Summary
Role prompting: instruct the model to adopt a persona ('You are a senior engineer,' 'You are a careful editor'). Sometimes helps, sometimes measurably hurts, often has no effect. The literature is mixed; the discipline is to test rather than assume.
Objectives
- 01State what role prompting does mechanistically.
- 02Cite one case where it helps and one where it does not.
- 03Explain why it is often overrated.
The Lesson
The technique
Prepend a persona: 'You are a Nobel laureate physicist.' In principle, this conditions the model's output distribution toward that persona's typical text. In practice, effect sizes are small and inconsistent.
What actually helps
Structured instructions with concrete constraints beat vague personas. 'Reply in bullet points, cite sources, keep under 200 words' outperforms 'You are a helpful assistant.' Test both against a baseline before committing.
Key Ideas
- Role prompting is a weak lever compared to structure.
- Test personas empirically; do not trust folklore.