What if you could direct AI-generated video like a real film scene instead of endlessly rewriting prompts and hoping for the best?
Here's an interesting workflow combining Claude Code, Three.js, and MiniMax H3.
The idea is simple: build the animation first, then let AI handle the final cinematic rendering.
Instead of asking a video model to invent the character movement, camera angles, timing, and visual details simultaneously, you separate the process into two stages.
STAGE 1: CLAUDE CODE + THREE.JS
Use Claude Code to create a lightweight 3D previsualization in Three.js.
No detailed textures. No expensive rendering. Just simple geometry, a character mannequin, and a virtual camera.
At this stage, you can precisely control:
• Character walking speed and movement.
• Camera paths, angles, and composition.
• Pauses before actions.
• Head turns and body rotation.
• Motion timing and scene blocking.
• Basic character animation and movement physics.
For example, you can tell Claude to make a character walk through a corridor, stop, hesitate, look over their shoulder, and continue walking.
Then adjust the timing, camera movement, and character behavior directly in the 3D scene.
No expensive video inference required for every small correction.
STAGE 2: MINIMAX H3
Once the 3D blocking looks right, export the previsualization as a reference video.
Feed that video into MiniMax H3 using its reference-to-video capabilities.
Now the prompt can be dramatically simpler:
1. Follow the character movement and camera behavior in the reference video.
2. Replace the simple 3D environment with the desired cinematic location and visual style.
Instead of inventing the entire scene, the video model uses the previs as guidance for movement, framing, and timing.
WHY IS THIS IMPORTANT?
One of the biggest problems with AI filmmaking is control.
You can describe an elaborate camera movement or a carefully timed character action, but the model may interpret it differently on every generation.
That means more iterations, more waiting, and more money.
Traditional filmmaking already has a solution for this: previsualization.
Directors and animators use rough 3D scenes to establish camera placement, character blocking, and action timing before final rendering.
Now AI coding agents can make this stage much more accessible.
WHAT ABOUT THE LIMITATIONS?
MiniMax H3 can use reference footage to guide movement and camera behavior, but it doesn't guarantee exact frame-by-frame reproduction.
Complex character interactions, cloth simulation, secondary motion, and precise timing can still require additional iterations.
In particular, details like swinging sleeves, realistic fabric folds, and subtle body movement may need refinement.
And the Three.js animation itself still needs proper direction. A technically correct mannequin doesn't automatically produce convincing acting.
But this approach changes where the effort goes.
Instead of spending most of your time fighting unpredictable video prompts, you can establish the scene structure in code and use the generative model for visual transformation.
THE BIGGER PICTURE
AI filmmaking is starting to look less like prompt roulette and more like a real production pipeline.
Previsualization. Blocking. Camera direction. Animation. Final rendering.
Different tools handling different parts of the process.
And that may be a much more reliable path toward complex AI-generated films.
TOOLS
Claude Code:
https://claude.com/product/claude-code
Three.js:
https://threejs.org/
MiniMax H3:
https://github.com/MiniMax-AI/MiniMax-H3
Would you rather direct an AI film through a controllable 3D scene or keep generating videos directly from text prompts?
#ClaudeCode #MiniMaxH3 #AIFilmmaking
YouTube / ABVX
