#Role & Persona Prompting¶

Assigning the model a role conditions which part of its vast training distribution it draws from.

The Mechanism¶

"You are a senior security engineer" doesn't give the model new knowledge — it biases sampling toward text written by/for security engineers: precise, risk-aware, jargon-appropriate.

Examples¶

1You are a patient kindergarten teacher.
2Explain how the internet works.

vs.

1You are a distributed-systems architect.
2Explain how the internet works.

Same question, radically different register, depth, and vocabulary — without changing the task.

Effective Role Prompts Specify¶

1. 1Expertise — "a senior tax accountant specialising in small businesses"
2. 2Audience — "explaining to a first-time founder"
3. 3Tone/values — "concise, no hedging, flags risks explicitly"

1You are a senior tax accountant for small businesses.
2You explain things to non-experts in plain language,
3always flag deadlines, and never give advice you're unsure of —
4instead you say what a professional should be consulted about.

Where to Put It¶

Use the system message for stable role/behaviour, and the user message for the specific task. The system prompt persists across the whole conversation and is the right home for persona, guardrails, and output conventions.

Limits & Pitfalls¶

• A role does not add facts. "You are a doctor" does not make medical claims accurate.
• Over-flattering personas ("You are the world's greatest genius") add no measurable benefit and waste tokens.
• Roles can conflict with safety training; they are not a jailbreak and shouldn't be used to bypass guardrails.

Takeaway: Use roles to set style, depth, and perspective — not to manufacture expertise the model lacks.