A small startup has built a system that takes a single prompt, hands it to Claude 5.5, and spits out a rendered explainer video in minutes. The company behind it, DiggerHQ, calls the process Ship a launch video. The whole thing runs on OpenComputer, a serverless platform, and the final output is an MP4 you can watch.
Here is how it works, step by step.
Prompt In, Video Out
The system takes a URL or a prompt as input. Claude 5.5 generates the script as code, then a serverless agent renders it into frames. The rendering replaces the browser’s clocks with a virtual clock, so every frame is a deterministic seek — meaning the same prompt always produces the same video. The result is 1920×1080 at 30 fps, compressed with libx264 and a crf of 18.
The agent holds no secrets. Each job gets a temporary Vercel Blob upload token, scoped to one path for three hours, which the finished MP4 lives on as a public URL.
The Tools That Make It Work
The agent is built from three tools:
- web_fetch: Reads the product’s site, pulling in page text, titles, headings, the most used hex colors, and Google Fonts.
- check_scene: Loads the generated HTML, reports errors, and captures visible text at sample timestamps.
- render_video: Runs Playwright’s headless Chromium, renders the frames, and uploads the MP4.
Each tool call happens in a fresh Amazon Linux 2023 microVM on arm64, with 4 vCPUs and 8 GB RAM, running Node 22. The VM installs Playwright and ffmpeg in about a minute, then is thrown away.
How Claude 5.5 Writes the Film
Claude 5.5 handles the heavy lifting of writing the film. The model takes roughly 90k input tokens and 15k output tokens per video, most of which is the HTML itself. The remaining tokens cover the script and scene descriptions.
The process is remarkably lean. There is no video model doing the rendering. The HTML is assembled by Claude, then handed to a headless browser that renders it frame by frame. The final output is a deterministic MP4 that anyone can watch.
What Makes This Different
This is a neat trick. A prompt turns into a rendered video, shot by shot, in a matter of minutes, with no editing needed. The deterministic rendering means the same prompt always produces the same output, which is a powerful property for anyone building documentation or tutorials.
The system is also remarkably lightweight. The agent is a single TypeScript file deployed with opencomputer deploy, with no framework, no queue, and no server of the company’s own. Everything runs in a transient microVM that disappears after each job.
The storage model is clever too. Each job gets its own temporary token, so the finished MP4 is publicly accessible but scoped to a single path for three hours. That means the video exists only as long as it is needed, which is a sensible design for a system that generates content on demand.
Our View On It
The system is a proof of concept for a larger trend: AI models writing code that renders itself. Claude 5.5 is good at explainer videos, and this demo shows exactly how that works in practice. The deterministic rendering is the star of the show — it is the feature that makes the whole thing useful.
The company has made the entire setup available as a repository, diggerhq/shipvideo, and it deploys with a single command. Anyone can try it out.
The demo has drawn attention from readers who follow the site closely.
Source material: “Opus 5.5 is good at explainer videos,” launchvideo.io.
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.