What we tested
We installed the skill from commit 3f7a55a into an empty project folder with one file: a 12-second, 960×540 MP4 of a paper-cut New Year title. The clip came from our own test of Code Animation Prompts, so we could publish both sides, and the agent had never seen how it was made. We asked Claude Code on Claude Opus 5.5 to clone the whole clip into plain HTML, CSS and JavaScript at the default motion-faithful level.
What we found
- The motion came out right. The agent wrote a 17-row motion spec from the video's frames and rebuilt the clip in about 300 lines of HTML and JavaScript. In its last round every animated part, from the ring and the corner rosettes to the two banners and the text typing in, started and ended within one frame (40 ms) of the original. The medallion's slide was the furthest off, by 10 px at the middle of a 233 px move.
- It measured instead of guessing. It wrote its own small scripts to track ring radii, banner edges and the slide, and used the skill's comparison sheets after each round. It stopped after three of the five allowed rounds, because the remaining difference was static styling.
- It was honest about the gap. The skill's whole-frame check failed its 5 % bar at 8.6 %, and the report said so, with the reasons: flat colours instead of paper texture and shadows, slightly different cut-out shapes, and Georgia standing in for a serif it could not identify. At the motion level those are allowed to differ.
- Setup needed a workaround. Bun 1.3.9 rejected the skill's lockfile ("Unknown lockfile version"). The agent installed the dependencies with npm and setup finished. The first run also downloads Chromium into Playwright's shared cache, which the README mentions, and in our run it left a static-ffmpeg download cache in the user Library folder and a few scratch files in /tmp that the agent's report did not mention.
- It is not quick. 15 minutes and $3.59 for a 12-second clip with no interaction. The author's live-site examples, with hovers, WebGL and workers, took 30 to 90 minutes each.
The result is a close, editable starting point. Asking for pixel-faithful would have made it chase the texture and the font as well, at more cost.




