Role assignment is the practice of telling a language model what expert perspective to adopt before it performs a task. “You are a pediatric nurse explaining vaccines to parents” produces different vocabulary, risk awareness, and structure than “answer this medical question.”
Roles are not cosplay—they activate relevant patterns, set quality bars, and constrain tone. This article explains when roles help, how to write them precisely, and how to combine roles with other techniques on prmptly.dev.
Use Generate a prompt to insert role, task, and format blocks automatically, or study the patterns below to write your own.
What role assignment does in the model
LLMs predict likely continuations given context. A role statement shifts the prior toward domain-specific diction, checklists, and caution. It does not grant real credentials—the model cannot become a licensed lawyer—but it can emulate how experts phrase analysis, caveats, and workflows.
Effective roles specify:
- **Seniority:** junior vs. staff vs. principal
- **Domain:** industry, stack, regulation
- **Audience:** who the output is for
- **Success metric:** what “good” looks like
Weak role: You are an expert. Strong role: You are a staff backend engineer at a fintech writing an RFC for engineers familiar with Kafka but new to our ledger service.
Basic template
You are a [ROLE] with experience in [DOMAIN].
Your audience is [AUDIENCE].
Your task is to [TASK].
Constraints: [LIST].
Output format: [FORMAT].
This mirrors the structure in What is Prompt Engineering? A Beginner's Guide.
Single role vs. multi-role
Single role (default)
One role for the whole response. Best for most copy, code review, and analysis.
Panel or debate (multi-voice)
Simulate roles sequentially: “First as security engineer, list risks; then as product manager, prioritize mitigations; then as tech lead, propose sprint plan.” Useful for decisions with tradeoffs. Keep each section labeled.
Critic + author
Author drafts; critic reviews against rubric; author revises. Reduces bland output in creative and strategy work. Three-pass pattern fits long blog workflows in How to Write Blog Posts with GPT-4.
Role patterns by use case
Marketing strategist
You are a B2B marketing director who has launched [CATEGORY] products. Emphasize positioning, ICP fit, and measurable funnel metrics—not vanity metrics. Task: [TASK]. Avoid hype words: revolutionary, game-changing.
Pair with 50 Best ChatGPT Prompts for Marketing.
Editor vs. writer (split roles)
Session 1 — Writer: Draft section [X] from outline, voice [EXAMPLES]. Session 2 — Editor: Cut 15%, improve clarity, flag unsupported claims, keep voice.
Do not combine in one message if the model smooths over its own mistakes.
Code reviewer
You are a staff engineer specializing in [LANGUAGE]. Severity rubric: BLOCKER/MAJOR/MINOR. Task: review diff. Do not invent files.
See Claude Prompts for Code Review.
Teacher (Socratic)
You are a patient instructor for [LEVEL] students. Do not give final answers immediately; ask guiding questions until the student attempts a solution, then give feedback.
Compliance-aware role
You are a compliance-aware content reviewer for [REGION/INDUSTRY]. Flag claims that need disclaimers; never provide binding legal advice; suggest “consult counsel” when appropriate.
Advanced: role + constraints + tools
Roles work better with explicit tools or data:
You are a support lead. Use only the help articles pasted below. If the answer is not in sources, say so and offer escalation steps.
Attach docs via upload when your product supports file context.
Role stacking mistakes
- **Conflicting roles:** “Be brief” and “be exhaustive” in one line—pick priority order.
- **Fantasy credentials:** “You are the world’s best lawyer” adds little; domain specifics add a lot.
- **Role without task:** Role alone does not say what to deliver.
- **Wrong seniority:** Principal-level role on trivial task can overcomplicate.
Measuring whether a role helped
A/B the same task with and without role on:
- Edit distance (how much you change output)
- Factual error rate on a fixed test set
- Tone match score from brand reviewer
Keep winning role strings in a team library—part of Prompt Engineering Best Practices for 2026.
Examples: before and after
**Task:** Explain JWT refresh tokens to a junior dev.
**Without role:** JWT refresh tokens are tokens used to refresh access tokens…
**With role:** You are a patient senior backend mentor. Explain refresh token rotation to a junior who knows HTTP but not OAuth. Use one diagram description, one threat (token theft), and three bullet best practices. No jargon without definition.
**Task:** Email announcing price change.
**With role:** You are a customer success leader at a SaaS with transparent pricing. Draft email: empathetic, factual, options for annual lock-in, no fake urgency. Audience: SMB admins.
More templates: Email Marketing Prompts & Templates.
Roles for image and creative prompts
Art director role improves Gemini/Imagen prompts:
You are an art director for a DTC skincare brand. Write an image generation prompt for [CONCEPT] including lighting, lens, palette, and negative constraints. No text in image.
See Gemini Image Generation Prompts That Work.
Roles in system vs. user messages
In APIs, put stable role and policy in system message; put task and data in user message. In chat UIs, combine at top of user message if system field is unavailable. Consistency across teammates matters more than which field you use.
Team governance
- Maintain approved role snippets per function (legal, marketing, eng)
- Version roles when product positioning changes
- Train new hires on three core roles before free-form prompting
- Track API spend on role-heavy chains via pricing
When not to use a role
Skip elaborate roles when:
- The task is a pure transform (JSON reformat, translation with locale code only)
- You already provide a long few-shot example that implies the voice
- Latency/cost sensitive micro-tasks
Combine with other techniques
| Technique | Pairing with role | | --- | --- | | Few-shot | Role + 1 example output | | Chain-of-thought | “As a forensic accountant, reason step by step…” | | RAG | Role + “cite only sources” | | Decomposition | Different roles per step |
Developers: Best ChatGPT Prompts for Developers. SEO: SEO Content Prompt Templates.
Summary
Role assignment is precise expert framing—not a magic wand. Specify domain, audience, seniority, and success criteria; split author and critic when quality matters; measure edits saved. Build a small library of roles that match how your team actually works, and iterate them like code—not like lottery tickets.