Role Prompting
also called Persona
Tell the model who it is supposed to be.
Setting an identity or expertise in the prompt — "you are a senior tax accountant" — to nudge vocabulary, standards and priorities.
Roles work because they shift the distribution of plausible continuations. They help most on tone and domain framing, and they are no substitute for actual information.
A role is a cheap way to shift the distribution: it changes which considerations the model surfaces and what it treats as important. What it does not do is supply information. The line to hold is that a persona sets a stance, never a fact.
A stance, not expertise
Code review comments are consistently too gentle to be useful.
"You are a sceptical staff engineer looking for failure modes, not a cheerleader" reliably changes what gets raised. It does not make the model better at knowing whether the concurrency is actually safe — that still needs the code and the tests.
Believing a role adds knowledge the model lacks. "You are a lawyer" does not make legal advice reliable.
- docsPrompting best practices ↗
The current house guidance on explicit instructions, output constraints, examples, XML structure and role prompting.
- docsPrompt engineering overview ↗
Ties each prompting technique to a success criterion and to empirical evaluation, rather than presenting them as a bag of tricks.