oso95/scroll-world

scroll-world is a skill for AI coding agents like Claude Code and Codex that builds a scroll-driven landing page where the camera flies through a series of three dimensional looking scenes, one connected to the next with no cuts. As you scroll down the page, it feels like flying into and through a small generated world built around a brand or industry, similar to how some product pages use a continuous camera movement tied to scroll position. The skill relies on Higgsfield, an external service, to generate the visuals: isometric diorama style still images for each scene and short video clips that show the camera moving into and between them. It also supports using the Codex CLI to generate the still images through a ChatGPT subscription instead of Higgsfield credits. Once the images and clips exist, a small vanilla JavaScript scroll engine plays them in sequence, tied to how far the user has scrolled, and works in plain HTML, Next.js, Vue, or any page served by Python since it does not assume any particular framework. When someone uses the skill, it first asks questions: what the site is for, the brand look and feel, the order of scenes, whether a separate portrait version should be made for phones, and how much to spend, since generating images and clips costs credits. It then generates the still images, the camera flight clips for each scene, and connector clips that join scenes together so the transitions match up cleanly. Finally it wires everything into the scroll engine, including serving the phone friendly version automatically on mobile devices. To use it you need the Higgsfield command line tool with an authenticated account and credits, ffmpeg for handling video frames, and Python with Pillow for building the portrait versions. The generated video and image files are created per project and are not included in the repository itself. It is released under the MIT license.