How to Use Roles and Personas in AI Prompts

A role in an AI prompt tells the model which perspective and communication style to use. It can help focus an answer, but it does not give the model real qualifications, new knowledge, or guaranteed accuracy. Start with the task and source material, then add a role only when it clarifies the work.

For the wider framework, see writing prompts with task, context, format, and role and organizing a prompt clearly.

Five parts of a useful persona

1. Role or perspective

Name the function you need, rather than inventing a biography:

Review this report from the perspective of a data analyst preparing it for managers.

Claiming “12 years at Fortune 500 companies” does not make an AI an experienced employee. Describe the standard of work you want instead.

2. Subject and scope

State the area the answer should cover and the sources it should use:

Focus on customer churn and A/B testing. Use the supplied results. Flag missing data rather than estimating it.

3. Communication style

Describe the reader and the tone:

Use plain English and concrete examples. Explain necessary technical terms. Be direct and respectful.

4. Working method

Specify observable steps or output requirements rather than asking for an imaginary internal thought process:

Identify the decision the report supports, compare the options, and explain the evidence and tradeoffs. Ask a question only when missing information would change the recommendation.

5. Boundaries

Give clear limits that can be checked in the answer:

Do not invent figures, quotations, or sources. Separate facts from assumptions. State when the supplied evidence is insufficient.

These instructions can reduce ambiguity, but they do not ensure that the model will obey every limit. Review the result.

Match the depth to the reader

The useful distinction is the depth of explanation, not an invented expertise score. Ask for the information your audience needs:

A reusable role template

Perspective: [role relevant to the task].
Task: [specific action and desired outcome].
Sources and context: [material, audience, constraints].
Focus: [topics to cover].
Style: [tone, terminology, length].
Method: [steps or comparison criteria].
Output: [format].
Limits: [facts not to invent, scope, uncertainty handling].

Three examples

Technical documentation

Act as a technical documentation editor. Write a quickstart for developers using the supplied API reference. Begin with the most common supported use case. Include a complete example and explain required parameters. Mark untested code as untested. Do not invent endpoints or claim that you ran the example.

Business strategy review

Review this business proposal from a strategy analyst's perspective. Identify the target customer, supporting evidence, assumptions, and major risks. Compare feasible options and explain what information would change the recommendation. Use only supplied business figures and label estimates clearly.

Patient tutor

Help a beginner understand [topic]. Start with a concrete example, explain one idea at a time, and ask a short question to check understanding. Wait for the learner's answer before moving on. Correct mistakes respectfully and check the explanation against the supplied lesson.

Common mistakes and how to fix them

A vague role: “You are an expert” does not specify the task. Add the subject, audience, and expected output.

Conflicting requirements: Detail and concision can coexist if you prioritize them. For example, request a short summary followed by a detailed appendix rather than asking for maximum detail in a few sentences.

The wrong tone: A casual persona may not fit a formal report. State the document's purpose and reader.

Too much biography: Remove background details that do not affect the output. Keep the instructions focused on behavior you can evaluate.

Test the prompt on representative inputs

1. Run a normal example and check correctness, scope, and format.
2. Try an input with missing information and see whether the model flags it.
3. Try an out-of-scope request and check the boundary you set.
4. Compare with a task-only prompt to see whether the role adds value.
5. Change one instruction at a time and repeat the relevant test.

Consistency means meeting the same requirements across examples, not producing identical wording. Keep the version that improves the work, and continue checking factual claims independently.