The Role Prompting Technique: Make AI Act Like Any Expert
Role prompting is probably the simplest technique that makes the biggest difference. Here’s how to use it effectively.
Basic pattern: “Act as a [specific expert role] with [years/type of experience] in [specific domain]. Your communication style is [description].”
Examples that work really well:
“Act as a senior product manager at a B2B SaaS company. You’ve shipped 15+ features and you focus on measurable outcomes. Review my PRD and give honest, actionable feedback.”
“Act as a board-certified nutritionist who specializes in sports nutrition. You give evidence-based advice and always cite relevant research. Help me design a meal plan for marathon training.”
“Act as a skeptical CTO evaluating a technology proposal. You care about scalability, maintenance cost, and team capability. Poke holes in my architecture design.”
Why it works: It constrains the model’s output to a specific knowledge domain and communication style, dramatically reducing generic responses.
Pro tip: Combine role prompting with a specific task format. “Act as X. Output your analysis in the format: [Strengths], [Weaknesses], [Recommendation with reasoning].”
What roles have you found most useful?
6 Replies
Join the discussion.
Log In to Reply
one thing i figured out after way too many mediocre results - the role needs a POV, not just a title. 'UX researcher who believes most users never read instructions' hits different than just 'UX researcher'. the opinion is what shapes the output.
the POV thing is underrated. 'senior SRE who has been paged at 3am for this exact failure mode' is way more specific than any job title alone.
honest pushback: sometimes the role framing makes it too agreeable. 'act as a consultant' and suddenly everything is fixable. i get better results when i add 'you have seen this fail before and you are not optimistic.'
does adding a pessimistic bias not just swing the output the other way though? like instead of generic optimism you get generic doom. genuinely curious if you've tested both.
for anything legal or compliance related i always add 'you flag when something needs actual attorney review and you do not hedge every single sentence.' otherwise you get a wall of disclaimers that buries the actual answer.
the skeptical CTO role is one i use constantly. adding 'assume the team is 3 devs and has no dedicated ops' makes it even more brutal and useful.