Taste Skills Playbook
Routing layer for the 13 taste-* skills. Pick one visual-style skill per project, layer universal enforcement and workflow modifiers on top.
Decision Tree
| Task | Primary skill | Layer with |
|---|---|---|
| "Build a landing page" (no style hint) | taste-core | taste-output if long file expected |
| "Premium SaaS landing page, Linear-tier" | taste-soft | taste-imagegen-web if visual-first |
| "Notion-style document UI", "workspace clone" | taste-minimalist | — |
| "Data dashboard", "terminal aesthetic" | taste-brutalist | — |
| "Image-first design then implementation" | taste-image-to-code | + style fork |
| "Generate website reference images" | taste-imagegen-web | feeds into taste-image-to-code |
| "Generate mobile screens" | taste-imagegen-mobile | — |
| "Brand kit", "logo concepts" | taste-brandkit | — |
| "Audit this site", "redesign without rewriting" | taste-redesign | + style fork for direction |
| "DESIGN.md for Google Stitch" | taste-stitch | — |
| Long output (multi-component) | taste-output | layered on every other skill |
The One Style Per Project Rule
Never mix visual-style skills (soft/minimalist/brutalist/gpt) in the same project. They are archetypes, not feature sets. Workflow modifiers (image-to-code, imagegen-*, redesign, output) compose freely on top.
Variance Dials
Three dials drive every taste skill. Defaults 8/6/4.
| Project | Variance | Motion | Density |
|---|---|---|---|
| Marketing landing page | 8 | 7 | 3 |
| Premium SaaS app | 6 | 5 | 5 |
| Data dashboard | 4 | 3 | 8 |
| Documentation site | 4 | 2 | 4 |
| Portfolio | 9 | 7 | 3 |
Skill Graph
ALWAYS-ON: taste-enforcement.mdc (variance dials, banned patterns)
BASELINE: taste-core + frontend-design-taste
STYLE: taste-soft | taste-minimalist | taste-brutalist | taste-gpt (pick ONE)
WORKFLOW: taste-image-to-code, taste-imagegen-web, taste-imagegen-mobile,
taste-brandkit, taste-redesign, taste-stitch, taste-output (layer freely)
Anti-Patterns
- Wrong-style adoption: agent picks
taste-softfor a brutalist dashboard - Style-skill stacking: loading multiple style skills simultaneously
- Skipping image-first: coding from text when visual reference would be better
- Cropping old images: always regenerate standalone images for re-analysis
- Truncated output:
taste-outputmust layer on any task with >2 deliverables