Video as code: what Claude Opus 5.5 is actually rendering

When most people say "AI video" they mean a diffusion model: you describe a scene, the model hallucinates pixels, and you take what you get. Skillry's Opus 5.5 collection is a different thing entirely, and the distinction is the whole reason it caught my attention. Nothing in it is generated as pixels. Every piece is a program — Three.js, Canvas, SVG, GSAP — that Claude Opus 5.5 wrote, and the video is simply what that program renders.

That sounds like a pedantic difference until you think about what it means for the output. A rendered program is deterministic, seekable, resolution-independent and, most importantly, editable. You are not stuck regenerating and praying. You open the source and change a number.

What's in the collection

Beyond those, there are 55 3D scenes plus a set of interactive pieces. The range runs from motion designer reels and product launch videos through to podcast animations, service explainers and game trailers — so it is less a tech demo and more a catalogue of the commercial formats people actually get paid to produce.

The stack underneath

• Three.js and WebGL for 3D scenes and camera work
• Canvas and SVG for 2D motion graphics and vector animation
• GSAP for timeline choreography and easing
• GLSL shaders where the effect needs to run per-pixel on the GPU
• HyperFrames and Remotion as the composition and render frameworks
• CSS and pixel art for the lighter, stylised pieces
• ElevenLabs, Gemini TTS and Lyrria for voiceover and audio

If you have written any of that by hand, the interesting part is not that a model can produce it — it is the volume. Choreographing a GSAP timeline so that every beat lands against a voiceover is fiddly, unglamorous work, and it is exactly the kind of work that benefits from a model that can hold the whole timeline in its head and iterate on it.

Why code-rendered video matters for engineers

The properties you get for free are the ones a diffusion model cannot give you. Text renders as actual text, so it is never smeared or misspelled. A brand colour is a hex value, not an approximation. The same source re-renders at 1080p or 4K without regenerating anything. And because the artefact is source code, it diffs, it reviews, and it lives in version control next to everything else you ship.

The tradeoff is just as real. Code-rendered video cannot give you photoreal footage of a place that does not exist. It is the right tool for motion graphics, data visualisation, UI walkthroughs, explainers and title sequences — and the wrong tool the moment you need something that looks filmed.

Worth knowing before you click

The gallery is browsable, but full access to the prompts and project breakdowns sits behind a subscription that starts around $9.99/month. I would rather say that up front than have you find out after you have spent ten minutes deciding you want the source. The site also notes the strongest results came from running Opus 5.5 at its highest reasoning setting, which tracks: this kind of long-timeline, many-constraint code generation is exactly where extra reasoning shows up in the output.

What I would actually take from it, subscription or not, is the pattern. If you already write TypeScript, you are roughly one Three.js scene away from being able to produce your own launch animations, and the collection is a good map of how far that goes before you hit the limits.