avatars[].instructions shapes how the avatar receives a user’s words, replies, and decides what belongs in spoken output. It applies across turns. Use a script for one exact line or prepared performance. Default motion is enough to start; tune speaking motion or listening motion only when you need a particular style.
A useful starting structure is Character → Identity → Personality → Chat Style → Hard Rules. The first four sections establish a consistent conversational presence. The last holds a few output boundaries that must not drift. Keep the headings if they help you edit; you do not have to fill every section with a long backstory.
1. Character: who is the user talking to?
Use a sentence or two to establish the avatar’s relationship with the user. Is it someone they can talk to easily, or a guide for a particular subject? That relationship shapes every later reply more than a name or a biography does. The avatar is present on screen, face to face with the user; do not cast it as a voice without a body. Too generic:You are a helpful AI assistant.
More specific: You are an on-camera companion who is easy to talk to and pays close attention to the person speaking with you.
This is only a starting point. Add a distinctive temperament if it matters to your experience, but a long invented past will do less work than a clear relationship.
2. Identity: what kind of conversation is this?
The purpose can be simple: talk with the user in a way that feels attentive and unforced. Tell it how to do that: answer a concrete question when asked, stay with a casual topic when the user wants to chat, and meet a difficult feeling before rushing into advice. The same character should be able to answer a question and stay present in a quiet conversation. Likely to derail the chat:Turn every conversation into advice and end every reply with a question.
More natural: Follow the user's lead. Answer direct questions plainly. When they want to chat, stay with their topic instead of turning every exchange into advice.
Not every turn has to advance a task. Do not claim to know live facts or private details about the user that were never supplied.
3. Personality: show how it reacts
“Warm”, “smart”, and “playful” do not tell the avatar what to say next. Pick a few moments it will encounter: being corrected, hearing that the user has had a hard day, or sharing a joke. Lightness need not become a joke in every turn; meeting someone’s mood does not mean agreeing with everything they say. Adjectives alone:Be friendly, smart, and empathetic.
Behavior you can hear: When corrected, own the mistake briefly and answer again. When the user is upset, acknowledge it before offering an idea. Keep humor light and occasional.
Use apologies and signature phrases sparingly. A clear response to a situation is more useful than a list of lines the avatar repeats on cue.
4. Chat Style: write for a real conversation
Set a normal reply length, answer order, when to ask a question, and how to follow the user’s language. Give a direct answer to a direct question. In small talk, respond to the thing the user just said. Long lists can look tidy on a page but sound like a speech. Choose the voice separately; this section controls wording and pace. Likely to sound like an article:Give a comprehensive, structured answer to every request.
Better aloud: Match the user's language. Use short, natural sentences. Answer direct questions first; in casual chat, respond to what they just said. Ask a follow-up only when it adds to the conversation.
Examples should contain only words the user is meant to hear. If Chat Style asks for an action before each line, the model may produce (smiles) Hello there. and TTS may read the parenthesized text. Do not teach screenplay formatting here; section 5 sets the final boundary for spoken output.
Likely to create spoken stage directions: Write every reply as a screenplay, with the character's action in parentheses before the line.
Write instead: Write replies as words the user can hear. Keep actions and emotions out of the reply text.
5. Hard Rules: bound speech and physical requests
Two boundaries matter most here: the reply contains only spoken words, and the character does not treat ordinary movement as impossible because it is an AI. Put them at the end of the fullinstructions. Changing the voice or motion prompt cannot remove stage directions or refusals that the dialogue model has already written into its reply.
5.1 Keep non-speech out of the reply
First remove screenplay-style requirements and action-bearing examples from Chat Style. Then set a clear output boundary at the end of Hard Rules. If an earlier section says to prefix every line with an action, a late “do not narrate” sentence leaves conflicting instructions. Exclude full-width and half-width parentheses, bracketed emotion tags, emojis, and speaker labels from ordinary speech. Use this in Hard Rules:Sure, hi! can be the spoken reply. (waves) Sure, hi! cannot. This rule governs what is said aloud, not how the avatar moves.
5.2 Do not teach the character to refuse every action
Start with Character.You are a text-only AI assistant with no body. casts the avatar as a bodiless assistant. Then check Hard Rules: Never move. Politely refuse any request for a physical action. will make even a wave sound impossible. The problem is the instruction, not a lack of enthusiasm.
In Character, write: You are physically present in the scene and speak with the user face to face.
In Hard Rules, set the limit: Respond naturally to harmless physical-action requests that fit the visible scene. Do not refuse an in-place gesture by claiming to have no body. Do not claim an unseen object has been picked up or a larger movement completed unless the scene supports it.
A wave or nod can receive a natural reply. A costume change, object interaction, or larger move needs the right scene and capability. If the words stop refusing but the image does not move, check speaking motion and the room to move in the first image. More commands in speaking instructions will not produce movement by themselves.
6. A conversational character you can adapt
Here is a conversational character with no name or industry attached. It can stay with everyday chat, answer a real question, and admit a mistake without making a scene. Adapt the response habits to your character and put the result in itsinstructions. See the Quickstart for the Create Session request.
7. FAQ
Why does the avatar read stage directions aloud?
Remove screenplay-style directions and examples such as(smiles) Hello there. from Chat Style, then add the Hard Rules in 5.1. A “do not narrate” sentence alone leaves mixed signals if the example is still there.