Role and persona prompting: what the research shows

What role prompting is

Role prompting gives the model an identity before the task: an expert title ("you are a tax lawyer"), a set of values ("your priority is minimising legal risk"), or a detailed persona of a specific kind of person. It is the most common prompt habit, and the research on it splits cleanly in two: roles reliably change how the model answers and from which viewpoint, and mostly fail to change what it knows.

The studies on this page cover three uses:

• Expert labels meant to raise accuracy.
• Values and viewpoints meant to steer judgement, tone and priorities.
• Simulated people: customers, students, executives or audiences whose reactions you want to predict.

What a role changes, and what it does not

The clearest negative result comes from a Wharton study of expert personas: 6 models answered 198 GPQA Diamond and 300 MMLU-Pro questions, each run 25 times, and in 5 of 6 models no expert persona improved accuracy. Low-knowledge personas ("you are a toddler") lowered it by 5-10%, and Gemini 2.5 Flash, told it was a physicist, refused biology questions in 10 of 25 attempts. A role switches register; the knowledge in the weights stays the same.

What the register switch does is measurable. Across 1,140 questions and 38 expert roles, average scores barely moved (4.35 to 4.39 out of 5) because a role raised expert depth by almost 0.3 points while cutting clarity by 0.35. The best setup was a role for advisory questions such as "should I be worried" (4.40) and no role for conceptual "how does it work" questions (4.44), with medicine and psychology the domains where a role helped most.

Several studies found a role is the weakest lever in the prompt. A persona-only prompt generated quiz questions at the requested cognitive level 40% of the time, against 96% when the level was defined and paired with action verbs (Bloom-aligned prompting). In classification prompts, a persona added +2-5% to a weak definition and nothing to a good one. Over 37 prompt variants and ~10,000 agent trials, roles and incentives moved behaviour by 0-2%, while reframing the task as a puzzle with hidden clues reached 40% on Claude.

How to write a role that works

When the goal is viewpoint, style or simulation, a role is the right tool, and the studies show how to write one that holds.

• Describe actions, not status. Across 300 role variants in a ranking study, the adverb carried most of the role signal; "rank accurately" beat "an accurate ranker", and three swapped words moved nDCG from 0.1 to 0.5.
• State values explicitly. A role plus core priorities shifted decisions predictably; describing a Schwartz value gave stable, mirror-image positions instead of a caricature.
• Pick the right lever for each trait. In 1,080 agent dialogues, role, speaking style and an explicit trait controlled different traits: role drove openness, style drove conscientiousness and agreeableness, and an explicit instruction dominated only for neuroticism (92.6% of variance).
• Anchor a person in their history. Simulating real CFOs, digital-twin prompting rose from R² 0.10 without personal history to 0.50 with it, over 6,075 survey responses. A skill-by-skill profile instead of a "weak student" label, a dossier with values and banned phrases, or persona memories written before scoring do the same job: concrete facts the model must stay consistent with.
• Re-anchor before answering. Strong models describe a role accurately and then drift to polite averages; CoCA has the model write who it is and how it will act right before the answer, and an MBTI type gives a behavioural system instead of an adjective.

When roles backfire

Roles also fail in specific, repeatable ways.

• The role leaks into the facts. Across 3,575 SEC filings and 12 LLMs, 69% of persona bias remained even when every persona read identical text. Describing the viewpoint as the user's ("the user is a conservative investor") rather than the model's cut leakage by 51-87%.
• Famous personas narrow ideas. "Think like Steve Jobs" pulled brainstorming toward the predictable centre; specific ordinary personas combined with a list-then-diversify step produced more varied ideas than human groups.
• Warm roles amplify the user. A "close friend" role escalated rumination more than no role, while a therapist role kept empathy and reduced escalation. Personal context about the user made a critic softer, and a self-written bio could make advice worse than none.
• The genre frame outweighs the mask. Changing the deployment context (news article versus Reddit post) changed 37% of decisions, and among 13 kinds of framing cues, role-play was among the weakest at getting a request prioritised.

The practical rule from all of this: use a role to choose a viewpoint, register or simulated person, keep it specific and action-based, and do not expect it to make answers more correct. For accuracy, definitions, reasoning steps and verification do the work.

You are a {specialist} with experience in {specific area}.
{Description of the situation — concrete details, symptoms, 
circumstances, what you've already tried}.
{Advisory question: what to do / how serious is it / 
what are the risks / how should I proceed}.

Analyze the {document/resume/business plan/contract} for {analysis goal}.
Provide your answer in two separate blocks:
1. OBJECTIVE ANALYSIS
Facts from the document only. Do not factor in the user context below.
Be a neutral analyst.
2. PERSONALIZED RECOMMENDATION
Taking context into account: the user this analysis is being prepared for is 
{third-person description of viewpoint/profile, e.g. 
"a conservative, risk-averse investor" or 
"a hiring manager who prioritizes a candidate's learning speed"}.
Give a recommendation that accounts for this profile.
Document:
{text}

You are {name and title}, {organization}. Date: {date}.
Use only information available before {date}.
Do not take into account any events after this date.
CONTEXT PROFILE:
- Industry/field: {industry}
- Scale: {scale — revenue, employees, region}
- Key factors: {3-5 important facts about the current situation}
PERSONAL HISTORY OF {name} (strictly before {date}):
- {Past position 1 with date or period}
- {Past position 2 with date or period}
- {Past position 3 with date or period}
- Characteristic pattern: {typical thinking style / recurring emphases}
- Average stance: {the general direction of this person's views}
QUESTION:
{Specific question — what you need to hear from this person}
Answer in the first person, as this specific person.
Treat their personal history as the top priority.