One camera that never cuts flies through a 3D world while a pixel avatar talks, with a station for every step you say.
AI-generated avatar + voice
What it madeFrom idea to render in one flightOne of my one-take reels, made with this engine: a single camera flies through a pixel city, one station per step of how a reel gets made.Voice only here. The finished edit also has a music bed and sound effects I can't share.
Give Claude a script and a voice track. It designs a world for the script (a dusk city, lit rooms, a teal lab, a neon terminal canyon, a sunrise library), places one station per beat, flies one continuous camera between them with arcs, lifts and orbits, and keeps a lip-synced pixel avatar fixed on screen like a chase-cam while the world moves around it. Every prop, sign and sound lands on a word of your voice. Gates read the camera and the scene at 30 fps to prove there is no cut, jump or flash before anything renders. Built from a short JSON spec in code with HyperFrames; five finished example specs are included.
Raw in
A script in your own words (20 to 45 seconds, 4 to 7 beats, a comment keyword at the end), a voice track of it (any recording or TTS, or your own ElevenLabs account) with word times, and optionally frames of your own videos for the props that show a phone. The pack ships a neutral demo avatar, a placeholder logo and placeholder frames you replace.
Finished out
A 1080x1920, 30 fps HyperFrames project: a world page built from a spec, an avatar lip-synced to your voice, props landing on your words, camera flights with whooshes on their peaks, an original-SFX mix at -14 LUFS, a verify block for the finished render and the gate reports. Nothing renders until you say go.
The look
a camera that never cuts, dissolves or fades
an avatar fixed on screen while the world flies around it
signs, cards, typed code and stamps that appear exactly on the word
five worlds from sunset city to neon terminal canyon
The first thing you paste
Use the one-take world skill. My script is ./script.txt and my voice take is ./voice.wav with word times in ./words.json.
Scaffold the project, design a world for this script (start from the closest example), build it, run the gates and show me
the contact sheets. Do not render until I say go.
Best for
tours, pipelines and method explainers told in 4 to 7 steps
reels where the order of the steps should be obvious from the motion
a continuous-camera look without a 3D app
a world you can re-skin by editing a spec
What is in the zip
SKILL.md (10 steps with a gate at each, five world recipes, the rules that cost a rebuild)
the world engine: builder, scenery, 26 prop types, camera curves
continuity, caption, near-plane and on-screen gates
reference of every spec key and prop type
a 7-second sample and five finished example specs with their scripts and word times
voice, word-time and lip-sync scripts, audio plan, mix and render finisher
a neutral demo avatar and the script that draws it