Role Prompting
Personas change vocabulary, framing, and priorities — not capability. When a role earns its tokens and when it's superstition.
A role prompt — 'You are a senior security engineer reviewing this pull request' — doesn't unlock hidden knowledge or intelligence. What it does is shift the distribution: which vocabulary is likely, which concerns surface first, which audience is assumed, what counts as a red flag. Ask a generic assistant to review code and you get style nits; frame it as a security review and injection risks and unvalidated inputs jump to the front. Same weights, same knowledge — a different region of behavior. That's the honest mental model: a role is a lens, not an upgrade.
Roles help most where perspective genuinely changes the right answer. Tone and voice: 'a pediatrician explaining to a worried parent' calibrates register better than a paragraph of style adjectives. Audience calibration: the same outage explained to executives versus engineers should differ, and a role sets that in one line. Review and critique: naming the reviewer's discipline — accessibility, security, legal — activates that discipline's checklist. And judgment calls with no single correct output, where whose priorities apply is the real question. In all of these the role is doing real semantic work: it's compressed context about audience and priorities.
Then there's the cargo cult. 'You are the world's greatest mathematician' does not improve arithmetic on Claude, GPT, or Gemini — controlled tests on modern models show flattery personas moving accuracy on objective tasks by roughly nothing, and occasionally hurting. Capability questions have capability answers: a stronger model, chain-of-thought, a calculator tool. Stacking superlatives — 'world-class, award-winning, 30 years of experience' — spends tokens on adjectives the model converts into confidence, not competence. The test is simple: if you can't say what the persona changes about the output, it's decoration.
Writing a role that earns its place: specify the job, the audience, and the stakes — not the excellence. 'A staff engineer writing an incident postmortem for non-technical execs who will set next quarter's budget' beats 'a brilliant writer' because every clause constrains the output: what to include, what to skip, what tone fits. Then let constraints do the rest. Role sets the lens; instructions, examples, and format do the heavy lifting. A persona is one line of a good prompt, never a substitute for one.