No build, no deps. Every page ships as a single self-contained HTML file. Open it, it runs.
Nine worlds. Run them here.
Each entry is a finished single-file page. RUN plays it inside its plate. COPY THE PROMPT hands you the full build spec: layer stack, palette, geometry, shader math, timings, and the self-verify loop, round-trip tested against the skill.
Type a brand. Get a world.
This is the input box. The plate on the right is live: it rebuilds as you type. When it feels like yours, copy the prompt. It carries your name, palette, and archetype into Claude Code.
Study anything. Copy nothing.
Clone the feel of a site
Paste a URL into the prompt. The agent browses the reference, storyboards it at five scroll depths, greps its bundles for tech signals, maps what it sees to this skill's recipes — then rebuilds that motion language as original code for your brand. It studies patterns. It never copies code, assets, or content.
The review every page must pass
Rendering correctly is not being designed well. Before anything ships, the skill makes the agent re-read its own screenshots against this checklist. It is why the plates above look the way they do:
- 5-second value scan — what, for whom, what do I get
- Squint test — the CTA pops first, headline second
- Risk-reversal ✓s touch the CTA, never orphaned
- No dead micro-text — big type or cut it
- Ghost wordmark readable, occluded, never cropped
- Visible light — blooms and vignette, no flat fills
- One winner per region of the frame
- Settled in ≤ 2.5 s, never blank under UI
- Every control works — honest labels, real routes
- Hover everything: surfaces respond like lit objects
Two commands, then just ask.
Install the skill into Claude Code with the commands on the right: one clone, one copy.
Copy any prompt from the index, compose one at the bench, or hand a reference URL to the field method.
Watch it verify itself: the agent screenshots desktop, tablet, and phone, and iterates until the page passes the review.