Content Creation

How Raphael Aubry Makes Motion Videos with Opus 5.5 in Claude Code

Built by Raphael aubry · · Source: X

Quick answer

Raphael aubry spent three days making motion videos with Opus 5.5 in Claude Code, then open-sourced the pipeline as a Claude Code skill. Opus writes code that paints each frame, a headless browser screenshots it, and ffmpeg stitches the video.

Result: Seven videos in 3 days with no paid tools, including a 24-second launch video for their SaaS Howseen that took 5 versions. The pipeline is open-sourced as a Claude Code skill.

What they built

Raphael aubry spent three days making motion design videos with Opus 5.5 in Claude Code: a 60-second voiced ad, an Apple keynote style film, a food app launch, a baguette launched like an iPhone, a K-pop paper zine remake, a showreel, and the launch video for their own SaaS, Howseen. They then packaged the whole pipeline as an open-source Claude Code skill at github.com/howseen-ai/claude-motion-design. They say the prompt is maybe 10% of the video and the harness around it is the other 90%.

How it works

Opus can't output an MP4, so it writes a program that paints frames. According to Raphael, everything hangs on one function, seek(t), which paints the exact frame for a given time with no timers and no state carried between frames. A headless browser calls it for every frame and takes a screenshot, and ffmpeg stitches the screenshots together. They say no After Effects, Remotion, video model or paid tool is involved, only Chromium, ffmpeg, Python, free music and free sound effects.

  • Fixed flow: gather the inputs (brand, photos, song, format), build a beat map so each scene starts on a beat, review 4 stills, then do the full render, a frame by frame check, and audio.
  • Springs: motion uses a closed-form damped spring, so any frame can be rendered without simulating the ones before it.
  • Motion blur: render 8 subframes per output frame and let ffmpeg average them. They say this costs about 1 to 1.5 minutes of rendering per second of film, and that 4 subframes left ghosting on fast moves.
  • Sound: they stopped trusting automatic beat grids after one was 2 beats off. Now the drop is found by measuring bass energy, each sound effect is placed on its measured peak, and the mix is normalized to minus 14 LUFS.
  • Self review: Opus renders contact sheets, scores its own frames from 1 to 10 on things like the hook, phone readability and sound sync, fixes the 3 biggest problems, and repeats until every score is 8 or higher.

To use the skill, they say to copy the folder into ~/.claude/skills/ and ask Claude Code for a video.

The result

Raphael is upfront that the one prompt claim is only half true. Some videos were usable on the first render, but the 24-second Howseen launch video took 5 versions, each with a few minutes of feedback. Reviewing stills first caught blurry text on the Crave remake, which they fixed in 2 minutes; found later, it would have meant a 10-minute re-render. That 15-second remake alone took 886 frames and 7,088 screenshots. They also admit the videos still look AI-made without a reference or a real product, and that a great motion designer using the same pipeline would beat them easily.

How to build something similar

  • Have your agent write the animation as a pure function of time, so a headless browser can screenshot any frame and ffmpeg can assemble the result.
  • Approve a handful of still frames before you spend time on a full render.
  • Give the agent a reference video, real product assets and real data, so it has something concrete to work from.
  • Make the agent look at its own rendered frames and score them against a checklist before you call the video done.

The original post

https://x.com/i/article/2104303014542512128 Most people trying motion design with Opus 5.5 hit the same wall: ↳ the first render is "mid": centered text, a gradient, everything fading in ↳ the music and the picture live separate lives ↳ one fix breaks three other scenes ↳ the caption says "one prompt", their prompt gets them nowhere near it I spent the last 3 days making motion videos with Opus 5.5 in Claude Code. A 60-second voiced ad, an Apple-keynote film, a food-app launch, a baguette launched like an iPhone, a K-pop paper-zine remake, a showreel, and the launch video for my own SaaS. Here's what I learned: the prompt is maybe 10% of the video. The other 90% is the harness around it. So I packaged the harness as a Claude Code skill and open-sourced it: github.com/howseen-ai/claude-motion-design This article is the whole pipeline, with the real numbers, the real code and the mistakes. Full disclosure first ↳ "One prompt" is half true. Some of my videos came out usable on the first render. The launch video for my SaaS took 5 versions. The Crave remake took a few fixes after the stills. ↳ Every number in this article is from my own renders and my own posts. ↳ Nothing here uses After Effects, Remotion, a video model or a paid tool. Chromium, ffmpeg, Python, free music and free SFX. 1. Opus doesn't make videos. It writes a program that paints frames. Opus 5.5 outputs text. It can't emit an MP4. Every video in this trend is code that something else turns into frames. The trick is one function: seek(t). Give it a time, it paints the exact frame for that moment. No timers, no CSS transitions, no state carried between frames. window.seek = async (t) => { // every style is computed from t, nothing else const u = P(t, 2.0, 0.5, E.inOut); // eased progress of a move box(card, 70, lerp(620, 420, u), 940, 440); // position = pure function of t rise(title, t, 2.1); // masked word-by-word rise return true; }; A headless browser calls it for every frame, takes a screenshot, and ffmpeg stitches them. Same input, same frames, every time. A fix is a one-line edit plus a re-render of the seconds that changed. 2. The flow that stopped the "mid" first renders The one-liner gets you a clip. A flow gets you a film. Mine is always the same, in this order: 1. Inputs: brand, photos, song, format. If the brief lists them, ask for them. 2. Beat map: every scene starts on a beat, and the music drop lands on the key visual moment. 3. 4 stills, reviewed before anything else. 4. Full render, then a frame-by-frame check, then audio. Step 3 saved me hours. On the Crave remake, the stills showed the message text going blurry during the send animation. Fixed in 2 minutes on a still. Found on a full render, that's a 10-minute re-render for nothing. Here's the one I'm proudest of. A baguette, launched like a new Apple product. Same flow, white background, classical music, real crunch sounds. https://x.com/RaphaelAubryy/status/2104178266990858739 3. Springs: why some motion looks expensive Cheap motion goes from A to B on a fixed curve. Expensive motion has mass: it accelerates, overshoots a hair, settles. The catch: it has to stay a pure function of time. So no physics simulation. A closed-form spring: function step(tau, f = 3, z = 0.6) { // damped spring, 0 -> 1 if (tau <= 0) return 0; const w = 2 * Math.PI * f, wd = w * Math.sqrt(1 - z * z); return 1 - Math.exp(-z * w * tau) * (Math.cos(wd * tau) + (z * w / wd) * Math.sin(wd * tau)); } // a value that changes target several times = one spring per change, summed function S(t, base, changes, f, z) { let v = base, prev = base; for (const [t0, to] of changes) { v += (to - prev) * step(t - t0, f, z); prev = to; } return v; } You can render frame 812 without simulating frames 0 to 811. That's what makes a 60 fps render deterministic. 4. Motion blur for free: render more frames than you keep The render step that makes it look like a real film: # 8 subfra
Raphael aubry (@RaphaelAubryy) · September 28, 2026 · View on X

Want to build something like this?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week.

Join Shipping Skool — Free

More use cases

← All AI agent use cases