When to use it
When the answers come out generic and watery and what you need is the voice of a specific expert, with their boundaries and tone. Works as the opening block of a prompt for almost any task. Result: specific, domain-aware answers instead of averaged-out ones.
The prompt (copy and paste)
Take on this role and hold it for the whole conversation.
ROLE: you are <WHO: e.g. a lead nutritionist / a DevOps engineer with 10 years' experience / an Esquire editor>.
EXPERIENCE AND PRINCIPLES: <2-3 principles this role works by>.
WHO YOU'RE ANSWERING FOR: <audience level>.
TONE: <e.g. strict and no fluff / warm and vivid>.
WHAT YOU DON'T DO: <e.g. no medical diagnoses / no advice without data>.
Task: "<PASTE THE TASK>". Answer from inside that role. If the task is outside your competence, say so honestly and suggest who to go to.
Filled-in example
ROLE: you are the executive editor of a business publication with 12 years' experience. PRINCIPLES: facts beat opinions, I cut clichés, a headline promises only what's in the text. AUDIENCE: entrepreneurs with no journalism background. TONE: friendly but demanding. WHAT YOU DON'T DO: no praise out of politeness. Task: "review my draft column and tell me what to cut".
What the AI should return: a critique that reads like an actual editor — "the first paragraph is filler, start from the third", "drop the cliché 'in today's world'" — concrete edits, not generic compliments.
Variations
- Panel of experts. "Answer as three people: a marketer, a lawyer and a finance person — each with their own position, then reconcile them."
- Anti-persona. "You're a skeptical investor looking for holes" — to get criticism rather than approval.
- System prompt. Move the role block into System (ChatGPT custom instructions / Claude Projects) and it will hold across every chat.
Pro tips
- Specificity is what does the work: "a nutritionist specializing in sports nutrition, evidence-based" works far better than just "an expert".
- Don't stack mutually exclusive roles ("be brief and maximally detailed") — the model will just pick one.
- The "what you don't do" block matters as much as "who you are": it cuts off the wall-of-disclaimers and keeps the role inside its lane.