Taste Lab

Taste takes a URL and runs it through a four-step agent pipeline that terminates in two artifacts: a `{domain}.md` and a `{domain}.json`. The first agent extracts measurements — every color, weight, spacing value, radius, and shadow, cited with exact px, hex, or ratio. The second detects 5–8 system-level rules, each with an evidence line and a stated design goal. The third infers four ‘taste principles’, each carrying a trigger, decision, reason, evidence citation, and explicit trade-off. The fourth agent acts as a quality gate: it runs an anti-slop grep, validates the JSON, and discards anything that didn’t survive scrutiny. At least one of the four principles is required to be a restraint principle — a documented choice about what the designer left out.

The differentiating claim is in what the output carries beyond tokens. A spec sheet tells an agent the border radius is 6px. A Taste DNA principle tells it that 6px was chosen for primary cards, that 2px was reserved for micro-elements, and that the decision trades warmth for precision — so when the agent encounters a component the spec never covered, it has the reasoning to make a consistent call rather than defaulting to its training data’s average.

The tool fits inside workflows where you are supplying design context to Cursor, Claude, or Gemini via rules files — the .md output is structured for that handoff. It is self-hostable, free, and the vendor describes it as a GitHub-installable skill with no paid tiers. Where it breaks is on sites with no coherent design system: the pipeline is built to infer intent, and on a site assembled from mismatched libraries or iterative patches, the principles it produces will be internally consistent but factually wrong about what the designer intended. Teams working against those targets will need to treat the output as a starting draft and edit before committing it to agent context.