scroll-world
github.com/oso95/scroll-world- Category
- AI Agents
- Rank
- No. 2065Tools index
Previous survey · No. 2074 ·
- Pricing
- Open Source
- Type
- TOOL
- Use case
- Design & Media
- Interfaces
- Agent Skill / Plugin
- Builder
- oso95
- GitHub
- 9.8k stars
- Date
About
scroll-world is an agent skill that generates an immersive, scroll-scrubbed landing page in which a camera flies continuously from outside each 3D scene into its interior and on to the next without cuts. It works with Claude Code, Codex, and any SKILL.md-compatible agent, installable as a Claude Code plugin or through the skills CLI.
What it does
A set of agent instructions plus helper scripts that turns a brand brief into a website hero made of pre-rendered video. It asks about the business, palette, art direction, camera style and budget, then generates one image per scene, one video clip diving into each scene, and joining clips between neighbours. Each joining clip is rendered from the real last and first frames of the clips on either side, so the seams match exactly. A small vanilla JavaScript player then maps page scroll to video time, so the reader controls playback by scrolling.
Why it's ranked here
The hard part of this genre is the seam between clips, and the procedure is built around it: joining clips must start and end on frames pulled from the rendered videos, never the original stills, and the docs call a mismatch the most common failure. The player is MIT licensed, has zero dependencies and loads each clip as a whole blob so it never relies on server byte-range support. Costs are measured and stated up front, with a six-scene 1080p chain quoted at about $27 on the default video backend.
What's good
Budget is a gate, not an afterthought: the skill states an estimated total with roughly 15% re-roll headroom and waits for approval before rendering anything. The phone handling is specific: seeks are coalesced so fast flicks cannot freeze the video, the poster stays up until a real frame paints, and reduced-motion visitors get cross-dissolving stills with no video loaded at all. The mobile option renders a separate native portrait chain rather than cropping the landscape film. The encode settings explain why a tighter keyframe interval makes phone scrubbing smoother.
Tradeoffs
Every build spends money on paid generation services: video through Monid by default, scene images through Higgsfield credits or a ChatGPT subscription via the Codex CLI, plus ffmpeg and Python with Pillow locally. Generations take minutes each and run in the background. The mobile version roughly doubles the video spend. The docs say seam behaviour when a chain is finished on a different provider is untested, and no test directory appears among the files.
How to use it well
Fits a marketer or developer who wants one cinematic scrolling hero for a brand and is willing to pay per clip for it. Use the cheap 480p or draft tier to preview the whole journey, approve it, then re-render final legs at full quality. Keep each scene's subject centred, check every clip's last frame before chaining the next, and keep five to seven scenes. It does not cover the rest of a site: the output is a scroll section, not a full page builder.
Technical notes+
skills/scroll-world/SKILL.md is the procedure (bootstrap, interview, budget, seam rule). skills/scroll-world/references/pipeline.md holds bash 3.2 safe scripts: gen_still / gen_dive / gen_conn against the Higgsfield CLI, ffmpeg boundary-frame extraction with -sseof -0.15, libx264 crf 20 GOP 8 encodes (mobile crf 23 GOP 4), and Monid helpers monid_frame_url and monid_wait that upload frames to sfs and poll run status. skills/scroll-world/references/scrub-engine.js exports mountScrollWorld(container, config), interleaves dive and connector segments, applies a lingerEase remap per section, and falls back to a crossfade when a connector is null. skills/scroll-world/references/knockout.py does a border-connected flood fill with a TOL env var (default 34) and a 1.4px Gaussian alpha blur.
Observed
- License
- MIT
- Form
- Agent skill: a procedure document plus reference scripts, prompt templates and a JavaScript scrub engine
- Install surface
- Claude Code plugin marketplace, the skills CLI via npx, or manual copy into an agent skills directory
- Languages
- Markdown instructions, bash scripts, vanilla JavaScript, Python
- Runtime dependencies
- Monid CLI (default video backend), Higgsfield CLI, ffmpeg and ffprobe, Python 3 with Pillow; Codex CLI optional
- Scrub engine dependencies
- Zero dependencies; does not require HTTP byte-range support
- Accessibility
- Under prefers-reduced-motion the engine never loads clips and cross-dissolves the stills
- Tests
- No test directory among the repository files
Read from README.md, LICENSE, .claude-plugin/plugin.json, skills/scroll-world/SKILL.md, skills/scroll-world/references/pipeline.md, skills/scroll-world/references/prompts.md, skills/scroll-world/references/scrub-engine.js, skills/scroll-world/references/knockout.py.
Tags
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.