Simon Willison reports that Claude Opus 5.5 generated an HTML5 Canvas pixel-art animation after being given three kākāpō photographs and a request for at least 20 birds, confetti, and interaction. He then asked a local Claude Code session to create a 15-second video from the downloaded HTML, delaying interaction until three seconds into the recording. The published Playwright script uses a 1280×720 viewport and schedules clicks at multiple canvas coordinates.
The mechanism is browser automation as a capture layer: Playwright opens the local page, reproduces state-changing inputs on a timeline, and records the browser context as video. This turns a web-native interactive into a conventional asset without manually screen-recording it. Why now is that model-generated artifacts increasingly arrive as runnable interfaces, while presentation and publishing workflows still often require video or image files.
The first implication is that teams can treat a generated UI as an executable source asset and automate its conversion into reviewable media. A second implication is that the specification of interaction timing and coordinates becomes part of creative direction, not merely QA. The evidence boundary is that this is a single successful, local demonstration with a simple canvas interaction; it does not establish reliability across dynamic applications, permissions, browser differences, or longer production pipelines. The first watch condition is whether agents can make this capture workflow robust across richer interactive applications without manual scripting.