Claude Makes Videos Without Generating Visuals, by Drawing With Code

Someone on GitHub has put together a list of 8 projects and Skills that get Claude involved in making videos, with well over 100,000 stars combined.

Let's be clear about what these things are built on. They don't rely on video-generation models. Models like Sora read text and spit out pixels directly; these 8 projects have Claude write code and render the visuals frame by frame. Music, voiceover, and editing are also done in code.

anatomy, about 24,100 stars. Karpathy has recommended it. It turns abstract concepts into interactive isometric explainer diagrams—engines, mechanical structures, and software concepts can all be broken down into parts, motion, and real-time readouts. It outputs SVG, WebGL, and 3D interactive visuals. It only handles the visualization step; a complete finished video still needs to be recorded or assembled separately.

github.com/wheresryan22/anatomy

lemo-opuscar, about 11,200 stars. 66 ready-made talking-head video styles, each paired with a short film made with code, plus style prompts and production guides. Choose a visual style, tell Claude what you want to talk about, and it follows that workflow. The visuals are drawn and rendered with code, and the music and voiceover are also part of the workflow. The author states it plainly in the repo: this is not an awesome list; every film was made by him and Claude Opus 5.5.

github.com/lemomo-ai/lemo-opuscar

pixelfx, about 3,540 stars, reaching that number four days after launch. 244 pixel effects: fire, explosions, lightning, magic, weather, and complete attack animations. Each effect is a small piece of code that can be previewed, recolored, mirrored, and exported as a sprite sheet. The methodology comes from Hugo Duprez's article on making pixel effects with Claude; its core is three lines: code is clay, trust your eyes, feed references. Some effects in the repo are modified from his open-source code.

github.com/patch-gg/pixelfx

huashu-motion, about 2,130 stars. It leans toward product launch videos and promotional motion graphics. Give it product materials, and it first proposes creative directions, then makes the visuals and music. The more interesting part is how you revise it: click a spot on the preview and say directly that this needs changing; the feedback includes the timestamp and element information. Titles, colors, numbers, and dwell time can also be adjusted in the editor.

github.com/alchaincyf/huashu-motion

animate, about 1,410 stars. It uses code to make explainer shorts, historical stories, and small animations, with 7 built-in styles: paper cutout, pen hatching, risograph, sketchbook, mathematical, pixel, and isometric. The confirmation step comes first: review the story pacing, then the style frames and storyboards, and only after confirmation does it make the animation. It suits people with clear requirements for both visual style and narrative.

github.com/cth9191/animate

teardown, about 1,280 stars. It lets Claude Code turn machines into 3D simulators you can rotate, inspect, and take apart. You can see how an engine runs and how landing gear retracts, with motion and parameters viewed side by side. The result is an offline interactive webpage, and it can also be recorded frame by frame into 60fps video. The repo includes three complete examples.

github.com/elvarofficial/teardown

vibe-knowledge-film, about 950 stars. Aimed at knowledge-focused short videos, it offers two directions: Nocturne and Fable. The former explains abstract, difficult knowledge; the latter dramatizes psychological and behavioral logic into stories. Give it a topic or script, and it organizes subtitles, storyboards, and sound effects, first making a sample of about 30 seconds, then continuing to the full film after confirmation. The reference samples in the repo are for learning only and cannot be posted directly as your own work.

github.com/HeyChengC/vibe-knowledge-film

motion-prompt, about 2,900 stars. 15 motion graphics design styles, each with a 10-second finished video, all made by Claude Opus 5.5 writing code.

github.com/showheyohtaki/motion-prompt

There are two kinds of experience in the comments worth noting.

One commenter says that no matter how many skills you install, before rendering, first have Claude output a timeline table: how many seconds each shot is and what is moving; after confirming that, write the code, and you'll have much less rework. Another says to get one Skill working on its own first before combining them; mixing them tends to lose control over style and duration.

Some also question it. Someone asked whether videos made this way can compare with Stable Diffusion, saying the quality feels far off. No one in the comments answered directly. The two handle different objects: SD generates the pixels themselves, while these projects generate code that describes the pixels; the ceiling for the visuals depends on how detailed the code is.

Someone else asked whether it can be used for commercial delivery, and no one responded to that either. At least judging from the licenses stated in the repos, anatomy and pixelfx are both MIT, lemo-opuscar's code is MIT, its bundled demo films are CC BY 4.0, and films users make themselves belong to them.

The value of this toolkit now may not be replacing video-generation models. It is more like turning the production process of explainer videos into code. Diagrams, structural teardowns, and parameter demonstrations are hard to make reliably with generative models; drawing them with code allows precise control, and changing one part doesn't require rerunning the whole segment.