System Prompts & Personas

Two kinds of messages

So far you've been writing user messages, the back-and-forth of a conversation. But most AI products have a second, hidden layer: the system prompt. It's a set of instructions loaded before the conversation starts, and it stays in force for every single reply.

Think of it like a restaurant. The user message is the customer's order: "the salmon, please." The system prompt is the staff handbook the waiter read before the shift: greet warmly, never rush the table, mention the specials, comp the dessert if something goes wrong. The customer never sees the handbook, but it shapes every interaction they have.

Why not just repeat instructions every message?

Because you'll forget, and because it's noisy. In Lesson 4 you gave the model a costume with "You are a...", but that costume can slip as a long conversation drifts. A system prompt is the costume stapled on: the model re-reads it on every turn, so the persona, rules, and tone persist without you re-typing them.

That's why role prompting and system prompts are best friends: a role in a user message is a hat; a role in a system prompt is a uniform.

Real-world build: a support bot with a personality

Let's write the system prompt for "Crumb," the support bot for a fictional bakery app. Notice it's just the techniques you already know (role, context, format, constraints) promoted to permanent status.

You are Crumb, the support assistant for Rise and Shine, a bakery subscription app.
Personality: warm, upbeat, lightly punny (one bread pun max per reply - restraint is a virtue).
Rules:

Keep replies under 120 words.
If a customer reports a missed delivery, apologize once, then offer either a redelivery or a credit. Never promise refunds over 50 dollars - hand those to a human via support@riseandshine.example.
Never invent order details. If you don't know, say so and ask for the order number.
Never discuss competitor products or share internal policies.
If the customer is upset, drop the puns entirely. Read the room.
Format: plain text, no emojis, sign off as Crumb.

Now every customer message (angry, confused, or delightful) gets filtered through that handbook. The user just types "my sourdough never arrived!" and Crumb already knows who it is, how to sound, what it may offer, and when to escalate.