Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)

Gemini built a 15-second, 13-scene video from a single prompt, in code, without video generation

A video is usually made with a video model, but Gemini was asked to do without one. From a single prompt, it produced a program that draws and voices 15 seconds of animation on its own. It is a web page with 13 scenes, the anime hero AO drawn with vector code, and music synthesized by the browser. Making an edit means changing a number in the code, not regenerating frames. For that reason, titles and motion graphics are cheaper to order from a model as code.

AI-processed from TG: @cgevent (Метаверсище и ИИще); edited by Hamidun News
Gemini built a 15-second, 13-scene video from a single prompt, in code, without video generation
Source: TG: @cgevent (Метаверсище и ИИще). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @cgevent (Метаверсище и ИИще) ↗
Editor-in-Chief
Canvas at 1280×720 (169), with a 24 or 60 fps mode
Key metric #2WebM export with sound and a timeline of cuts on the page
Key metric #3The finished page can be published to Google Sites with one button

Usually, to get a short video from a neural network, people use a video model. It draws the frames, each attempt costs money, and a small fix like "move the title 0.3 seconds earlier" means paying for a new generation. The author of the Telegram channel "Metaverseshche and AIshche" showed a different route. He banned Gemini from video generation and gave it a single prompt. In response, he got a program that draws and voices 15 seconds of animation from 13 scenes, featuring an anime hero named AO. So the video here is not generated as video. It is written as code and plays right in the browser.

For the past few weeks, feeds have been full of clips where Claude Code builds motion design: titles fly in, a grid rotates, everything is synced to music. The author took the prompt from his previous post and gave it to Gemini in the Gemini app. He set strict conditions: no agents and no studios, one request, and no video generation.

What the model actually generated

According to the model's own description, it assembled a single HTML file, ao_reveal_trailer.html. It contains several layers:

  • An HTML5 Canvas at 1280×720 in 16:9 format. This is an ordinary drawing surface in the browser, where everything is rendered frame by frame.
  • A procedural vector rig for the character AO. The hero is not a drawn image but is described in code: shapes and "bones" that can be moved.
  • A custom 3D projection for typography and the grid. Volumetric text and the perspective grid are computed with formulas.
  • A Web Audio API sound engine. The electronic soundtrack and effects are synthesized on the fly; there are no ready-made audio files.

The video consists of 13 scenes with a total length of exactly 15 seconds. It has a switch between 24 and 60 frames per second, an interactive timeline of cuts, and a button to export the finished video with sound in WebM format.

Why this matters for practitioners

Conventional video generation works like a lottery: each attempt requires computation, and a precise edit is nearly impossible without regenerating everything. The "video as code" approach works differently. The expensive part happens once: the model writes the program. After that, playback runs on your device, and any edit comes down to changing a number or a line of text.

For motion design, this is especially convenient. Typography, lower-thirds, titles, and interface animation map well onto formulas and vectors. A photorealistic cinematic shot cannot be obtained this way; video models are still needed for that.

Another detail from the post: after generation, a button appears in the app for publishing to Google Sites. The author published the finished result, with control buttons at the top and bottom of the page. The chain of "idea, code, interactive publication" closes within a single app.

What to keep in mind

This is a personal demonstration post, not an independent test. The author gives no token counts, no generation time, and no number of attempts. There is also no comparison with what Claude Code produced under the same conditions, so it is too early to say who did better. The "Gemini Neon" mode name in the post is also not explained, so nothing definitive can be said about the specific model or plan.

WebM export from the browser usually depends on the browser's own capabilities: recording the canvas and audio happens in real time. That means a 15-second video takes about 15 seconds to record, and frame drops are possible on weak devices. This is a general property of the approach, not data from the post.

The author also notes that Claude Motion, by his account, is currently available only to the enterprise segment, and he believes Anthropic has reason to worry. This is his assessment, not a confirmed fact.

Practical takeaway

If you need short brand intros, announcements, explainer animations, or teasers with clear geometry, try asking a model to build them as a web app rather than generating video. Specify the resolution, duration, number of scenes, frame rate, and export requirement right in the prompt. That keeps the result reproducible and makes edits cheap. For photorealistic scenes you still need video models, but for motion graphics the programmatic route looks like a serious and inexpensive alternative.

Story

Key Questions Answered

What the model actually generated?

According to the model's own description, it assembled a single HTML file, ao_reveal_trailer.html. It contains several layers: An HTML5 Canvas at 1280×720 in 16:9 format. This is an ordinary drawing surface in the browser, where everything is rendered frame by frame. A procedural vector rig for the character AO. The hero is not a drawn image but is described in code: shapes and "bones" that can be moved. A custom 3D projection for typography and the grid.

Why this matters for practitioners?

Conventional video generation works like a lottery: each attempt requires computation, and a precise edit is nearly impossible without regenerating everything. The "video as code" approach works differently. The expensive part happens once: the model writes the program. After that, playback runs on your device, and any edit comes down to changing a number or a line of text. For motion design, this is especially convenient.

What to keep in mind?

This is a personal demonstration post, not an independent test. The author gives no token counts, no generation time, and no number of attempts. There is also no comparison with what Claude Code produced under the same conditions, so it is too early to say who did better. The "Gemini Neon" mode name in the post is also not explained, so nothing definitive can be said about the specific model or plan.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…