The taste stack
AI agents can produce interfaces on demand — and almost all of them feel the same. The bottleneck moved from writing code to knowing what good looks like. This is a working stack for that problem: one tool generates with taste built in, one judges it against a declared target, one measures it in public, and one learns from every build. Each piece works alone; together they close the loop.
01 · Generate
Gives any coding agent (Claude Code, Cursor, …) the craft to build cinematic, scroll-driven sites: pinned chapters, parallax, 3D, editorial pacing — run through a 5-phase pipeline with taste guardrails and a transform/opacity performance budget. Every build is gated by its own quality scorer before it ships.
5 live demos · 11 themes · self-gated at ≥90/100
02 · Judge
Turns taste into a declared, checkable target: a 30-axis design grammar, 101 brands published as machine-readable vectors, and a deterministic judge that scores any URL against a brand, a live reference, or a pinned grammar — returning per-axis failures and a prescriptive fix-list, not just a number.
101 brands · 30 axes · no-auth API + MCP + CLI
03 · Benchmark
One rubric over the web's best-known sites: pacing, performance, motion accessibility, craft — measured passively (one ordinary page load, robots honored, truthful UA), median of three runs, deterministic scoring. Then run the exact same scorer on your own site with one command.
61 sites ranked · npx cinematic-bench <url>
04 · LearnStudio · paid
The self-improving layer: it distills patterns from sites you're authorized to study into an originality-firewalled pattern library, tracks which motifs actually get reused, scaffolds new variants against a declared taste target, and promotes nothing to canon without human sign-off. The skill that gets better every time you build.
pattern IR · reuse telemetry · judged variants · human-gated canon
A generator without a judge drifts generic. A judge without a benchmark has no public ground truth. A benchmark without a learner is a scoreboard nobody improves against. Wired together: builds declare a taste target before a line is written, get scored against it after, land on a public rubric anyone can verify — and the patterns that survive feed the next build.
Everything measurable here is deterministic and open: published weights, reproducible scorers, audit trails. The judgment stays human where it should — nothing is promoted to canon by a machine.