Can AI rebuild a website pixel-perfect — and then take it over?

Hypothesis

Our old website was locked inside a site-builder: slow, unownable, hard to extend. Could AI agents reverse-engineer it — fonts, layout math, scroll effects and all — into a modern codebase pixel-perfect enough to silently replace production?

What we built

The site you're reading. Agents crawled the original builder-hosted site, extracted its design tokens (custom fonts pulled out of a page mirror, exact image dimensions and stagger margins, background parallax behaviour), and rebuilt all thirteen routes as a Next.js app. The cutover happened quietly: the rebuilt site became www.thrivbe.com, and everything since — this labs notebook included — builds on it. The process was packaged as a reusable cloning skill, so the next site takes hours, not days.

Learnings

• Pixel-perfect is a verification discipline, not a vibe: screenshot diffs at fixed viewports, and computed-style probes for behaviours a static screenshot can't distinguish (a scroll effect that's frozen looks identical to one that works).
• Site-builders hide their design system in plain sight — mirrored HTML contains the fonts, sizes, and layout constants as literal values. Treat the artifact as source, don't reinterpret it.
• Owning the codebase changed the site's metabolism: experiments, a labs notebook, and API routes appear in days — none of which the builder could host.

Log

• 2026-06-06 — Rebuild became the production site; cloning process packaged as a skill.
• 2026-06-01 — Crawl + design-token extraction of the original site.