01 · A Claude Code skill
/reskin
Takes any Next.js project. Outputs ten wildly different visual reskins as interactive HTML mockups. Each one commits to a specific design archetype with hard tokens — exact fonts, hex palettes, motion DNA — and gets reviewed by a slop-critic before shipping to the gallery.
The problem
Every AI-built website looks the same. Inter font. bg-indigo-500. Rounded-2xl cards. Centered hero + three feature cards. "Build the future of X."
This isn't a v0 problem or a Bolt problem — it's a base-rate problem. LLMs predict the median of the training data, and the median of AI training data is Tailwind UI defaults.
"Be creative" doesn't move the needle. The fix is to remove the option of generic output entirely.
The fix
- 01
Anchored archetype library
28 named design movements, each with hard tokens (specific font + Google Fonts URL, hex palette, exact radii, motion DNA, real reference URLs). The model can't default when there's no default to fall to.
- 02
Reference grounding
Each parallel sub-agent WebFetches its archetype's reference URLs (Awwwards winners, real production sites) and uses them as visual ground truth. Exemplars do work prose can't.
- 03
Forbidden-patterns blocklist
48 concrete slop signals (Inter, indigo-500, rounded-2xl everywhere, 'Build the future of X' copy) hard-coded as system-prompt anti-rules. Every fork has an explicit NO.
- 04
Slop-critic pass
A separate agent scores each output on 8 axes (typography commitment, palette commitment, layout primitive strength, copy specificity, motion personality, reference fidelity, video-cut recognizability, decision density). Anything that scored below threshold gets regenerated.
Slop is the absence of decisions, not the presence of any particular pattern. At every fork — font, palette, radius, density, motion, voice — commit.
Explore