Role and persona

Concept

Role prompting is the second lever after clear instructions. Done well, it shifts the model into a specific voice, vocabulary, and set of priors. Done badly, it adds words without changing anything.

A role prompt works when it carries information the model would otherwise have to guess. "You are a senior procurement officer reviewing a tender response" tells the model what to scrutinise, what jargon is acceptable, and what counts as a red flag. "You are a helpful assistant" tells it nothing it did not already assume.

The most useful role prompts pair a role with an audience. "You are a tax accountant explaining VAT registration to a first-time freelancer" works because it sets both the expertise and the level of pitch. The model now knows to avoid technical shorthand without dumbing the content down.

Three things to watch for. First, do not stack roles. A prompt that says "you are a CEO, a copywriter, and a developer" forces the model to average across three voices and produces something none of them would write. Pick one. Second, make sure the role matches the task. Asking "you are a poet" to write tax advice will produce flowery tax advice, which is not what anyone wanted. Third, role prompts decay. After a few turns of conversation the role drifts; if it matters, restate it.

When role prompting fails, the fix is almost always to add what the role should be doing, not to add another role. "You are a marketing strategist" plus "your job in this conversation is to challenge weak positioning claims with one direct question per claim" is far more useful than a longer character sketch.