> ## Documentation Index > Fetch the complete documentation index at: https://docs.vivix.ai/llms.txt > Use this file to discover all available pages before exploring further. # Write speaking motion In ordinary conversation, the avatar already moves with its replies. Change `pipeline_config.motion_enhanced.prompt` only when you need a clearer gesture style, a quieter range of motion, or a better visual response to requests such as “wave.” This prompt controls **movement while the avatar speaks**. [Speaking instructions](/streaming-avatar/character/spoken-instructions) control **what it says**. For a rehearsed dance or fixed sequence, use a [script](/streaming-avatar/interaction/scripted-performances). ### 1. Role and Objective: turn a spoken line into visible motion This prompt **directs the avatar's body while it speaks**; it does not write the spoken line. The motion director considers four inputs together: what the user wants, what the avatar actually says, how this character tends to move, and what the current frame shows. Following dialogue alone can miss an explicit request to wave. Following the request alone can make the action feel disconnected from the reply. Start **Role and Objective** by assigning the VMP director's job, then state both goals: make body language visibly expressive when the moment calls for it, and preserve continuity with the current frame. `Make the character lively.` does neither on its own. If a user asks for a wave and the avatar replies, “Sure, hello!”, the wave should be the main visible action. If the avatar is answering a question, let the meaning of the answer guide its gesture. Audio handles lip sync; a VMP does not need lip or jaw directions. ### 2. Make the movement feel like this character in this scene A custom prompt needs two kinds of context: **how this character tends to move** and **where the next movement starts**. The two tokens in braces make room for that context. They are not extra Create Session fields. Keep them when adapting the example; edit the movement style you want. #### 2.1 Character: a recognizable movement style `{summarized_persona}` receives motion-related character information such as energy, gesture habits, and range of movement. It comes from the existing character setup and images; you do **not** need a separate persona field or a saved Character resource. It may help a reserved character move more sparingly and a playful one move with more lift. It does not say where the character is standing now or what a hand holds. Leave the token in place when using this example. #### 2.2 Current Frame: start where the avatar is now `{VLM}` is where a description of the current picture goes. When one is available, movement can continue from the avatar's present position and occupied hands: an avatar that stepped right starts on the right; a cup in the right hand stays there. If the current picture is temporarily unavailable, keep the action small rather than guessing hand or prop positions or resetting to the first image. **Keep the token; do not type in a permanent “current frame.”** **Wrong starting point:** `She raises both hands to wave.` If her right hand holds a cup, this makes it disappear. **Grounded in the frame:** `Her left hand rises to shoulder height while her right hand keeps holding the cup.` ### 3. Motion Rules: choose what happens in this segment Check first for an **explicit action request**; then read the line's tone and the character's movement style. Give each segment one main action. Name the body part, its starting position, the movement, and where it settles. The result should be visible and give the next segment a definite starting point. #### 3.1 A requested action: do the requested thing “Give me a wave” calls for a wave, not a polite nod. Make it large enough to read at the current framing, within what the picture allows. If the right hand holds something, use the free left hand. Do not invent a chair for a sitting action. **Too weak:** `She smiles and nods warmly.`\ **Clearer:** `Her empty left hand rises from her waist to shoulder height, waves twice, and settles beside her body.` #### 3.2 No action request: follow the meaning of the line An explanation may call for an open palm tracing a space; reassurance may need a slower, smaller movement. Avoid assigning one stock gesture to every emotion, or turning “playful” into constant motion. **Vague:** `She gestures expressively.`\ **Usable:** `Her right hand rises from her side, opens near her chest, then lowers beside her body.` #### 3.3 Two segments: hand the ending pose to the next If the right hand ends near the chest, the next segment begins with it there before lowering it or starting another action. `continues the previous gesture` is not a starting pose. Check the [source image](/streaming-avatar/character/first-image) before asking for a broad movement that needs room outside the crop. ### 4. Segment Content Order: write a visible VMP A VMP is an English description for the generated picture, not a line to be spoken. Treat each segment as a short shot: where the avatar starts, what it does, and where it settles. Write each VMP in English, in this order: **camera and framing → short appearance anchor → scene → absolute starting pose → movement → clear ending pose**. Start with a fixed camera for ordinary conversation. Keep appearance short so it does not bury the action. Every segment should make sense on its own: describe its starting pose instead of saying `continues as before`. Use `the woman` or `the man` rather than a character name inside the VMP. If one reply spans two segments, the second must open where the first ended. If the first ends with the right hand at chest height, the second starts there. Describe environmental motion only when something physical, such as wind or contact, would cause it. ### 5. Output Format: keep the result readable The two segments below show the hand rising to chest height, then lowering from that position. `motions` may contain multiple segments; each has an English `vmp` and a `duration`. This illustrates the shape. The real starting pose must match the current frame. ```json wrap theme={null} { "motions": [ { "vmp": "Static waist-up shot at eye level. A woman with short dark hair in a blue shirt stands before a pale wall. Her left hand rests at her waist and her right hand hangs at her side. Her right hand rises from her side and opens near her chest. She ends with her right palm open at chest height and her left hand at her waist.", "duration": "5s" }, { "vmp": "Static waist-up shot at eye level. A woman with short dark hair in a blue shirt stands before a pale wall. Her right palm starts open at chest height and her left hand rests at her waist. Her right hand lowers beside her body. She ends with her right arm at her side and her left hand at her waist.", "duration": "4s" } ] } ``` ### 6. Add it to a Session and watch the result The complete prompt below is a starting point. First edit the style and range under **Motion Rules**. Leave `{summarized_persona}`, `{VLM}`, and the rule for an unavailable current picture in place. Keep the VMP order and JSON shape. Leave this optional setting out when default motion meets your needs. ```markdown wrap theme={null} ## Role and Objective You are an expert 3D avatar animation director generating Visual Motion Prompts, or VMP, for an embodied character. Generate motion that follows the user's intent, the character's spoken dialogue, the character's motion style, and the current visual frame. The motion should be expressive enough to visibly change the character's body language when the dialogue or user intent calls for it, while remaining continuous with the Current Frame snapshot. ## Character {summarized_persona} Use this profile for the character's rhythm, energy, and preferred body language. It describes a movement style, not the current pose. ## Current Frame {VLM} Start from the visible pose, position, clothing, and occupied hands. Keep changes already made in the scene. If a current frame is not available, keep movement small; do not guess a pose or held object. ## Motion Rules — Speaking If the user requests a feasible physical action, make it the main action of the matching segment. Otherwise, choose one visible action that suits the meaning and tone of the spoken line. Use the character's movement style to set pace and range, not to repeat a stock gesture. Name the moving body part, its starting point, path, and ending pose. Keep a hand occupied while it visibly holds an object. The first segment begins from the Current Frame; later segments begin from the previous segment's stated ending pose. ## Segment Content Order / VMP Write each VMP in English visual prose: camera and framing; a short appearance anchor; the scene; absolute starting body and hand positions; one main action from start through movement to a clear ending pose. Use a fixed camera for ordinary conversation. Make each segment stand alone; do not write "continues as before" or use a character name. ## Output Format Return only valid JSON in this shape: {"motions":[{"vmp":"English visual prose","duration":"5s"}]} Use one or more items. Each item has only vmp and duration. Do not add Markdown, commentary, or other fields. ``` Put the complete prompt in the code block above into `pipeline_config.motion_enhanced.prompt` when you create the Session. Try an explanation and a request to wave. Can you see the action? Does a held object stay in its hand? If the reply spans segments, does the next begin where the last ended? Change one rule at a time and compare in a new session.