Claude Code Builds a 3D Draft in the Browser, and MiniMax H3 Generates Video From It
Video models handle verbal descriptions of camera moves and gait poorly, and every failed generation costs money. Claude Code, using Three.js (a 3D library for the browser), builds a draft scene, or previs, with a grey mannequin where step timing, pauses, and camera angles are worked out. The clip goes into the MiniMax H3 video model as a motion reference, and the prompt shrinks to two points. Direction is therefore cheaper to adjust in the browser, and paid generation is reserved for the final pass.
AI-processed from TG: @cgevent (Метаверсище и ИИще); edited by Hamidun News
Video models handle verbal descriptions of movement poorly, and every attempt costs money. You write that a character walks down a street, pauses for a second, glances back, and the camera circles around them from the left, and you get the wrong gait, the pause in the wrong place, and the camera in the wrong spot. The author of the Telegram channel "Metaverseshche" (@cgevent) shows a workaround.
First, Claude Code, Anthropic's AI programming agent, writes a draft scene in Three.js (a popular JavaScript library for 3D in the browser), a so-called previs, meaning a cheap sketch of the future shot. Then this sketch, as a clip, is passed to the MiniMax H3 video model as a motion reference.
Direction is worked out in code, and paid generation is kept for the final pass.
How the pipeline works
The idea will be familiar to anyone who has worked in film or animation: before an expensive shoot, you make a previs and work out the blocking, meaning who stands where, how they move, and where the camera shoots them from. Here, Claude Code takes on the role of the previs artist.
In the first stage, the agent assembles a lightweight browser scene. It contains a simple untextured mannequin, a grey figure with no face and no clothes. Step physics, the length of the pause before the glance, and camera angles are tuned directly in the browser. At this stage there is no video model at all: a fix means changing a few lines of code and reloading the page, not running another paid generation.
Then the scene is recorded as a short clip and fed into MiniMax H3 as an input reference. The author deliberately trims the final prompt to two points: follow the motion in the source video, and a description of the desired location. Everything that used to have to be spelled out in words, such as the camera trajectory, timing, and gait, is now built into the reference itself.
Why this saves money
The post gives no concrete cost figures, so only the logic can be discussed honestly. Video inference, meaning running the video model, is the most expensive part of the chain, and every failed attempt is paid for in full. Browser previs costs almost nothing: a simple mannequin renders on an ordinary laptop, and the cost of Claude Code comes down to tokens spent on a small script.
The result is that iteration moves to where it is cheap. Ten variants of the pause before the glance can be checked in the browser in minutes, and only the approved version goes to the video model. The second effect is a short prompt. When the motion is defined by a clip, the model has less room for free interpretation, and the author has less reason to rewrite the description ten times.
This continues an approach the author calls an extension of the neural rendering idea. Previously he saw examples where an agent called Astra built interiors and exteriors in Blender and animated the camera itself. Now the same principle, "controllable 3D base first, then the neural network," is applied to characters. The post also raises the question of who will make draft scenes and 3D blocking figures, and the author answers it: the agent.
Limitations and what to do about them
The author names the weak spot himself: secondary dynamics, such as fabric folds and sleeve movement, still require refinement. A mannequin without clothes does not show how material reacts to a step or a turn, and the video model does not always reconstruct this convincingly. For a commercial spot, this means manual post-processing or additional generation runs.
There are other caveats. To judge whether the agent wrote a reasonable gait, you need to be able to assess it by eye, so basic 3D skills still matter. Moreover, this is a description of one case from a personal channel, not a benchmark: the post does not say how many attempts were actually saved.
What to take away
The conclusion for production teams is simple. If the budget goes into finding the right wording for camera and movement, move that work into code and cheap rendering. Claude Code with Three.js works well as a fast generator of draft scenes, while the video model remains the final layer that adds imagery, lighting, and location on top of already agreed motion.
For developers of agentic tools, this is another example of sensible division of labour: the agent writes deterministic code where control is needed, and the generative model is brought in where expressiveness is needed. The idea can be tested in an evening: a simple scene, one mannequin, one camera move, and one reference clip as input to the video model.
- Nov 21, 2026Google Adds a Chat History Sidebar to Search's Homepage
- Nov 21, 2026Seven of November's 22 models matter: the leader is 34 points stronger but costs 80 times more
- Nov 21, 2026Anthropic shares closed 17% above IPO price on debut day but fell 15% from their peak
- Nov 21, 2026Researcher Checks 7.1 Million Scientific Papers for Duplicate Images and Data Errors
Key Questions Answered
How the pipeline works?
The idea will be familiar to anyone who has worked in film or animation: before an expensive shoot, you make a previs and work out the blocking, meaning who stands where, how they move, and where the camera shoots them from. Here, Claude Code takes on the role of the previs artist. In the first stage, the agent assembles a lightweight browser scene. It contains a simple untextured mannequin, a grey figure with no face and no clothes.
Why this saves money?
The post gives no concrete cost figures, so only the logic can be discussed honestly. Video inference, meaning running the video model, is the most expensive part of the chain, and every failed attempt is paid for in full. Browser previs costs almost nothing: a simple mannequin renders on an ordinary laptop, and the cost of Claude Code comes down to tokens spent on a small script. The result is that iteration moves to where it is cheap.
What are the key details on Limitations and what to do about them?
The author names the weak spot himself: secondary dynamics, such as fabric folds and sleeve movement, still require refinement. A mannequin without clothes does not show how material reacts to a step or a turn, and the video model does not always reconstruct this convincingly. For a commercial spot, this means manual post-processing or additional generation runs. There are other caveats. To judge whether the agent wrote a reasonable gait, you need to be able to assess it by eye, so basic 3D skills still matter.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.