Creator demos Claude Code skill that builds scroll-driven 3D brand sites using Higgsfield video and Three.js

• Orchestrator: Claude Code running a private skill. No Claude model version is named.
• Video: Higgsfield calls "Kling" for 5-second 1080p clips (04:07). The Kling version is not given.
• 3D: ChatGPT in the Codex app makes a product photo on a white background, then Higgsfield turns it into a 3D model (05:35). Three.js renders it on the page.
• Access: Higgsfield's connector needs a paid plan, and every agent-run generation spends credits (Higgsfield help). The skill is only available behind the paywall.

• Pipeline as described: a five-scene storyboard (problem first, product as the answer, calm ending), then prompts written like camera direction, then video clips, then an image-to-3D model, then a site build, then QA on desktop and mobile that looks for flicker and overlap frame by frame (04:30, 09:47).
• Skills mechanics: only the skill's description stays in context until it is invoked. Anthropic's skill-creator plugin can run evals with and without a skill (Claude Code docs). The video reports no such eval.
• Design rules in the skill: use real website structure (menu, hero, product section, FAQ, footer), give each brand its own palette, font and symbol, and use five ways to embed video, such as video inside text and pinned split screens (08:01).
• Reproducibility: the skill, prompts and code are not public.

• "Four sites in a couple of days," 34 scenes, zero hand-written code: the creator's own account only (10:52). The video also says "we build all of this in the morning," which conflicts with "a couple of days" (10:55). The creator profits from skill sales.
• "Agency-level" quality: no outside review, performance or accessibility data.
• Higgsfield makes 3D models: not confirmed by the reviewed Higgsfield docs or by ClaudeFast's MCP overview, which cover image and video only (ClaudeFast).
• Independent corroboration: none found.

Open-source versions of the same idea already exist:

• scroll-cinematic-claude uses Higgsfield clips cut by FFmpeg into about 180 frames, scrubbed on a canvas with Lenis smooth scrolling. It claims about $1–2 in credits per site.
• scroll-world uses GPT Image 2 with Seedance.

A tutorial points out that the Apple-style "3D scroll" effect is often a canvas image sequence, not Three.js (Tenten tutorial). The video's design adds an actual Three.js product model on top of that.

• The creator says the first versions looked "amateur": busy code-drawn particles and labels, then pages full of video that had no menu or shop (06:42). Human review of all 34 clips was still needed (05:28).
• Higgsfield job statuses include failed and NSFW-flagged outputs, and video can take several minutes per job (ClaudeFast).
• Claude Code docs note that skill shell commands go through permission checks and that organisations can switch skill shell execution off (Claude Code docs). That matters when installing skills from third parties.
• The page weight of video-heavy sites is not addressed in any reviewed source.

• Cost shifts from labour to credits: agent-run generation always spends credits (Higgsfield help). Budget for regenerating clips, not just one pass.
• The skill itself is a weak moat: free open-source skills cover most of the same pipeline, so check them before paying for gated ones.
• Taste still decides the result: storytelling, brand systems and QA rules mattered more than the code, by the creator's own account.
• Before production: test load time, mobile data use and accessibility yourself. None of the reviewed sources measure them.