← ClaudeAtlas

tasterolllisted

Rapid prototyping through constrained randomness. Audits what exists, fixes what is broken, generates fresh design candidates from context, then rolls the dice between valid alternatives. Lock what works, re-roll what doesn't, converge on a direction. Use for design exploration, rapid prototyping, or when a user wants surprise within guardrails.
KyaniteLabs/tastecheck · ★ 7 · Web & Frontend · score 71
Install: claude install-skill KyaniteLabs/tastecheck
# Tasteroll (roll for taste, keep what sticks) Most design generators either ask you everything (the full interview) or nothing (press a button, hope for the best). Tasteroll is the middle path: it audits what exists, asks at most three questions, fixes what is broken, then rolls between valid alternatives. Lock what works, re-roll what doesn't. The thesis: **constraints produce taste, even when the choices are random** — but only if the constraints come from the actual project, not a fixed list. ## The pipeline ``` 1. AUDIT → scan what exists → mandatory findings (craft-only) 2. INTAKE → infer from context/repo OR 3-question mini-interview 3. FIX → resolve all findings (seed-independent, non-negotiable) 4. GENERATE → fresh candidates for open dimensions (context-aware) 5. ROLL → seeded pick between valid alternatives 6. LOCK → keep what works, re-roll what doesn't, converge 7. GATE → relaxed gate (readable + WCAG AA + self-contained + inert) ``` ## 1. Audit (findings are mandatory fixes) Scan rendered surfaces, source CSS, repo files, and content structure. Findings are **non-negotiable** — they get resolved before the dice touch anything, regardless of seed. The audit scope is **craft-only**: - **Contrast**: body < 4.5:1 or UI < 3:1 on actual surfaces - **Spacing**: no scale, arbitrary margins with no token system - **States**: interactive elements missing hover/focus/active/disabled - **Tells**: named AI slop patterns present (gradient,