Opus 5.5 Replaces After Effects? The 9,500 Characters and 12 Hours the Viral Video Didn't Show You

A tweet presented Opus 5.5 as an After Effects replacement, with a demo video and a PDF prompt. In the comments, someone replied: “Motion designers are cooked.” First, watch the video.

The video looks like it was generated from one sentence. seeco then published a long post breaking down the full production pipeline behind this kind of viral video, titled Opus 5.5 Motion Design: What the Viral 'One-Prompt' Videos Are Hiding. After spending a week taking the videos apart, his conclusion was blunt: a few videos really do start from 30 words, but the best ones start from a 9,500-character brief, a skills folder, two API keys, and a 12-hour run. Thariq, who is on the Claude Code team, put it even less diplomatically: the post says one-shot, but the actual prompt runs to 10,000 characters and comes with skills, examples, and keys.

First, let’s establish a fact: Opus 5.5 cannot export MP4 directly. Every frame you see in the feed is drawn by a program it wrote running in the browser, then assembled into a local video by ffmpeg. At the core is a seek(t) function: you input seconds, and it outputs the screen image for that second. 15 seconds at 60fps equals 900 frames; a headless browser is called 900 times, frames are captured, and then encoded. There is no timer, so frame 240 can be redrawn independently. If the title appears too early, change one number and re-render.

This is also why results from a chat window alone are mediocre. seeco’s rig is Claude Code, Playwright, and ffmpeg, with CLAUDE.md at the project root hard-coding the rendering contract: window.seek(t) is responsible only for drawing frame t; in render mode, CSS transitions, setTimeout, and requestAnimationFrame are disabled; noise must be seeded; Math.random is banned. Setting Effort to medium is only suitable for small fixes; use xhigh for new videos, and max for anything whose first 3 seconds need to survive publication. A chat window can write animation, but only an agent with a shell can render, listen to audio, and look at its own frames.

Brief is another dividing line. If you ask the model to make “a crazy product video,” it will fill every blank with the safest default, such as a centered gradient title and all elements fading in together. A production document should at least include FILM, ASSETS, VISUAL, HARD LINE, and DELIVER. The hard line should state clearly: do not invent product interfaces or metrics. Tony Dinh’s TypingMind video added three more lines: the product URL, a requirement to use real screenshots and logo assets, and a requirement for music. Opus gathered assets from the site itself and finished within 30 minutes. A similar launch video a year earlier cost him about $1,200.

Reference frames determine whether a video looks like someone else’s video. Rexan Wong went through dozens of viral hits and kept seeing the same issue: naming a style is more effective than describing it; one reference frame or clip can directly communicate pacing, typography, and transitions. seeco’s process is to use ffmpeg to extract one frame every 0.5 seconds, write docs/style_guide.md, then write docs/shotlist.md, borrowing only the visual grammar, not the subject, logo, or copy.

A state list matters more than piling on effects. A container goes from logo to CTA button, email input, loader, success check, dashboard, self-drawn chart, tooltip, ⌘K panel, toast, and back to logo; every beat needs an entry state, an exit state, and a reason to exist. Springs use closed-form solutions to ensure seek(t) remains a pure function of time. Sound and picture share the same clock; if there is an existing track, run beats.json first so state changes land on the beat.

Then let Opus look at its own frames. Use ffmpeg to generate a contact sheet at 2 frames per second in 6 columns; cut a 12-frame strip around the fastest action; scale down to 360px for a phone readability test; loop playback twice to check seams; render the same frame twice and compare hashes. The Critic task judges only rendered frames, scores 1 to 10, lists the 3 worst defects with timestamps, evidence, and local fixes, then re-renders after fixes until every item is above 8. This step is often behind viral videos. Drew’s launch-day video reached 1.7 million views but still had obvious cleanup rounds.

This also lines up with how coding agents are evaluated. Artificial Analysis launched Coding Agent Index v2.0; Claude Opus 5.5 running in Claude Code ranks first at 61.7, covering 290 tasks. The new version replaces DeepSWE with Cognition’s FrontierCode v1.1; the average score across 21 configurations dropped from 54.8 to 47.4, with 20 configurations declining. The Index scores the model and the agent harness together, which explains seeco’s article exactly: the prompt accounts for only a small part of the result; the surrounding rig, files, gates, and review determine the ceiling.

Overnight runs need gates. Reference and product assets approved; style guide and shot list approved; stills from every shot critiqued; a 960x540 animatic with placeholder audio; a full pass; format and loop checks—each step can block progress. Donald dictated for five minutes and went to sleep; when he woke up, he had a 142-second music video, which later reached 2.1 million views; his prompt was about 9,500 characters. Another Pip robot video used a 19,000-character brief. If an interface screenshot is missing, stop and ask; do not let the model draw a realistic-looking fake dashboard.

Finally, fold the process into a skill folder: brief.md, style_guide.md, shotlist.md, assets, src, reviews, out. Next time, swap in the product brief and assets, and the command shrinks to one line. achxvi has already been selling a similar service: a 38-second vertical product video finished in 45 minutes, with music, a mascot, feature points, and an end offer, with up to 3 revisions.

One-sentence generation is good for testing a rig, not for judging the industry’s endgame. What actually carries over is the files, gates, and critique loop, so the tenth video will be better than the first. Real product assets cannot be invented, and the contact sheet cannot be skipped. Some commenters say motion designers are doomed; hold on. Run the contact sheet once, then decide who is doomed.