Glossary
Role prompting
In short
Role prompting is a task in which you first give the AI a role, such as lawyer or editor, so that it answers in the tone, vocabulary and depth you need.
Prompts · role prompting
MMOOZZGG.com editorsChecked: 1 min read
For a child
When you play hospital with friends, one becomes the doctor and another the patient. The one playing the doctor starts talking differently right away: asks where it hurts and advises tea with honey. You can also tell an AI “you are a doctor” or “you are a teacher”, and it will start answering in a similar way. But that does not make it a real doctor.
In plain words
An AI can write in any style, and you need to tell it which one. To do this, write at the start of the request who it should be: “You are an experienced accountant” or “You are the editor of a children’s magazine”. The role decides the words, the tone, what the model explains in detail and what it skips. The accountant will answer about entries and reporting deadlines; the editor will retell the same thing in short phrases. The role changes the delivery, but it does not add knowledge the model lacks.
For business
Companies use a role so that the AI speaks in the brand’s voice: a polite support agent, a strict analyst for the finance team, a mentor for new staff. It costs almost nothing and takes one line in the system instruction. Know the limit in advance: the study by Zheng et al. (EMNLP 2024) found no accuracy gain on factual questions from a role alone, and the effect of different roles was hard to predict. So the role is responsible for style, while accuracy comes from data and checks. Before launch, compare answers with and without the role on real requests and keep the version clients rated higher.
For a specialist
Role prompting, also called persona prompting, sets a persona for the model in the system message or at the start of the user request. The role shifts the distribution of continuations toward texts typical of that persona in the training data: vocabulary, register, depth and the order of reasoning change. Kong et al. (2023) proposed role-play prompting and reported higher zero-shot accuracy for ChatGPT on most of twelve datasets, for example from 53.5% to 63.8% on AQuA. Zheng et al. (Findings of EMNLP 2024) found no gain from personas on a set of factual questions compared with a control without a role; automatic persona choice was often no better than random. Anthropic’s documentation recommends setting the role in the system message to control behavior and tone. The practical conclusion: a role helps with style, and its effect on accuracy should be tested on your own sample.
Example
A lawyer asks the model to review a supply contract and starts with a role: “You are a lawyer who protects the buyer’s interests.” The answer is built around the buyer’s risks from the start: penalties, acceptance terms and the return procedure.
Related terms
Questions and answers
Sources
- Kong et al., Better Zero-Shot Reasoning with Role-Play Prompting, 2023arxiv.org
- Zheng et al., When "A Helpful Assistant" Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models, Findings of EMNLP 2024arxiv.org
- Anthropic, Prompting best practices, documentation, checked 16 October 2026platform.claude.com
Where companies use it
Not sure whether your company needs this?
AI solutions map