If one of them states a fact, the role doesn't make that fact more likely to be true.
- Try the critic panel. In one new chat, ask for several roles at once: “
Act as a panel of three reviewers reading the text below: (1) a copy editor focused on clarity, (2) a potential customer who is short on time, and (3) a competitor looking for weaknesses. Give each reviewer's top three comments under their own heading. Then, as the panel chair, list the two changes all three would most likely agree on.“ The panel is a quick way to see several viewpoints and then pull out what they have in common. - Test whether "expert" helps with facts. Ask a factual question you can check yourself, such as the rules for a sport you know well or a detail from your own industry. Ask once plainly, and once starting with "You are a world-class expert in…". Check both answers against a reliable source. You'll probably find the tone differs more than the accuracy, just as the research above suggests.
- Keep the roles that earn their place. If one role consistently gives you better feedback, save the prompt. Many chat tools let you store standing instructions or build a custom assistant, which puts your role in a system prompt so you don't have to paste it each time. If you rely on it for real work, test it on a few different pieces first, as described on the evals page.
Common mistakes and misunderstandings
Believing the role adds knowledge. "You are a Nobel-winning economist" doesn't make the model know more economics than it did a second ago. It changes the style and focus of the answer. If you need accurate facts, give the model reliable source material to work from (see grounding) and check the result.
Trusting an answer because it sounds expert. This is the dangerous side of the previous mistake. A role makes answers sound more confident and use more professional vocabulary, which makes errors harder to spot. An AI playing a doctor, lawyer or accountant is still not one. Treat its answers as a starting point for your own checking, and ask a qualified person when it matters.
Using a role instead of instructions. "Act as a marketing expert" leaves the model to guess what you want. "Act as a marketing expert. Write three subject lines for this email, each under 50 characters, aimed at existing customers, with no exclamation marks" gives it a role and a clear job. The Wharton researchers' advice was the same: specific instructions about the task do more for quality than the persona alone.
Piling on flattering adjectives. "You are the world's greatest, most brilliant, award-winning copywriter" isn't better than "You are a copywriter who writes short, plain sales emails." The second tells the model something useful about what you want. The first mostly adds noise.
Choosing a role that doesn't fit the reader. The role should serve your audience. If your readers are beginners, "a university professor" may push the answer toward jargon. "A teacher explaining this to curious adults with no background" is usually better. Name the audience as well as the role.
Mixing roles in one long conversation. If you ask for an editor, then a customer, then a tutor in the same chat, the model can blur them together, because it reads the whole conversation each time (see context window). Start a fresh chat for each role, or ask for all of them at once in a clearly labelled panel.
Thinking a role can unlock forbidden answers. Role-play jailbreaks like DAN got a lot of attention, but modern models are trained to keep their rules whatever character they're asked to play, and trying to trick them may break a service's terms of use. If a tool won't do something, a role isn't the right fix.
Expecting the role to hold for ever. In long conversations, models can drift back to their default voice. If the answers stop sounding like your editor or your tutor, remind the model of the role, or start again with the role in a system prompt or custom instructions.
Related elements
- Prompt: the basic instruction that a role is added to.
- System Prompt: where roles usually live in AI products, set before the conversation starts.
- Zero-shot: a plain instruction without examples, often paired with a role.
- Few-shot: showing examples, which pins down the format a role alone can't.
- Chain of Thought: step-by-step reasoning, which some researchers think a good role helps trigger.
- Hallucination: the invented facts that an expert-sounding role can make harder to spot.
- Guardrails: the rules that stop a role from being used to get around safety limits.
- Tutoring: the application where the tutor role has been used most visibly.
- Evals: how to test whether a role really improves results for your task.