↩ tastecheck
Theme
design intelligence · 20 skills · roll for taste

Make it yours, not the average.

Every AI builds the same site. tastecheck interviews you or audits what you already have into a real design system first — then nineteen skills make sure your agent ships taste, not slop.

Landing page audit ready.

Works with Claude Code · Codex · Gemini · Cursor · MIT

v1.4.1 20 skills 3 P0 fixed 8 P1 fixed 7 P1 deferred 0 P2 regressions

Same content, eight committed design systems → see the gallery

20 concerns, represented

Not a claim of execution. Each cell shows what the skill owns.

This whole page was built by them. Hover, type, toggle — they’re real.

00

design-system-interview

“‘Modern’ is a non-answer — name one you’d resemble.”
basalt + glaze
→ tokens committed

Grills you into a system + tokens before any code.

🎲

tasteroll

seed 42
ref vibe mode sig
seed 42 · click roll

Audit fixes what is broken. Roll picks the rest. Lock what works, re-roll.

01

improve-existing-website

this page: nav, CTAs, type, color, states, proof
→ existing system tightened on the live landing page

Reads the real site, separates signal from drift, then makes the partial system true.

02

color-system

OKLCH · copper × verdigrisAA 5.4:1 ✓

Perceptual ramps that pass contrast — math, not vibes.

03

web-typography

AaRedaction · perfect fourth · 66ch

Scale, measure, rhythm — fluid & legible.

04

component-states

click → cycles real states

Hover, focus, active, disabled, loading — real.

05

form-ux

Enter a complete address; it checks when you leave the field.

Labels, truthful validation timing, specific inline errors, recovery.

06

empty-states

Loading preview.

Loading · empty · error — handled.

07

theming

LightDark
light + dark + high-contrast · one token source

Supported mappings from one token set; explicit preference and forced-colors behavior.

08

micro-motion

Hover me ↗transform · 180ms · reduced-motion safe

Purposeful feedback with interruption and reduced-motion paths.

09

responsive-layout

intrinsic · container queries

Survives any width — it just flows.

10

a11y-pass

WCAG 2.2 AA ✓

Keyboard, focus, contrast — fixed, not bolted on.

11

deslop-ui

Get started
✕ pill · purple
Get started
✓ 5px · copper

Strips the named AI tells; commits to a direction.

12

humanize-copy

“In today’s fast-paced landscape we leverage synergies to delve…”
“Deploy logs in plain English — what shipped, what broke, what fixed it.”

Strips the ChatGPT accent — a checkable kill-list.

13

data-viz

Quarterly revenue reaches $4.2M Seven quarterly points rise from 3.1 million dollars to 4.2 million dollars, a 31 percent increase. $4.2M
Quarterly revenue values
Q4 2024$3.1M
Q1 2025$3.2M
Q2 2025$3.35M
Q3 2025$3.5M
Q4 2025$3.9M
Q1 2026$3.75M
Q2 2026$4.2M
Revenue +31% · honest axis · direct label, no legend

Tufte-informed charts: data-ink, no chartjunk, real table fallback.

14

cognitive-a11y

TL;DRplain language · one idea per line · predictable layout

Readable for ADHD, autism & dyslexia — beyond WCAG.

15

spacing-system

one --space-* scale · section rhythm

Consistent spacing & vertical cadence from one ramp.

16

i18n-ready

Get started
Comienza ahora mismo
EN/ES first-class · longest approved locale fixture · logical properties

Multilingual-resilient: text expansion, Intl, native voice per language.

17

art-direction

one committed icon set · real OG/favicon

Imagery & icons as decisions — no gradient-blob AI graphics.

18

tastecheck-pass

✓ tokens · ✓ contrast ≥ 4.5:1 · ✓ keyboard
✓ cold-load render · ✓ no slop tells · ✓ structural
✓ 20 skills audited · ✓ 11 findings fixed
✓ blind eval rehearsal · 160 sim / 0 real
Gate: PASS ✓

The ship gate: runs every self-check, reports a pass/fail table — with a pasteable auditor.

+

plain markdown

Claude · Codex · Gemini · Cursor · Kimi
plain markdown · no runtime

Canonical .agents links plus detected skill dirs.

v2

blind evaluation

■ 48 generation ■ 96 judgment ■ 16 anchor
160 simulated · 0 real · zero spend · sealed authority

Machine-only blind comparison: rehearsed, adversarially tested, production not started.

what taste refuses

There isn’t one AI slop. There are three.

You just watched the skills work. Here’s the principle behind them: slop isn’t only the look — the skeleton repeats, and so does the copy. We refuse all three planes.

The look

Purple gradient or editorial serif — same average, one season apart.

The structure

Centered hero, 3 cards, equal grid — same bones under any skin.

The copy

delve · in today’s fast-paced world · it’s not just X, it’s Y

proof

This landing page is the integration surface.

The GitHub Pages homepage represents every skill concern directly and demonstrates named interactions: existing-site audit, inferred design system, OKLCH themes, type, states, form validation, empty/error/retry paths, chart/table parity, accessibility, human copy, motion, and responsive layout. The twentieth skill, tasteroll, is a procedural design generator — it audits what is broken, rolls committed design candidates from a seeded source of randomness, and locks the direction that works. A machine-only blind evaluation protocol (effectiveness v2) is built, rehearsed at 160 simulated calls with zero real calls, and adversarially tested — but production evaluation has not started.

install

One command. Then just build.

Links canonical skills and mirrors them into detected agent homes.

$ git clone https://github.com/KyaniteLabs/tastecheck
$ ./tastecheck/install.sh