A design discipline that rewires LLM defaults — replacing centered heroes, Inter font, purple gradients, and placeholder code with testable, named bans and parametric controls.
LLMs have hardcoded priors toward mediocre output: centered heroes, Inter font, purple-blue gradients, 3-column equal-width card grids, h-screen, and rounded-full on every button.
Generated code trails off with // ... placeholders, "for brevity" hedges, and "rest of the code here" elisions. The output looks complete but isn't.
"Elevate Your Workflow." "Seamless integration." "Next-gen platform." LLMs reach for the same startup-slop brand names and marketing clichés every time.
Same prompt, same model — 8 banned patterns in a single hero component: centered hero, h-screen, purple gradient, Inter font, "Elevate" cliché, "Seamlessly" cliché, "next-gen" cliché, rounded-full CTA.
Named, hard bans that are checkable in 5 seconds by visual inspection or grep. Three parametric dials that act as global configuration variables. No vibes — only switches.
50+ named forbidden patterns merged across all archetypes. Each ban has a 5-second test: grep the source or inspect the rendered page. Inter font, purple gradients, h-screen, placeholder names, AI copywriting clichés — all banned by name.
DESIGN_VARIANCE (1–10): layout symmetry. MOTION_INTENSITY (1–10): animation richness. VISUAL_DENSITY (1–10): content packing. Threshold crossings activate or ban specific patterns automatically.
Minimalist (Notion/Linear), Brutalist (Swiss Industrial), Soft-Editorial (Ethereal Glass), GPT-Strict (Codex/GSAP). Mutually exclusive overlays on a shared base. Pick one, get a coherent design language.
Always-on behavior banning // ... placeholders, "for brevity" hedges, and structural elision. Defines a clean pause/resume protocol for token-limited outputs. Mount on every session that produces code.
The bundle didn't add new code paths or libraries — it shifted the model's defaults. Same prompt, same complexity, completely different output quality.
Each archetype overlays the shared base with its own palette, type stack, motion budget, and layout posture. Mutually exclusive — never compose two together.
Notion / Linear. Off-white backgrounds, Cabinet Grotesk + Newsreader pairing, Phosphor Bold icons. Gradients banned. Shadows < 0.05 opacity. Editorial restraint.
Swiss Industrial / CRT. 90° corners everywhere. border-radius banned. Monospace-heavy. Uppercase structural headers. Dark substrates with high-contrast data.
Ethereal Glass / Luxury. Doppelrand double-bezel architecture. Floating glass pill nav. backdrop-blur-2xl required. Ultra-light Phosphor/Remix icons. Button-in-button CTAs.
Codex / GSAP-heavy. XML design plan. Standalone — replaces the base. Hero headlines capped at 2–3 lines. grid-flow-dense mandatory. No cheap meta-labels.
gpt-taste standalone (replaces base).Base design-taste, 4 archetype overlays, redesign, output-discipline, design-context-generator
Component Designer (implements from specs) and Design Critic (audits against ban list)
Image-to-code, design-system-bootstrap, redesign-audit
Dials, ban list, Bento 2.0, creative arsenal, variations, output-discipline research
Prompt-level design rules. The base skill provides architecture, performance guardrails, and bans. Archetypes override the aesthetic dialect.
Component Designer implements designs faithfully from analysis docs. Design Critic audits with structured violation reports and file:line citations.
Multi-stage workflows with approval gates. Image-to-code: generate references → analyze → implement. Design-system-bootstrap: interview → archetype-pick → DESIGN.md.
Generate N section reference images via imagegen MCP → systematic visual analysis (typography, spacing, palette, anti-drift flags) → approval gate → faithful frontend implementation. Three stages with phase-gating. No shortcuts, no simplification.
Interview the user about their project → recommend an archetype with reasoning → approval gate → write a self-contained DESIGN.md with concrete hex codes, rem values, and component conventions. One-time setup, permanent reference.
Scan existing codebase against the full ban list → structured violation report with file:line citations → approval gate → targeted fixes. Upgrades existing projects in place without migrating frameworks.
The image-to-code recipe replaces a 1,228-line monolithic upstream skill with a 3-stage declarative YAML workflow — because sequential pipelines with phase-gates are recipes, not skills.
Global variables that drive branching logic throughout the skill body. Threshold crossings are switches, not suggestions.
Default: 8 · Range: 1–10
1–3: Symmetrical grids, centered layouts
4–7: Offset overlapping, varied aspect ratios
8–10: Masonry, fractional grid units, massive empty zones
> 4: Centered hero is banned
Default: 6 · Range: 1–10
1–3: CSS :hover and :active only
4–7: Custom cubic-bezier transitions, cascades
8–10: Scroll-triggered reveals, Framer Motion hooks
> 5: Magnetic micro-physics enabled
Default: 4 · Range: 1–10
1–3: Art gallery, huge section gaps
4–7: Normal web app spacing
8–10: Cockpit density, font-mono for numbers
> 7: Card containers are banned
Adapted from Leonxlnx/taste-skill — a curated SKILL.md collection of Awwwards-tier engineering rules. This bundle ports the original into Amplifier-native skills, behaviors, context files, and recipes without changing the substantive design guidance.
Primary source: michaeljabbour/amplifier-bundle-taste repository (cloned from GitHub, May 2026).
Leonxlnx/taste-skill (MIT-licensed). All design rules, parametric dials, named bans, archetype palettes, the Bento 2.0 paradigm, and the Doppelrand pattern are Leonxlnx's original work.context/ai-tells-design.md — the "50+" count reflects universal bans plus archetype-specific bans across all 6 design skill files.context/dials.md. Default values (8, 6, 4) and threshold rules are quoted directly.recipes/image-to-code.yaml (263 lines), recipes/design-system-bootstrap.yaml (99 lines), and recipes/redesign-audit.yaml.Methodology: Repository cloned via gh repo clone. All statistics derived from git log, git shortlog, find, and wc -l. No numbers are estimated — all are direct measurements from the source.