Three areas have been improved:
More reliable data collection
The extractor used to be all-or-nothing: if one part failed, everything failed. Now each section runs independently if button detection breaks on a weird site, you still get colors, typography, and everything else. The report tells you which parts degraded so you know what to trust.
Taste Lab
PicWish
@sunlinsen love it.
@sunlinsen The 'delete any principle that could've been written without seeing this site' rule is the sharp part. How do you stop it from over-fitting, treating one page's quirks as the brand's actual taste? Curious where you draw the line between a site's real signature and just noise on that specific page.
Brew
Very cool! Does it provide md files?
Taste Lab
very cool idea! does it capture interaction and motion too (transitions, easing), or is it focused on the static layer for now?
The "why" behind design decisions is exactly what's missing when you hand off to AI agents. Tried it on a few sites and the output is way more useful than just extracting colors and fonts. Nice work, upvoted!
The part I'd try first is running this on a messy personal portfolio, then using the breakdown as a prompt for a redesign instead of starting from a blank style guide. I like that it tries to explain why the spacing and type choices work, not just list tokens. Does Taste Lab keep a reusable brief/history for a site, or is each scan meant to be a fresh one-off export?
Love the concept. Have you experimented with analyzing entire product ecosystems rather than individual pages? I'd be curious to see how Taste Lab handles consistency across marketing sites, product dashboards, and mobile experiences. 🤔
Websites are becoming increasingly animated, how does it captuer animations styles? and unconventional animations?