nateherkai/scroll-craft

Selection

Select when

• You need a distinctive scroll-driven marketing or portfolio page.
• You want Claude Code to follow a structured design-to-verification workflow.
• The site needs layered heroes, video scrubbing, pinned sequences, reveals, or horizontal rails.
• You have authentic brand media or accept optional third-party generation cost.
• You can install Node, ffmpeg, Chrome, and playwright-core in the build environment.
• You want semantic HTML rather than runtime-generated page structure.
• You need explicit mobile composition and reduced-motion checks.
• You want safeguards against repeating prior site structures and interactions.
• You can perform human visual, keyboard, device, and deployment review after automation.

Boundaries

Avoid when

• You need a conventional static page with minimal interaction.
• You want a no-code visual editor or turnkey hosted site builder.
• You cannot install the documented media and browser dependencies.
• You require verified native macOS production history from the supplied evidence.
• You require automated proof of real iPhone behavior.
• You need broad browser, assistive-technology, or formal accessibility certification.
• You want common template conventions that the skill explicitly refuses.
• You cannot review generated assets, screenshots, motion quality, and factual content.
• You need deterministic design outcomes without model or human judgment.
• You cannot send prompts or assets to kie.ai but also lack suitable source media.

Strengths

Capabilities

Risk profile

Risks and limitations

• Uncertainty: only the README, plugin manifest, and SKILL.md were supplied, so scripts, engine code, references, templates, tests, CI, and package integrity were not directly inspected.
• Uncertainty: repository metadata is dated in 2026 relative to this analysis context, so activity, adoption, and recency cannot be independently reconciled.
• Uncertainty: the repository reports ten approved immersive websites, but their source packages, acceptance criteria, and production outcomes were not independently examined.
• The author states that completed builds have only been run on Windows, although scripts search Windows, macOS, and Linux locations.
• Node 18+, a full ffmpeg build, Chrome, and locally resolvable playwright-core are required for the documented workflow.
• Real iPhone video decoding, autoplay policy, Low Power Mode, and touch scrolling are outside the headless verification harness.
• The harness cannot judge composition quality, motion smoothness, emotional effect, factual accuracy, or whether the page is meaningful.
• Optional kie.ai generation introduces third-party cost, network transfer, credential handling, and provider availability concerns.
• The workflow is intentionally opinionated and may conflict with established design systems or requested conventional patterns.
• Persistent build artifacts, screenshots, briefs, and fingerprint records may contain project-sensitive material and require normal workspace controls.
• No minimum Claude Code version, cross-host compatibility matrix, automated integration result, security audit, or privacy policy was supplied.
• The evidence does not establish comprehensive browser, assistive-technology, performance, or production deployment coverage.

The skill permits shell, file-editing, search, and user-question tools, while its workflow adds browser automation, local serving, media encoding, and optional remote generation. First-party text calls for disabling native pointer lock and capture during headless checks, preserving real markup, and reporting unverified behavior. Uncertainty: implementation-level path handling, subprocess safety, dependency pinning, secret protection, network scope, and script integrity were not inspected, and no audit or threat model was supplied.