Agent Skills: vs-design-diverge

Create non-generic, high-entropy frontend interfaces using Verbalized Sampling (VS) to break mode collapse and maximize design creativity.

UncategorizedID: xavierchoi/xavierchoi_skills/vs-design-diverge

Install this agent skill to your local

pnpm dlx add-skill https://github.com/xavierchoi/xavierchoi_skills/tree/HEAD/plugins/vs-design-diverge/skills/vs-design-diverge

Skill Files

Browse the full folder contents for vs-design-diverge.

Download Skill

Loading file tree…

plugins/vs-design-diverge/skills/vs-design-diverge/SKILL.md

Skill Metadata

Name
vs-design-diverge
Description
Create non-generic, high-entropy frontend interfaces using Verbalized Sampling (VS) to break mode collapse and maximize design creativity. Use when the user wants to explore multiple distinct design directions, compare aesthetic options before committing, or explicitly asks to avoid generic AI-looking design (e.g. "make it less generic", "show me bolder options", "explore different design directions"). For a single polished implementation without direction exploration, use frontend-for-opus-4-5 instead.

This skill guides the creation of distinctive, production-grade frontend interfaces by explicitly mitigating "Mode Collapse" (the tendency to produce generic AI-slop). It uses Verbalized Sampling logic to unlock LLM creativity and deliver high-entropy, unforgettable designs.

Phase 0: Context Discovery (do this first)

Before any design analysis, gather deep context from the user — every later phase depends on it. Use the AskUserQuestion tool to probe the following dimensions:

Discovery Dimensions

  1. Emotional Tone: What feeling or atmosphere should the design evoke? (e.g., "trustworthy", "edgy", "playful", "luxurious")
  2. Target Audience: Who will see this? What are their expectations and technical sophistication?
  3. Reference Points / Anti-References: What does the user want to emulate? What do they explicitly want to avoid?
  4. Business Context: What problem does this UI solve? What's the usage scenario?

Context Signals

  • Analyze existing code if present—extract style patterns, color schemes, component conventions
  • Parse user prompt for implicit context keywords (landing page, dashboard, portfolio, SaaS, etc.)
  • Ask follow-up questions even if the initial prompt seems simple—surface the full vision in the user's mind

Do not proceed to Phase 1 until you have sufficient context. A vague prompt calls for more questions, not assumptions — guessing here produces exactly the generic output this skill exists to avoid.


Phase 1: Identify the Mode (The Generic Baseline)

  • Verbalize the most predictable, high-probability (P ≈ 0.95) design for this request.
  • Identify "AI-slop" markers based on the context. Common examples include:
    • Inter/Roboto/System fonts
    • Rounded blue/purple buttons
    • Standard F-pattern layouts
    • White backgrounds with gradient accents
    • Generic hero sections with stock-style imagery
  • Do not choose this baseline — it is the exact output this skill exists to avoid.

Phase 2: Sample the Long-Tail (Probability Mapping)

Generate three distinct aesthetic directions and assign a "Typicality Score" (T-Score) from 0 to 1.0 (where 1.0 is most generic):

  • Direction A (T ≈ 0.7): Modern/Clean but safe.
  • Direction B (T ≈ 0.4): Distinctive/Characterful (Specific niche style).
  • Direction C (T < 0.2): Experimental/Bold (High-entropy, unexpected).

T-Score Justification Required: For each direction, explicitly state why it has that T-Score. What makes it more or less typical? Reference specific design choices.


Phase 3: Commit to Low-Typicality

Select the direction with the lowest T-Score that still meets:

  1. Functional requirements from Phase 0
  2. All Quality Guardrails (see below)

Commit to this bold aesthetic with intentionality. The choice must be deliberate, not accidental — a half-committed direction reads as noise rather than character.


Quality Guardrails (always required)

These principles must hold regardless of how experimental the design becomes — they are what separate a distinctive design from a broken one. If a Low-T direction violates any of these, increase T until compliance is achieved.

| Guardrail | Description | |-----------|-------------| | Visual Hierarchy | Clear priority ordering—the user's eye must know where to go first, second, third | | Contrast & Legibility | Text must be readable against its background (WCAG AA minimum) | | Internal Consistency | The design must follow its own logic—a coherent system, not random choices | | Functional Clarity | Interactive elements must be recognizable as such; affordances must be clear |


Anti-Patterns (explicit failures)

If your design exhibits these patterns, it has missed the skill's intent:

1. Accidental Design

  • Making things "different" without intentionality
  • Randomness masquerading as creativity
  • Unable to articulate why a choice was made
  • Test: If asked "why this color/font/layout?", you must have a coherent answer

2. Frankenstein Style

  • Mixing incompatible aesthetics without synthesis
  • No unifying design language
  • Elements that feel like they belong to different projects
  • Test: Could you describe this design's "personality" in one sentence?

