Generate a Motion Video in 10 Minutes: What Opus 5.5 Is Really Doing Is This Pipeline
A motion-design video completed in 10 minutes spread on X. Under the hood, Opus 5.5 does not directly output an MP4; it writes a program that can be called frame by frame, the browser draws frames, and ffmpeg encodes them. What takes more time is the brief, references, storyboard, audio-visual sync, and review loop.
nicco posted a motion-design video on X, saying it took 10 minutes with Opus 5.5. The video is here:
He threw the PDF at Opus 5.5 and told readers to try it themselves. In the comments, someone posted a vertical version, and others said this does not look like something a single prompt could produce. Both reactions are right. The first version is fast; reliable delivery does not take the same path.
seeco (@seeconvm) has broken down this kind of viral video. His conclusion is specific: Opus 5.5 does not have the ability to export an MP4 directly. The piece you see on the timeline is backed by a program written by the model; the browser renders the program into frames, and ffmpeg compresses the frames into a video. There is only one core function, seek(t), which specifies what should be drawn on screen at second t. For a 15-second 60fps piece, a headless browser calls it 900 times and captures 900 frames. There is no timer, so frame 240 can be redrawn independently. If the title lands half a second early, change one number and re-render.
This is also why simply asking the model to write animation in a chat box tends to produce something mushy. It cannot see the frame it just rendered, and it cannot hear the sound. Only a shell-capable agent like Claude Code can run rendering, read images, change code, and render again. A minimal environment is Node, ffmpeg, Python, Playwright, and Chromium. Put CLAUDE.md in the project root and hard-code the rendering contract: window.seek(t) is responsible for drawing frames; do not use CSS transition, setTimeout, or requestAnimationFrame; noise must be seeded; do not touch Math.random. Also, do not leave effort at medium. Viral pieces commonly use xhigh or max; medium is only suitable for fixing small issues.
The more critical gap is in the brief. Ask the model to make a crazy product video, and it will fill in the blanks with the safest defaults, so you get a centered title, gradient background, and everything fading in together again. That famous showreel works with one sentence because it gives the genre: 15 seconds, 6 to 8 shots, the model itself as the protagonist, and no product facts that can be written incorrectly. Running it once to test the pipeline is fine; do not treat it as a production template. Hundreds of people posting the same sentence will end up with pieces that rhyme with one another. The term seeco cites is brief contagion.
When it comes to a real product video, the brief should read like a production document: who the audience is, duration and aspect ratio, the real logo, real UI screenshots, brand fonts and colors, audio, which references to keep, which clichés to avoid, and which line must not be crossed, such as no inventing product interfaces or metrics. Tony Dinh's TypingMind piece added only three lines: the product URL, a requirement for real screenshots and assets, and music being mandatory. Opus went to the site and grabbed assets itself, finishing in under half an hour. A similar launch video a year earlier would have cost him about $1,200. Rob Hallam then asked for a product ad in the same session and got it faster because the renderer, audio synthesis, and export pipeline were already in place.
References determine whether the piece looks like you. Without a visual grammar, Opus brings its own default aesthetic, and pieces from different projects start to look like relatives. Rexan Wong went through dozens of viral pieces and saw the same pattern: giving a style name is more useful than describing the style, and one reference frame or reference clip can deliver rhythm, typography, and transitions all at once. Pleometric's TikTok feed piece grew out of one frame from another viral animation, and his Donald-style piece points to his own PC-98 image library, a library no one else can copy.
Do not write the storyboard as a list of effects either. Zoom in, morph, add particles: these are actions, and they do not tell the audience what they should know at each second. A more useful approach is a list of states: 0 to 2 seconds hook, 2 to 5 seconds problem, 5 to 9 seconds product enters through real interaction, 9 to 13 seconds one screen or number as proof, 13 to 16 seconds resolution and CTA. @twoclipping's UI morph used this structure: 907K views, 19K saves. One container changes size, corner radius, and color between states; the cursor is driven by real clicks; the final frame connects back to the first, so the piece can loop.
Cheapness often comes from motion curves and sound. Fixed easing slides from A to B; motion with mass accelerates, slightly overshoots, and then stops. Closed-form springs preserve seek(t) determinism; when the value changes target multiple times, stack a spring for each change instead of restarting. Sound and picture share one clock. Rob Hallam thought audio was added in post, but it came directly out of code. @oozn's Steve Jobs biopic synthesized its score in Node, with 23 transitions locked to 120 BPM. If you already have a music track, measure it first: run librosa for bpm, beats, downbeats, and hits; put state changes on beats, big moments on downbeats, and clicks and whooshes on hits. Listen once with your eyes closed; if the sound has no shape, the sound design is not finished.
Review is the dividing line. Opus 5.5 can read images, so it can look at the contact sheet it just rendered. Do not watch the full export ten times in a row; let ffmpeg produce evidence: 2 frames per second in a 6-column contact sheet, a 12-frame burst near the fastest motion, scaled to 360px wide for a phone-readability test, loop playback twice, and render 300 frames twice to compare hashes. Then give it review tasks with teeth: look only at rendered frames, ignore what you intended to do, score on 2-second hook, phone readability, motion, variation, audio-visual sync, and brand accuracy; list the 3 worst defects with timestamps, evidence, and local fixes; fix those seconds, re-render, and repeat until every item scores above 8. @mablesjoseph's watercolor short used 163 model calls and nearly 7 hours. Drew's launch-day piece reached 1.7M views and was still going through visible patch rounds.
Finally, collect all of this into a skill. The directory contains brief.md, style_guide.md, shotlist.md, assets, src, reviews, and out. For overnight runs, set gates: references and product assets approved; style guide and shot list approved; each shot's still frame approved; 960x540 motion storyboard and placeholder audio track; full pass; format and loop checks. Donald dictated for 5 minutes and went to sleep, then woke up to a 142-second music video that later got 2.1M views. His prompt was about 9,500 characters; @pradeepXkapoor's Pip robot piece used 19,000 characters. John Heibel's PDoom repo, with 1.1K stars, demonstrates the sub-agent version: Opus writes ANIMATION_GUIDE.md and divides nine chapters among parallel workers.
People are already selling this as a service. achxvi takes on music, mascots in any style, product features, end offers, multiple languages, and up to 3 revisions; a 38-second vertical product video takes 45 minutes.
There is also a vertical test video in the replies:
A single prompt can produce a demo; that is fine. A deliverable piece depends on files, gates, and review. The first version looks like progress, and that is the step that fools people most easily.
The original post and PDF are here:
发布时间: 2026-10-30 10:43