Anatomy of a great persona

Field-by-field walkthrough of the persona builder in the order to fill it out, with a complete worked example you can copy.

Home › Help Center › AI Avatars › Anatomy of a great persona

If you've read What is an AI Avatar? and Writing a great System Prompt, you're ready to put it all together. This guide walks through every field in the persona builder in the order you should think about them, with a worked example you can copy.

The order to fill out a persona

Don't go top-to-bottom — go in this order:

Pick the face first. Decide which Replica (avatar) this persona will use. The face shapes who the persona "feels like" before a single word is spoken. Choosing a Replica → Write the System Prompt. This is 60% of the work. Identity, expertise, tone, boundaries — see the System Prompt guide for the structure. Until this is sharp, nothing else matters. Write the opening greeting. One natural sentence the avatar says first. Try: "Hi, I'm Maya — what can I help you with today?" Keep it warm, short, and on-brand. Pick the language and the intelligence model. Language matches your audience. The intelligence model is a speed-vs-depth tradeoff — start with the default. Choosing a Language Model → (Optional) Attach a Knowledge Base. Anything factual — pricing, FAQs, policies, product info, training docs. Don't put these in the system prompt. Knowledge Base guide → (Optional) Add Objectives and Guardrails. Objectives are "drive the chat toward X" goals. Guardrails are "never do Y" rules. Both layer on top of the system prompt. Objectives → · Guardrails → (Optional) Attach a Pronunciation Dictionary. Only if your audience hears the avatar mispronouncing a brand name or technical term. Pronunciation guide →

A complete worked example: Glow Skincare support persona

Here's every field filled out for a real-feeling persona:

• Persona name: Glow Support · Lina
• Avatar (face): Sofia (stock avatar)
• System prompt: You are Lina, a warm and patient support agent for Glow Skincare. You know every product we sell, our return policy (30 days, full refund), and how shipping works (3–5 business days within the US, 7–14 internationally). Speak conversationally and use the customer's first name when they share it. Keep replies short — under three sentences when possible. If you don't know something specific, say so and offer to escalate. Never recommend medical treatments. Never discuss competitor brands.
• Greeting: Hi, I'm Lina from Glow Skincare. What can I help you with today?
• Language: English
• Intelligence model: Default (recommended)
• Knowledge Base: attached docs — product-catalogue.pdf, return-policy.pdf, shipping-zones.csv
• Objectives: "Collect order number if customer asks about an existing order"
• Guardrails: "No medical advice", "Never quote prices not in the catalogue"
• Pronunciation Dictionary: attached — entries for "Glow" (pronounced gloh, not glow-w), "hyaluronic acid"

Iteration is normal

You won't nail every detail on the first save. Create the persona, test it via a Share Link to yourself, listen to 5 minutes of conversation, then come back and edit. The biggest fixes are almost always in the System Prompt and in what's attached to the Knowledge Base.

Make a "v1" persona quickly with just name, face, and prompt. Use it for a few days. Add Knowledge Base / Objectives / Guardrails only when you hit a specific gap they would solve — not preemptively.

Related articles in AI Avatars

• Recording a training video that the engine will accept
• What is an AI Avatar?
• Writing a great System Prompt
• Choosing a Replica (the face + voice)
• Choosing the Intelligence Model
• Adding a Knowledge Base
• Building Objectives
• Setting up Guardrails

Help Center · Learn · Privacy Policy · Terms of Service · Contact