amplifier-bundle-taste

May 4, 2026 · View on GitHub

An Amplifier bundle that ports the anti-slop frontend design discipline from Leonxlnx/taste-skill into Amplifier-native skills, behaviors, context files, and a recipe.

Credit

This bundle is an adaptation of taste-skill by Leonxlnx.

All design rules, parametric dials, named bans, archetype palettes, the Bento 2.0 paradigm, the Doppelrand pattern, the image-to-code workflow, and the research-backed output discipline are Leonxlnx's original work. This bundle reformats those skills into Amplifier's bundle/skill/behavior/recipe conventions without modifying the substantive design guidance. Original work is MIT-licensed; please support the author at https://tasteskill.dev and via GitHub Sponsors.

If you use this bundle, star and credit the original repo first.

What it does

LLMs have hardcoded statistical priors toward mediocre frontend output: centered heroes, Inter font, purple-blue gradients, 3-column equal-width card grids, h-screen, // ... placeholder code, "Acme"/"Nexus" placeholder names, and "Elevate"/"Seamless" copy clichés. This bundle rewires those defaults at the prompt level with testable, named hard bans plus three parametric dials (DESIGN_VARIANCE / MOTION_INTENSITY / VISUAL_DENSITY) that act as global configuration variables.

It also provides an always-on output-discipline behavior that is domain-agnostic — bans // ... placeholders, "for brevity" hedges, and defines a clean pause/resume protocol for token-limited outputs. You should mount this behavior on every Amplifier session that produces code, not just frontend ones.

Add to Amplifier (from GitHub)

This is an Amplifier bundle. Install it directly from GitHub:

# Add the bundle (tracks main branch — auto-updates on `amplifier bundle refresh`)
amplifier bundle add git+https://github.com/michaeljabbour/amplifier-bundle-taste@main

# Or pin to a specific tag for stability
amplifier bundle add git+https://github.com/michaeljabbour/amplifier-bundle-taste@v0.1.0

# Verify it loaded
amplifier bundle list | grep taste

Use it as an active app bundle

Once added, you can either compose it into your own bundle, or run Amplifier directly with it as the active bundle:

# Run an interactive session with this bundle mounted
amplifier --bundle taste

# Or one-off prompts
amplifier --bundle taste run "Build a minimalist landing page hero in Next.js"

Compose into your own bundle

In your project's bundle.md:

includes:
  - bundle: taste:behaviors/output-discipline   # always-on completeness

Skills register with tool-skills automatically and are loadable on demand via load_skill. The model can invoke them based on description matching — pick an archetype at invocation time rather than hard-wiring it into the bundle.

Install for development (editable, from local clone)

git clone https://github.com/michaeljabbour/amplifier-bundle-taste.git ~/dev/amplifier-bundle-taste
amplifier bundle add file://$HOME/dev/amplifier-bundle-taste

Quickstart — your first session

After installing, mount the bundle and try one of these:

# Option 1: Generate a new UI in a specific aesthetic
amplifier --bundle taste run "Build a Notion-style dashboard hero section in Next.js with Tailwind"
# → loads design-taste + minimalist-taste, generates faithful code

# Option 2: Audit an existing UI codebase
amplifier --bundle taste run "Audit src/components/ for design taste violations"
# → delegates to taste:design-critic agent, produces structured report

# Option 3: Bootstrap a design system for a new project
amplifier --bundle taste run "I'm starting a new B2B SaaS app — help me lock in an aesthetic"
# → runs taste:recipes/design-system-bootstrap interview

If you're unsure which archetype to use, just describe your project. The bundle's instructions tell the model how to disambiguate based on keywords ("Notion-style" → minimalist, "Swiss design" → brutalist, "luxury" → soft-editorial, "GPT-strict" → gpt-taste).

Before / after — what the bundle actually changes

A concrete example. Same prompt, same model, with and without the bundle.

Prompt: "Build a Next.js hero section with a headline, subheadline, and CTA button."

Without taste bundle (typical LLM defaults)

// hero.tsx
export function Hero() {
  return (
    <section className="h-screen flex items-center justify-center bg-gradient-to-br from-purple-600 to-blue-600">
      <div className="max-w-4xl mx-auto text-center px-4">
        <h1 className="text-5xl font-bold text-white mb-4 font-[Inter]">
          Elevate Your Workflow
        </h1>
        <p className="text-xl text-white/80 mb-8">
          Seamlessly integrate next-gen AI into your business with our cutting-edge platform.
        </p>
        <button className="bg-white text-purple-600 px-8 py-3 rounded-full font-semibold hover:bg-gray-100">
          Get Started — It's Free
        </button>
      </div>
    </section>
  )
}

Counts of LLM tells: centered hero (1), h-screen iOS-bug (1), purple→blue gradient (1), Inter font (1), "Elevate" cliché (1), "Seamlessly" cliché (1), "next-gen" cliché (1), rounded-full on CTA (1) = 8 banned patterns in one component.

