What you'll learn
A single line — "You are a patient math tutor" — can change a model's tone, vocabulary, and depth for an entire conversation. That line is a role, and its natural home is the system prompt: the highest-leverage place to set behavior that should persist across every turn.
By the end of this module you'll be able to:
- Explain what assigning a role actually changes — and what it doesn't
- Write a strong system prompt that sets role, tone, and rules
- Tune output for a specific audience with a persona
- Recognize when a role is decoration rather than help
What a role actually changes
Giving the model a role primes it toward the region of language associated with that role — the word choices, assumptions, and level of detail an expert of that kind would use. It doesn't give the model new knowledge or new abilities; it focuses the knowledge it already has.
Prompt
Explain what a database index is.
Response
Prompt
You are a patient tutor explaining to a total beginner. Explain what a database index is using one everyday analogy, in under 60 words.
Response
The system prompt
Recall the three message roles from Module 4. The system message is where you set standing behavior — it applies to the whole conversation, not just one turn, and the model treats it as higher-priority guidance than an individual user message. In a chat app it's usually hidden; through the API you write it yourself.
System
You are the support assistant for TaskFlow. Be concise and warm. Never invent features. If asked something you can't verify, direct the user to help@taskflow.app.
Prompt
Does TaskFlow have a dark mode and can it read my mind?
AI response
Key idea
Personas & audience
A role can describe who the model is; a persona can also describe who it's talking to. Naming the audience is often more powerful than naming the expert, because it fixes the vocabulary and depth precisely.
| Add to the prompt | Effect on the answer |
|---|---|
| “Explain to a 10-year-old” | Short sentences, simple words, concrete analogies |
| “Explain to a hiring manager” | Outcomes and impact over technical detail |
| “You are a meticulous copy editor” | Careful, rule-driven, flags small errors |
| “Audience: senior engineers” | Assumes fundamentals, goes deeper, uses precise terms |
Rules that persist across turns
Because the system prompt applies to the whole conversation, it's the right place for constraints you never want to repeat: output format, forbidden topics, brand voice, or a language requirement. State them once, clearly, and prefer positive phrasing (Module 5).
Note
When roles help — and when they don't
Roles shine when the style, depth, or perspective of the answer matters. They do far less for tasks with one correct answer. "You are a brilliant mathematician, what is 17 × 24?" is no more accurate than just asking — the flattery is wasted tokens.
- Helps: tone, audience level, writing style, which trade-offs to emphasize, perspective.
- Doesn't help much: factual lookups, arithmetic, and anything with a single objective answer.
Tip
Recap & quick check
Key takeaways
- A role focuses the knowledge the model already has — it changes style and depth, not its abilities.
- The system prompt is the highest-leverage place for role, tone, and rules that should persist all conversation.
- Describing the audience ('explain to a beginner') controls vocabulary and depth as well as any expert role.
- System rules are strong but can be overridden over long chats or by adversarial input.
- Roles help with style and perspective, not with single-answer factual or arithmetic tasks.
Quick check
1. What does assigning a role like 'You are a patient tutor' actually do?
2. Where should durable rules like tone and output format live?
3. For which task is a role LEAST likely to improve the answer?
4. Which is the better role instruction?
You've told the model what to do and who to be. Sometimes, though, the clearest instruction is a demonstration. Next up: Module 7 — Zero-shot & Few-shot Prompting.