% open teaching-claude-taste
Teaching Claude Taste
Gave Claude design taste, then ran v1 vs v2 to prove it landed.
PROBLEM
AI makes building a website easy, but the output has a "stock AI feel" — same fonts, same layouts, the generic look you can smell. Prompting harder doesn't fix it.
WHAT I BUILT
A foundation of reusable design + SEO skills for Claude — taste it keeps and reuses on every build — then I re-ran a real site build through it to see if it actually made a difference. The site was our Korean restaurant, WATDA.
HOW IT WORKS
v1 was a quick build to ship and test. Then I built the foundation: a web-design skill (real design principles, a vetted toolstack, drop-in style tokens, a pre-ship checklist) plus two existing design skills, with SEO/structured-data on top. v2 is the same content, photos and brand, rebuilt through the equipped stack — a like-for-like comparison.
WHERE IT'S AT
Both built. v2 is live at watda.co.nz and replaced v1. It visibly drops the stock-AI sameness — better spacing, better type, and knowing when to leave things alone. The skills now carry over to every future build (including this site).
WHAT I LEARNED
Taste isn't a prompt, it's a foundation. Once it's encoded as skills Claude reuses, every later build starts ahead instead of from the same generic baseline. The leverage is in building the system once, not re-prompting forever.
STACK
Claude Code skills (web-craft + 2 design skills) · Astro 5 · Tailwind 4 · Cloudflare Pages
% cd ..