Structure & format
Role (Persona) Prompting
Role prompting, also called persona prompting, assigns the model an identity such as “You are a senior tax accountant” to set the vocabulary, priorities, tone and assumed audience of its answers. It reliably shapes style and framing, but research shows it does not reliably make a model more accurate on factual questions.
- Best for
- Tone, framing and domain priorities — not accuracy
- Difficulty
- Beginner
- Updated
What is Role (Persona) Prompting?
It is the most used technique and the most oversold. “You are a world-class expert” does not unlock hidden knowledge. A large study testing 162 different roles in system prompts found personas did not improve performance on objective questions compared with no persona, and that the best role for a question was hard to predict.
What a role does do is compress a lot of instructions into a few words: what to prioritise, what to leave out, how technical to be, which risks to flag. The useful version of a persona is therefore not a title but a specification — who the expert is, who they are talking to, what they care about and what they would never do.
When to use it — and when not to
Use it when
- Setting register and depth: explaining to a nervous first-time buyer versus to a fellow engineer.
- Pulling in a domain’s priorities: a security reviewer looks for different things than a copy editor.
- Simulating a stakeholder to stress-test your work: “a sceptical CFO reading this proposal”.
- Keeping a consistent voice across many outputs, such as a support assistant.
Skip it when
- As a substitute for facts, examples or retrieval. The role adds no knowledge.
- Impersonating real, named people — it invents opinions and quotes they never gave.
- When the role raises the model’s confidence beyond its evidence, such as a “doctor” persona giving firm diagnoses.
Copyable template
You are [ROLE] with [SPECIFIC EXPERIENCE, e.g. "ten years reviewing SaaS contracts for mid-size companies"]. Your reader is [AUDIENCE], who needs to [WHAT THEY WILL DO WITH YOUR ANSWER]. What you care about most: [2–3 PRIORITIES, e.g. "liability caps, auto-renewal terms, data ownership"]. What you never do: [e.g. "give a definitive legal opinion", "pad answers with generic advice"]. Tone and format: [e.g. "direct, plain English; label each risk High / Medium / Low"]. Task: [TASK] If part of the task falls outside this role's expertise, say so plainly instead of guessing.
Notice that most of the value is in the lines after the first. Specifying the audience is often more useful than specifying the expert.
Replace everything in [SQUARE BRACKETS] before you run it.
Worked example: A contract review for a founder
A founder wants a vendor contract checked before signing. “You are a lawyer, review this” produces a generic summary. The specified persona produces a decision aid.
You are a commercial contracts manager who has reviewed hundreds of SaaS vendor agreements for 20–200 person companies. Your reader is a non-lawyer founder deciding today whether to sign or push back. What you care about most: auto-renewal and notice periods, liability caps, data ownership on exit, price increases. What you never do: give a definitive legal opinion, or summarise clauses that pose no risk. Format: a table of Clause | Risk (High/Medium/Low) | What to ask for. Then one line: sign, or push back first. Task: review the contract below. [CONTRACT TEXT]
| Clause | Risk | What to ask for | | 12.1 Auto-renews for 36 months, 120-day notice | High | 12-month renewal, 30-day notice | | 9.3 Liability capped at 1 month of fees | High | Cap at 12 months of fees | | 14.2 Data export "on request, fees may apply" | Medium | Free export in CSV within 30 days of termination | Push back first: the renewal and liability terms are well outside market norms for a contract this size.
Why it works: The role sets priorities and leaves out low-risk clauses; the audience line turns a legal summary into a decision. None of it depends on the model “becoming” a lawyer.
Common mistakes
- Magic-words personas. “You are a world-class genius” changes little. Specify experience, audience, priorities and limits.
- Expecting more accuracy. Roles shape framing, not facts. For accuracy add sources, examples or a verification step.
- Impersonating real people. The model will fabricate their views. Describe the qualities you want instead of naming a person.
- Stacking roles. “You are a lawyer, marketer and data scientist” gives muddled priorities. Use one role per prompt, or chain separate roles.
- Authority without humility. Expert personas can sound certain when they shouldn’t. Tell the role to flag what is outside its expertise.
Origin & sources
Persona instructions have been in use since the first instruction-tuned chat models, and “You are a helpful assistant” became the default system prompt. The best-known systematic test is by Mingqian Zheng, Jiaxin Pei and David Jurgens (University of Michigan), first released in 2023 and published in Findings of EMNLP 2024 as “When ‘A Helpful Assistant’ Is Not Really Helpful”. Across 162 roles and several model families, they found personas in system prompts did not improve performance on objective questions.
Try it on The Prompt Index
Frequently asked questions
Does telling ChatGPT “you are an expert” make it more accurate?
Not reliably. Research testing 162 personas found they did not improve accuracy on objective questions. Personas are useful for tone, priorities and framing; for accuracy, give the model sources, examples or a checking step.
What is the difference between role prompting and persona prompting?
They are the same technique. “Role” usually refers to a job or expertise and “persona” to a fuller character with a voice, but both assign the model an identity to shape its answers.
Should the role go in the system prompt or the user message?
Either works. Put a stable persona that applies to the whole conversation in the system prompt, and a one-off role for a single task in the user message.