What they did
The author, artiomleo, needed an overview video and seven short how-to clips for a web app they work on. Claude Code wrote roughly nine sentences of narration per clip, then ran Kokoro, an open 82M-parameter voice model, through a Python script on the laptop to read each sentence and log when it starts. It also wrote a 163-line Node recorder that drives a windowless Chrome over the DevTools Protocol, captures a frame on every repaint, draws a fake mouse pointer, and fires each click or keystroke when its sentence begins. ffmpeg joins the frames and audio. Each clip is a few text files: the script, the voice output and a list of 9 to 32 actions.
How it went
All eight clips were made without anyone filming the screen. Chrome only delivered about 42 JPEG frames a second, so the agent slowed the page's clock to half speed with an injected script, giving roughly 84 frames per finished second, then ffmpeg evened it out to 60 fps at 2880x1620. The author picked the voice from 13 options, watched each take and asked for changes. A planned hosted voice service was dropped for the free local model before anything was voiced.
Worth knowing
Only what happens inside the browser tab can be recorded. File pickers and sign-ins on other sites were avoided by passing files straight to the page and using a demo account with fake, fixed data and links already connected. The screencast command used is still marked experimental in Chrome.