Since Claude Opus 5.5 launched on September 22, 2026, a distinctive kind of video has flooded X: product films, explainers, music videos and animations that no camera shot and no video model generated. A language model wrote them — as code.
The honest mechanics: Opus 5.5 cannot emit an MP4. What it does unusually well is write a program that draws — a function that paints the exact frame for any moment in time. A headless browser calls that function hundreds of times, screenshots each result, and ffmpeg compresses the frames into a video. Same code in, same video out, byte for byte.
Press play — these play right here. Every one was rendered from code the model wrote; each card links through to the creator's original upload.
A full music video modeled entirely in code — figures, city, data center, stadium, planet. The wave's flagship proof that code can carry a whole film.
Music video · Remotion + three.js/WebGL · no image or video generation · Watch on YouTube ↗
A 5.5-minute 3D film following one prompt from power and data centers to tokens and the Transformer — built to teach how AI actually works.
Explainer film · Remotion 4 + React Three Fiber · one Claude Code session · Watch on YouTube ↗
Every lyric word tracked to the beat and placed inside a rain-soaked neon city as signage and holograms — 890 words, zero stock footage, zero cameras.
Music video · custom 3D engine · 22,128 frames at 60 fps · Watch on YouTube ↗
The video explains itself: the same function-of-time pipeline that produced it, broken down level by level while you watch it happen.
Meta explainer · pure JavaScript + Playwright + ffmpeg · zero paid tools · Watch on YouTube ↗
One prompt produces a 10-second explainer (water to steam); two edited lines retime it to a voiceover — the editability argument in miniature.
Motion graphics · one prompt in Claude Code · Vox-style explainer · Watch on YouTube ↗
中文圈最完整的一期：先讲清「一个函数、一个浏览器、一个 ffmpeg」的原理，再给三种做法——从一句话出片到让 Opus 5.5 当总导演。
中文讲解 · 原理 + 3 种做法 · JavaScript animation skills · Watch on YouTube ↗
Wisely Chen's documented run: one sentence turned a blog post into a 10-minute lesson and a 3-minute vertical cut. The lasting lesson is the verification loop — Whisper re-listens to the narration, screenshots audit the frames.
Claude Code · ffmpeg + headless Chrome + TTS · 2026-09-25 · Published workflow write-up
A techno-optimist hype edit assembled from found clips: the creator picked moments and gave feedback while Claude handled editing and animation. Under two hours of the creator's own time, per the release.
X · Arcmira MCP + Claude editing · reposted by Marc Andreessen and Garry Tan per the release · Video on X · via EIN Presswire report, 2026-10-14
The reference product film of the wave: a vintage turntable rebuilt in 3D and shot like a commercial — notable for taking nine days of human art direction on top of the model's code.
Opus 5.5 · 3D rendering · via the open index · Curated via the awesome-claude-5-5-videos index
Proof the pattern fits education: formulas, graphs and narration-timed reveals, all typeset by code (Manim) instead of drawn by a video model.
Opus 5.5 · Manim · via the open index · Curated via the awesome-claude-5-5-videos index
Guizang's skill rebuilds a product's real components and design language in code, then directs the launch film around them — storyboard, original score and sound effects included.
github.com/op7418 · ★852 · No license stated — curation link only
The largest shot library in the wave: 152 named, reusable camera and motion recipes for product videos, so a film is assembled from proven shots instead of improvised ones.
github.com/Vincentwei1021 · ★13,072 · MIT-listed collection — check per skill
The explainer pattern packaged: give it any topic and it produces a narrated, code-rendered explainer video end to end.
github.com/Vincentwei1021 · ★2,694 · License not stated — workflow replication candidate
A Claude Code / Codex video skill that proposes three directions with still previews first, then renders the chosen one frame by frame across Canvas, WebGL, Three.js or Remotion.
github.com/Kianzzz · ★244 · No license stated — curation link only
Fifty-nine ready film styles plus director and technique guides: style as a library, so the agent art-directs from a menu instead of inventing a look per project.
github.com/lemomo-ai · ★5,800 · No license stated — curation link only
The most careful index of the wave: 424 source posts catalogued as of October 2026 (301 of them Opus), with production paths, prompts and cost notes linked back to the originals.
github.com/athemeroy · ★730 · CC BY 4.0 (index itself)
Licenses are as stated in each repository on October 23, 2026; a missing license means all rights reserved — we link, we do not copy. Two entries are flagged as workflow-replication candidates for our own catalog: video-shotcraft's shot-recipe pattern and anything2explainer's topic-to-explainer flow.
Our own shelf: four free, downloadable Agent Skills built on the same Remotion pipeline — each one a finished template your agent customizes with your words, names and numbers.
Skilldraw(t)ffmpeg assembles the filmNo. Opus 5.5 is a language model: it reads and writes text and code. Creators point it (usually inside Claude Code) at a frame-drawing setup: the model writes a program that paints each frame, a headless browser or renderer steps through time, and ffmpeg assembles the frames into an MP4. Every video in this wave is code first, footage second.
Determinism and editability. The video is source code: change one line — a word, a color, a timing — and re-render exactly the same video with that one change. Pixel-generating video models cannot do that. Code rendering also nails text, charts, diagrams and UI, which generative video still garbles.
The recurring stacks are Remotion (React components rendered frame by frame), plain Canvas 2D or WebGL/Three.js in a single HTML file drawn by a seek(t) function, Manim for math and science explainers, and HyperFrames. A headless browser captures frames; ffmpeg stitches, mixes narration and burns in captions.
Start from a finished template instead of a blank prompt. The four free SkillU skills below are standard SKILL.md packages built on a Remotion pipeline — animated titles, terminal typing, pipeline diagrams and animated statistics. Install one, give your agent your words and numbers, and render your first code-drawn video today.
They are starting points — the same production pattern (Remotion compositions, frame rendering, ffmpeg) that the curated works use, packaged so any compatible agent can run it. Creators then extend the templates with their own scenes, voices and styles.