scroll-world (GitHub Repo)

There's a specific kind of website that always gets shared around: the one where scrolling feels like piloting a camera through a miniature world, gliding past a building's exterior straight into its lobby, then drifting on to the next scene without a single hard cut. Scroll-world, a new open-source agent skill from developer oso95, is an attempt to make that effect something anyone can order up rather than something only a studio with a big budget builds by hand.

It works as a plug-in skill for Claude Code, Codex, and anything else that speaks the SKILL.md format, installable through Claude Code's plugin marketplace, Vercel's skills CLI, or just by copying the folder into your agent's skills directory. Once invoked, it doesn't just start generating — it interviews you first, asking about your industry, your pitch, your brand kit, the art direction, and the exact sequence of scenes the camera should visit. It also asks whether you want a mobile version, which matters because the skill renders a separate portrait-format chain for phones rather than cropping the landscape footage down.

Under the hood, the heavy lifting is outsourced to Higgsfield: isometric diorama stills come from GPT Image 2 (either through Higgsfield or, if you've got the Codex CLI, billed to a ChatGPT subscription instead), while the actual camera-flight footage comes from Seedance or Kling image-to-video models — chosen specifically because they can frame-lock a seam so two clips can be stitched invisibly. The skill generates one still and one dive-in clip per scene, then builds connector clips from the neighboring scenes' actual rendered frames so the transitions line up exactly. The scrubbing itself — tying scroll position to video playback time rather than actually cutting between clips — is the same trick behind Apple's product pages, here repackaged as a portable vanilla-JS engine that drops into plain HTML, Next.js, Vue, or a Python-served page.

What's notable is how upfront it is about cost. Generating N scenes needs roughly N image generations and close to 2N-1 video generations, doubling on the video side if you opt into the mobile chain. Since Higgsfield doesn't expose per-generation pricing through its CLI, the skill instead checks your live credit balance and gives you a cost estimate to approve before it spends anything, then runs the whole batch in the background while polling for completion.

The repo itself is fairly small and deliberately unopinionated about your stack — it ships the prompt templates, the batch pipeline scripts, the scrub engine, a minimal HTML template, and a Python script for knocking out backgrounds on floating scenes, all under an MIT license. The generated video and image assets aren't included, since they're produced fresh per project and tied to your own Higgsfield credits.