AI Does UI Animation: Layout Passes, but Easing Still Gives It Away
Meng To shared on X a list of UI animation skills for agents, including HyperFrames, Emil Kowalski's skills, Remotion, and seven in total. The comments pinpointed the key issue: agents get layout right now, but the easing effects still look obviously AI-generated at a glance.
AI Does UI Animation: Layout Passes, but Easing Still Gives It Away
Meng To shared a list of Motion skills on X, specifically for AI agents building UI animation. Seven entries, from video rendering to Blender workflows. One comment nailed the current state: agents get layout right now, but the easing still gives them away as AI at a glance.
7 skills
| Tool | Purpose |
|---|---|
| HyperFrames | HeyGen's open-source AI video editor. Write HTML, render MP4. Apache 2.0, no React required |
| Emil Kowalski's skills | 19 skills covering the full process from animation vocabulary to design engineering |
| DesignBookmark Motion | Copyable animation components you can hand straight to Claude or Codex |
| Remotion skills | Write videos with React |
| Blender Skills | Turn Blender workflows into agent skills |
| Motion | Layout transitions and spring animations |
| Motion Canvas | Code-driven animated explainer videos |
The common thread across these tools: let agents directly produce video or animation, with code merely as an intermediate artifact. HyperFrames is the most direct approach. HTML is the language agents are best at; write HTML, render video.
The comments are more informative than the list itself
Nash replied underneath: agents get layout right now, it's the easing that still gives them away. Layout is no problem now; it's the easing that gives them away.
serx agreed: all my agent animations use the same ease out, and it's obvious at a glance.
Meng To admitted he has been practicing a lot lately. He added that it is now far beyond easing: blender, storyboard, skills, avatar 6, genjutsu are all in use. The point is that the problem has expanded from a single easing function to the entire animation workflow.
Someone asked what the best two are. Meng To's answer was very practical: emil, hyperframes, and remotion all work; let the agent decide which is appropriate.
Side note: the Liquid Glass skill
An interesting item also appeared in the comments. Sougata released an open-source skill that uses WebGL2 to create Apple-style Liquid Glass.

The core idea is worth remembering: most web glass is just backdrop blur, but Apple's Liquid Glass is refraction. Light bends as it passes through rounded edges, each color shifts by a different amount, and there are Fresnel highlights at the edges. None of that can be done with blur.
His solution is a 400-line, zero-dependency WebGL2 renderer plus a tuning lab. The lab contains your real background, real components, and live sliders for every property. Adjust the parameters, and once you're satisfied, copy the JSON to the agent; the agent carries the parameters into the next component. The idea is: stop making the agent guess your taste, and stop trying to describe requirements that can't be quantified--a little thicker, fewer edges, softer light.
Back to the main post. A commenter noted that 'write html, render video with agents' sounds wild in 2026. But looking back at this list, HTML really is the intermediate language agents are best at. Layout is no longer the bottleneck; the next competition is over easing, choreography, and the agent's judgment about what looks good.
发布时间: 2026-11-22 22:17