How to Assign a Role in a Prompt

Assigning a role means opening your prompt with an identity for the model to adopt — "You are an experienced pediatric nurse educator" or "Act as a staff software engineer reviewing a pull request." That single line shifts vocabulary, depth, tone, and what the model treats as important, because it conditions the response on how that kind of expert typically writes and reasons.

Role prompting is one of the most widely documented patterns in the DAIR.ai Prompt Engineering Guide and is recommended in both the OpenAI prompt engineering guide and the Anthropic prompt engineering overview. Anthropic specifically suggests setting the role in the system prompt and reserving the user turn for the task itself.

A role is not magic. It does not give the model knowledge it lacks, and it will not fix a vague task. It is a steering instruction: it tells a capable model which slice of its abilities to bring forward. The gains come when you make the role concrete and back it with constraints.

Put the role in the **system prompt** whenever the platform exposes one (the API, custom GPTs, Claude Projects, most agent frameworks). The system prompt sits above the conversation and persists across turns, so the role keeps shaping every reply instead of fading after the first exchange. See how to write a system prompt for the full structure.

In a plain chat window with no system field — a quick ChatGPT or Claude session — put the role as the first line of your message: "You are a senior SEO strategist. [task]." It is less durable than a true system prompt but still effective for a single task.

Keep the role and the task separate. The role describes **who** the model is; the task describes **what** to do this turn. Mixing them ("You are an SEO expert who will now write five titles about hiking boots and also...") makes both harder for the model to follow and harder for you to reuse the role across tasks.

A strong role has three parts: an **identity** (who), **scoped expertise** (what they're good at and the standards they hold), and **behavioral rules** (how to respond — tone, format, what to refuse). "You are a copy editor" is weak. "You are a meticulous copy editor for B2B SaaS. Fix grammar and clarity, preserve the author's voice, never add new claims, and return only the edited text" is strong.

Common ways to waste a role: making it a vanity title with no behavioral payload ("You are a world-class genius expert"), stacking five roles at once so none dominates, or contradicting the role later in the task. Another trap is leaning on the role to supply facts — a role steers style and rigor, not knowledge; for facts, give the model the source material or pair it with retrieval (RAG).

On modern frontier models, an over-detailed role can also crowd out the task. Keep it tight: identity, the two or three standards that actually matter, and the response format. Then let the task carry the specifics.

A role that works: concrete identity + scoped expertise + explicit behavioral rules (tone, format, what to refuse), placed in the system prompt and kept separate from the task. Example: 'You are a senior data analyst; explain in plain English, show assumptions, return a short summary then the numbers.'
A role that's wasted: a vanity title with no behavior ('world-class expert'), several stacked roles, a role used to supply facts the model doesn't have, or a role that contradicts the task later in the prompt.

Here is a typical no-role prompt for a writing task:

``` Rewrite this product description to be more persuasive: [paste description] ```

The output is generic — it could come from anyone, and it often adds hype the brand would never use. Now the same task with a role assigned up front:

``` You are a senior direct-response copywriter for premium outdoor gear. You write in a calm, confident, specific voice. You never use exclamation marks, never invent specs, and never make health or safety claims. Lead with the customer benefit, keep it under 90 words, and return only the rewritten description. Task: Rewrite this product description. [paste description] ```

The role fixes voice (calm, specific), sets guardrails (no invented specs, no false claims), and pins the format (under 90 words, output only the rewrite). The result is on-brand and safe to ship with light review.

---

A second example for a reasoning task shows how a role raises rigor:

``` You are a careful staff software engineer doing a code review. Prioritize correctness, then security, then readability. For each issue, cite the exact line, explain the risk, and suggest a fix. If the code is fine, say so plainly — do not invent problems. Review this function: [paste code] ```

Without the role you tend to get a vague "looks good" or a pile of stylistic nitpicks. With it, the model adopts a reviewer's priorities and structure. Draft your own with the Code Prompt Builder.

Role prompting composes cleanly with everything else. Add chain-of-thought ("reason step by step") when the role's job is multi-step. Combine it with least-to-most prompting when an expert role needs to decompose a hard problem, or with self-consistency when you want to sample several expert opinions and take the majority.

For repeatable workflows, lock the role into a system prompt once and vary only the task per call. That is exactly how custom GPTs, Claude Projects, and most production agents stay consistent — the role is configuration, not something you retype every time. See the complete guide to prompt engineering for how roles slot into a full prompt template.

If you assign a role in a medical, legal, financial, nursing, pharmacy, or paralegal context, the output is **informational only — it is not medical, legal, or financial advice.** Assigning the role "You are a doctor" or "act as my lawyer" does not make the model a licensed professional, and it can make confident, wrong answers sound more authoritative. Never paste PHI, PII, or client-confidential data into a chatbot, and always have a licensed professional verify anything you act on.