With `design-taste` + `minimalist-taste`

// hero.tsx
export function Hero() {
  return (
    <section className="min-h-[100dvh] grid grid-cols-12 gap-x-6 px-8 py-24 bg-[#FBFBFA]">
      <div className="col-span-12 md:col-span-7 md:col-start-2 self-end">
        <h1 className="font-[Cabinet_Grotesk] text-[clamp(3rem,7vw,6rem)] leading-[0.95] tracking-[-0.03em] text-[#111111] mb-6">
          Notes that think with you.
        </h1>
        <p className="font-[Newsreader] text-xl leading-[1.4] text-[#2F3437] max-w-prose mb-10">
          A document editor that reads alongside your work — pulling references, surfacing precedents, drafting passages on request. Ships next quarter.
        </p>
        <button className="px-5 py-2.5 bg-[#111111] text-[#FBFBFA] text-sm font-medium rounded-md hover:bg-[#2F3437] transition-colors">
          Join the private beta
        </button>
      </div>
    </section>
  )
}

What changed:

min-h-[100dvh] not h-screen (iOS Safari viewport bug)
Asymmetric 12-column grid, not centered
Cabinet Grotesk + Newsreader (editorial pairing) instead of Inter
Off-black #111111 instead of #000000, off-white #FBFBFA background
Concrete copy ("Notes that think with you", "ships next quarter") instead of "Elevate Your Workflow"
rounded-md not rounded-full on the CTA
Single-color treatment, no gradient

The bundle didn't add new code paths or libraries — it shifted the model's defaults. Same prompt, same complexity, completely different output quality.

Migrating from upstream `taste-skill`

If you're already using Leonxlnx/taste-skill directly via npx skills add and want to switch to this Amplifier-native version, the mappings are:

Skill names — upstream taste-skill → bundle design-taste; gpt-tasteskill → gpt-taste; minimalist-skill → minimalist-taste; brutalist-skill → brutalist-taste; soft-skill → soft-editorial-taste; redesign-skill → redesign-existing-projects; output-skill → output-discipline (now mounted as a behavior, auto-loads on session start).
image-to-code-skill is now a recipe, not a skill (recipes/image-to-code.yaml). The 1228-line upstream skill is split into 3 stages (generate references → analyze → implement) with an approval gate. Run via amplifier recipe run taste:recipes/image-to-code.
stitch-skill is replaced with the format-neutral design-context-generator skill — generates a generic DESIGN.md consumable by any AI session, not Stitch's proprietary semantic-design-language format.
Parametric dials (DESIGN_VARIANCE, MOTION_INTENSITY, VISUAL_DENSITY) are unchanged in semantics. They live in context/dials.md and are referenced by skills via @-mention.
The Python random.choice() theater in upstream gpt-tasteskill is replaced by the explicit anti-default rule in context/design-variations.md — "Choose a combination not used in your immediately prior output. Do not default to the first option." The model doesn't actually execute Python.

For substantive design-rule changes (a ban is wrong, a dial threshold is off), file at the upstream repo — we re-port after the original author accepts. For Amplifier-specific bugs (recipe wiring, frontmatter, behavior YAML), file here.

What's inside

Path	What it provides
`skills/design-taste.md`	Base premium frontend skill — 3-dial parametric config + Bento 2.0 paradigm + Creative Arsenal
`skills/minimalist-taste.md`	Notion / Linear editorial overlay
`skills/brutalist-taste.md`	Swiss Industrial / CRT Tactical overlay
`skills/soft-editorial-taste.md`	Ethereal Glass / Editorial Luxury overlay (with Doppelrand)
`skills/gpt-taste.md`	GPT/Codex-optimized variant — XML design plan, GSAP-heavy
`skills/redesign-existing-projects.md`	Audit-then-fix for upgrading existing projects
`skills/design-context-generator.md`	Generates a generic `DESIGN.md` reference doc (concept extracted from stitch-skill, format-neutralized)
`skills/output-discipline.md`	Auto-loaded — bans `// ...` placeholders, "for brevity" hedges, defines pause/resume protocol
`behaviors/output-discipline.yaml`	Behavior wrapper that auto-loads the output-discipline skill on session start
`context/dials.md`	The 3 parametric dials and their threshold logic
`context/ai-tells-design.md`	Merged hard-ban list across all design archetypes
`context/bento-2.md`	Bento 2.0 reference — 5 named card archetypes + Framer Motion physics
`context/design-variations.md`	Combinatorial variation menus
`context/output-discipline-research.md`	Distilled research backing for output-discipline
`recipes/image-to-code.yaml`	3-stage image-first workflow: generate references → analyze → implement
`recipes/redesign-audit.yaml`	4-stage audit-then-fix workflow with approval gate (scan → diagnose → approval → fix). Operationalizes the redesign-existing-projects skill.
`recipes/design-system-bootstrap.yaml`	3-stage onboarding: interview → archetype-pick (approval gate) → DESIGN.md authoring. Use once per new project.

