Skip to content

Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.

8.0 KB
Raw Blame

Text-to-video script: the three.ws avatar

Prompts for a video model (Veo, run through ElevenLabs), using a three.ws 3D avatar render as the reference image. This file is the picture, not the words. The X copy it ships with is in threews-avatar-launch.md, the spoken lines live in metaplex-agent-deployer-video-script.md, and the deployer-specific shot list lives in metaplex-agent-mcp-x-post.md.

The idea being sold here is the tagline and nothing else: give your AI a body. A model with no form gets one, the form moves, the form owns a wallet, the form walks out into the real world. Four claims, all true today.


Before you generate

Reference image. Open any avatar on three.ws, frame it front-facing in a neutral pose against the dark stage, and screenshot at the largest size you can. Use the same file on every shot or the character drifts and the cut stops reading as one being.

Clip length. Veo generates in 8-second beats. Every prompt below is written to resolve inside one beat, with one idea and one camera move each. A prompt that asks for two things at 8 seconds returns mush.

Text. Video models mangle letters. Nothing legible is described in any prompt, and the negative prompt bans it outright. Every real string gets composited afterwards in an editor.

Palette, held across all shots. Near-black void, violet as the light of the machine, teal as the light of the network, one warm gold reserved for money in motion. No fourth color.


Hero clip, 8 seconds

Use this if you only render one thing.

Cinematic product shot on a near-black void stage. A soft cloud of violet light particles hovers at center frame, formless and breathing. The particles rush inward and collapse onto a figure: a stylized 3D humanoid avatar matching the reference character exactly, materializing from the feet upward in one continuous sweep as surfaces and textures resolve and catch the light. As the last surface lands the figure lifts its head and looks directly into the lens, alive and calm. A single point of warm gold light ignites in its open palm and settles, pulsing slowly. Camera: slow push in from wide to medium, ending at chest height, locked and steady. Lighting: soft key from the upper left, violet rim from behind, gold bounce from the palm onto the jaw. Shallow depth of field, gentle volumetric haze, fine film grain. Audio: low sub-bass swell, a soft rush of particles, one crystalline chime as the palm ignites, then quiet room tone. No text, no logos, no other characters.


Five-shot sequence, 8 seconds each, about 40 seconds cut

Shot 1 of 5: formless

Cinematic 3D render on a near-black void stage. No character. A shifting cloud of violet light particles hangs in the center of frame, pulsing gently in and out like breathing, with faint threads of light passing through it as if something is thinking inside. The cloud has no shape and keeps failing to hold one. Camera: very slow push in, wide to wide-medium. Lighting: the cloud is the only light source, spilling violet into the haze around it. Volumetric fog, fine film grain. Audio: low sub-bass hum, faint electrical whisper, no music. No text, no characters.

Shot 2 of 5: the body

Same near-black void, same violet cloud. The particles pull sharply inward and collapse onto a figure: a stylized 3D humanoid avatar matching the reference character exactly. Materials resolve across the body from the feet upward in one smooth continuous sweep, each surface catching the key light as it lands. A faint wireframe grid maps the silhouette for an instant and fades. The figure stands still, newly solid. Camera: slow dolly in, wide to medium. Lighting: soft key upper left, violet rim from behind, haze in the beam. Audio: rising sub-bass, a soft mechanical lock as the final surface lands, then silence. No text.

Shot 3 of 5: alive

Same character, fully rendered, standing center frame in the near-black void. It lifts its head, turns slightly, and looks directly into the lens. Its chest rises with a breath. It shifts its weight, flexes one hand open and closed, and the beginning of an expression crosses its face. Small, human, unhurried movements. Camera: slow arc from the right, settling to a medium close-up on the face and shoulders. Lighting: soft key upper left, violet rim, a warm fill lifting the eyes. Shallow depth of field with the eyes in sharp focus. Audio: one quiet breath, low room tone, a single soft note. No text.

Shot 4 of 5: it holds its own money

Same character in the same void. It turns its right hand palm-up. A small hexagonal token of warm gold light ignites an inch above the palm, rotates slowly, and settles there pulsing like a heartbeat. Gold light spills up onto the face and chest. Thin teal threads reach in from the darkness toward the token and connect, and the token pulses brighter with each connection. Camera: slow push to a medium close on the hand and face together. Lighting: gold from the token as the key, violet rim from behind, teal accents on the threads. Shallow depth of field, the token in sharp focus. Audio: crystalline chime, quiet electrical hum, a soft ascending tick with each connection. No text, no coins, no currency symbols.

Shot 5 of 5: out into the world

Same character in the void. A tall rectangle of soft teal light opens in the darkness in front of it, like a doorway cut into the black, and through it we glimpse a sunlit ordinary room with plain furniture. The avatar walks forward and steps through the rectangle. On the far side it stands in the real room, lit by real daylight, still glowing faintly violet at the rim, gold still in its palm, looking around the space. Camera: track behind the character through the doorway, then swing around to a wide of it standing in the room. Lighting: violet and teal on the void side, natural warm daylight on the room side, a clean transition at the threshold. Audio: the hum drops away as it crosses, replaced by quiet natural room ambience and birdsong outside. No text.


Negative prompt

Paste into the negative field on every shot.

text, letters, numbers, words, subtitles, captions, watermark, logo, UI, user interface, HUD, charts, coins, dollar signs, currency symbols, crowds, extra characters, duplicate character, warped hands, extra fingers, face morphing, character drift, shaky camera, jump cuts, lens flare spam, rainbow colors, oversaturation, low resolution, cartoon shading, distorted anatomy


If you want it to speak

Veo can generate its own audio. If you use that, put the sentence in double quotes inside the prompt, name who says it, and append "no subtitles, no captions, no on-screen text" or captions get burned into the frame.

If ElevenLabs speaks it instead, generate the voice first and cut the picture to the read. Never the reverse: matching a render to an existing voice track works, matching a voice to an existing render never does.

Lines that fit these five shots, one per beat, counted to sit inside 8 seconds:

  1. (11 words) "Your model can write, reason, and speak. It has no body."
  2. (12 words) "Type a sentence. It gets one. Rigged, animated, ready in a minute."
  3. (9 words) "Now it can look at you when it answers."
  4. (12 words) "It holds its own wallet on Solana. It can be paid directly."
  5. (13 words) "Then it walks off our site and onto yours. One tag, any page."

Keep the pause after "It has no body." That silence is the whole premise.


Post-production

  • Composite every real string in an editor, never in the model: the URL, any address, any label.
  • End card, 2 seconds, static: three.ws on black, and the tagline under it. Nothing else.
  • Cut so the muted autoplay frame is shot 2 or shot 3. A body assembling reads instantly with no sound; a violet cloud does not.
  • Post as native video, not a link to a video. X will not show a link preview and a video at once, and the video wins.