You brief the model properly. Context, goal, audience, constraints, the lot. The answer comes back accurate, organised, and somehow still reads like the consensus of everyone who has ever written about the subject.
That is because it is. Nobody was speaking. You supplied the situation and left the vantage point empty, so the answer got drawn from the middle of everything the model has read.
Ask how should I think about failure? of a Socratic philosophy professor and you get questions back: what did you expect, why did you expect it, what would have counted as success. Ask a startup growth operator and you get a post-mortem template and a plan for the next two weeks. Both answers are good. They are not two phrasings of one answer, they are genuinely different answers, and nothing in your question decides which one turns up.
A persona does not make the model act like an expert. It decides which region of the model's training the answer is drawn from.
This is measurable rather than mystical. Research presented at NAACL 2024 found that role-play prompting consistently outperformed standard zero-shot across a spread of reasoning benchmarks, with one task climbing from 23.8% to 84.2% accuracy. Nothing was retrained in between. The model was simply asked to answer as somebody in particular.
The word that ruins most personas
"Experienced." Along with "senior", "expert" and "world-class". They feel like they are doing work and they specify nothing, because there is no body of writing produced by people who are merely experienced.
Compare:
- Weak. You are an experienced lawyer.
- Strong. You are a Berlin-based corporate lawyer who has advised 50+ startups on GDPR compliance.
The second has a desk, a jurisdiction and a stack of past cases behind it. That is the material the model reaches for. The first has a job title and a compliment.