A skill for your AI agent. Hand it your screenshots and it writes one spec file, then renders a 16:9 cut and a 9:16 short in this style. Rendered by the kit, not edited after.
The look
Captions rise out of a blur one word at a time, with one accent word.
Dark chapter wheels between light scenes.
Screenshots in a floating window with numbered badges, a camera push-in, and one card per screen.
Typing, a terminal, a browser frame, a closing lockup.
In the zip
SKILL.md: a 14-step workflow from your screenshots to a checked render.
The kit: 11 scene types, 3 colour themes, render scripts and a preflight check.
A caption-shake check and a -16 LUFS music master.
Two working demos (16:9 and 9:16) to render first.
Six reference docs, three example videos and 17 stills.
You need
Node 20+ and ffmpeg. About 1 GB of free disk.
An AI coding agent that runs shell commands and reads images, such as Claude Code.
Your own screenshots and copy. No API keys, accounts or video editor.
Good to know
Tested on Apple Silicon macOS with Node 26 only.
The 65 s demo renders in about two minutes.
Two examples show the Claude app (Anthropic’s UI). Watch them, do not reuse them.
Uses HyperFrames (Apache-2.0) and GSAP (own free license; read it before commercial use). Fonts: SIL OFL.