Recommended usage

Use case	What to load
Universal output discipline	`behaviors/output-discipline.yaml` — mount on every session
Generic premium frontend	`skills/design-taste.md` (auto-invokes on UI work)
Notion / Linear aesthetic	`design-taste.md` + `minimalist-taste.md`
Swiss Industrial / Terminal aesthetic	`design-taste.md` + `brutalist-taste.md`
Editorial luxury aesthetic	`design-taste.md` + `soft-editorial-taste.md`
GPT/Codex strict mode	`gpt-taste.md` (replaces base — its own structure)
Upgrading an existing project	`redesign-existing-projects.md`
Image-first design pipeline	`recipes/image-to-code.yaml`

Prerequisites for `image-to-code` recipe

This recipe has two runtime dependencies that are not bundled:

Imagegen MCP server — required for generating reference images in stage 1. Install: see https://github.com/microsoft/amplifier-mcp-imagegen (or set archetype and provide your own reference images).
(Optional) amplifier-bundle-design-intelligence — if installed, the recipe will use its component-designer agent. If not, the bundle's own taste:component-designer agent (defined in agents/) handles implementation.

The recipe degrades gracefully if these are missing — but you'll get clearer errors with them installed.

Composition rule (the bricks-and-studs pattern)

The four aesthetic archetypes (minimalist, brutalist, soft-editorial, gpt) are mutually exclusive — never compose two together (minimalist bans gradients; soft-editorial requires them). Compose the base + exactly one archetype, or use gpt-taste standalone (it replaces the base).

When two skills are simultaneously loaded into a session's context, the agent resolves conflicting rules via natural-language instruction-following — more specific archetype constraints take precedence over the base's generic rules. This is NOT the same as Amplifier's bundle-level composition merge (which entirely replaces parent instructions). Skills accumulate; bundles replace.

The output-discipline behavior is always compatible with any combination — mount it globally.

What changed from the source

This adaptation makes minimal substantive changes — the design rules are intact. What was reshaped:

Frontmatter rewritten to follow the Agent Skills specification and Amplifier extensions (description is now a "Use when…" trigger statement, not a workflow summary; added model_role for routing).
Output discipline split into two layers — a skill (skills/output-discipline.md) plus a behavior wrapper (behaviors/output-discipline.yaml) so it auto-loads on session start.
Image-to-code skill converted to a recipe — the 3-phase pipeline (generate references → analyze → implement) is inherently sequential with phase-gates, so a recipe (declarative YAML workflow) is the right form, not a skill.
Stitch DESIGN.md skill replaced with a format-neutral design-context-generator — the original generated Stitch's proprietary semantic-design-language format; this version generates a generic DESIGN.md consumable by any AI session.
Variation engine simplified — replaced the Python random.choice() theater (the model doesn't actually execute Python) with the explicit rule "choose a combination not used in your immediately prior output."
Reference material extracted — the Bento 2.0 specs, 3-dial logic, merged ban lists, and variation menus moved into context/ files that skills @-mention rather than embed inline. Keeps each skill scannable.

The taste-skill base, all 4 archetypes, the redesign skill, the output skill, and the entire research corpus from research/laziness/ are otherwise preserved.

License & attribution

MIT, matching the upstream license. See LICENSE.

Original work © Leonxlnx at https://github.com/Leonxlnx/taste-skill. Star the original repo — that's where the substantive work lives.

Issues, bugs, contributions

Adaptation issues (frontmatter, bundle structure, recipe wiring, Amplifier integration): file here at https://github.com/michaeljabbour/amplifier-bundle-taste/issues
Substantive design guidance (a ban is wrong, a dial logic is off, an archetype rule needs revision): file at the upstream repo so the original author can fix it once and we'll re-port.