Frontend Aesthetics Guidelines (VS-Enhanced)

Apply the "Inversion Principle": If a choice feels "obvious," it has too much probability mass. Consider the lower-probability, higher-impact alternative—but only if it serves the context.

Typography (Context-Aware Selection)

  • AI-slop examples: Inter, Roboto, Arial, System fonts, Space Grotesk in default usage
  • Low-T approach: High-character display faces paired with refined, unexpected body typefaces. Variable fonts, unusual weights.
  • Keep in mind that the "right" font depends on context. A brutalist portfolio may use industrial sans-serifs; a luxury brand may use refined serifs.

Color & Theme (Chromatic Divergence)

  • Avoid even distributions. Commit to a cohesive but "dissonant-yet-beautiful" palette.
  • Use CSS variables for systematic theming.
  • Prefer sophisticated textures, noise, and lighting over flat fills—when appropriate to context.

Spatial Composition (Anti-Grid Logic)

  • If a standard grid is P=0.9, consider P=0.1 layouts: Asymmetry, overlapping elements, diagonal flows, or editorial-style whitespace.
  • Caveat: Data-heavy interfaces (dashboards, tables) may require conventional grids for usability.

Motion (High-Impact Entropy)

  • Focus on "Micro-delights": staggered reveals, scroll-bound transformations, custom eased transitions.
  • Avoid motion for motion's sake—every animation should have purpose.

Strategic Frameworks (Context-Dependent)

When the user's goal involves persuasion or conversion (landing pages, marketing sites, product showcases), apply the AIDA Framework to structure the visual narrative:

AIDA Framework

| Stage | Goal | Design Application | |-------|------|-------------------| | Attention | Stop the scroll, create immediate visual impact | Hero section with bold typography, unexpected imagery, or striking color contrast | | Interest | Build curiosity, encourage exploration | Reveal information progressively, use visual storytelling, highlight unique value | | Desire | Create emotional connection, make them want it | Social proof, benefits visualization, aspirational imagery, micro-interactions | | Action | Drive conversion with clear next steps | High-contrast CTAs, reduced friction, urgency cues (when appropriate) |

When to Apply AIDA:

  • Landing pages and marketing sites
  • Product launch pages
  • Portfolio pieces meant to attract clients
  • Any UI where the primary goal is persuasion/conversion

When not to Apply AIDA:

  • Dashboards and data-heavy interfaces (prioritize usability)
  • Documentation or content-focused sites (prioritize readability)
  • Internal tools (prioritize efficiency)

VS + AIDA Integration: Apply Low-T aesthetics to each stage of AIDA. A generic hero section fails at Attention; a predictable CTA fails at Action. The AIDA structure provides the "what," VS provides the "how."


Implementation Standards

  • Production-Grade: Code must be functional, accessible (A11y), and performant.
  • Complexity-Typicality Balance: As the design becomes more "Low-Probability," implementation complexity must increase proportionally to maintain quality.
  • No Refusal on Complexity: Claude is capable of extraordinary creative work. Do not simplify the vision for brevity.

Execution Process Summary

[Phase 0] Context Discovery
    ↓ (AskUserQuestion tool - gather all dimensions)
[Phase 1] Identify the Mode
    ↓ (verbalize the generic baseline)
[Phase 2] Sample the Long-Tail
    ↓ (three directions with justified T-Scores)
[Phase 3] Commit to Low-Typicality
    ↓ (select lowest T that passes Guardrails)
[Phase 4] Execute Code
    ↓ (implement with meticulous attention)
[Phase 5] Surprise Check
    ↓ (does this look like typical AI output? if yes, refactor)

Final Validation

Before delivering, ask yourself:

  1. Intentionality: Can I justify every major design decision?
  2. Consistency: Does the design follow its own internal logic?
  3. Guardrails: Are hierarchy, legibility, consistency, and clarity preserved?
  4. Surprise: Would this stand out in a lineup of AI-generated designs?

The goal is to maximize "Surprise Score" while maintaining "Production Quality." Break the mold—deliberately.