AI Tells and Forbidden Patterns Banned by taste-skill: A Complete Guide

The taste-skill repository explicitly prohibits visual clichés, generic copy, cluttered layouts, and UI noise—collectively known as "AI tells"—that signal LLM-generated design rather than human craftsmanship.

The open-source taste-skill collection defines strict design governance rules to distinguish human-crafted interfaces from generic AI output. These guidelines identify specific recognizable shortcuts and clichés that are banned across mobile, web, and code-generation workflows unless explicitly requested by the user.

Visual AI Tells Forbidden in taste-skill

According to skills/imagegen-frontend-mobile/SKILL.md, the Visual AI tells section prohibits specific aesthetic shortcuts that dominate LLM-generated mockups. The banned visual patterns include:

• Purple-blue fintech gradients applied indiscriminately across interfaces
• Random glass cards without functional purpose
• Ambient blobs lacking contextual relevance
• Fake neon premium looks that simulate depth without hierarchy
• Generic Dribbble-style floating widgets disconnected from user needs
• Oversized corner radii applied universally to all elements
• Over-rendered glossy surfaces that obscure information hierarchy

These elements create an immediately recognizable "AI aesthetic" that taste-skill seeks to eliminate from frontend generation workflows.

Layout AI Tells and Structural Anti-Patterns

The repository bans specific layout AI tells that result in cluttered, unfocused interfaces. As defined in the mobile skill documentation, forbidden structural patterns include:

• Fake chart-dashboard spam—grids of irrelevant data visualizations
• Repeated stat cards with no product justification
• Homepages resembling widget battlegrounds where 12+ elements compete for attention
• Cloned screens in a flow—identical layouts repeated across user journeys
• Giant empty cards containing weak or placeholder content
• Phone-shaped websites masquerading as native app screens

These patterns indicate template-driven generation rather than intentional information architecture.

Copy AI Tells and Generic Brand Slop

The Copy AI tells section of skills/imagegen-frontend-mobile/SKILL.md targets hollow marketing language and fake brand names that scream LLM generation. Prohibited phrases include:

• "Elevate your life"
• "Unlock your potential"
• "Next-gen finance"
• "Seamless control"
• "Smarter than ever"
• "Transform your day"

Additionally, taste-skill bans generic brand slop—placeholder names such as Acme, NovaCore, Flowbit, Quantix, and VeloPay that frequently appear in AI-generated mockups. The system requires either real brand identifiers or deliberately neutral placeholders, not these recurring AI tropes.

UI-Clutter Tells to Eliminate

Under UI-clutter tells, the repository prohibits decorative elements that add noise without function:

• Excessive pills, badges, and tiny labels crowding the interface
• Fake system markers and meaningless avatar rows
• Random chart inserts lacking data context
• Decorative toggles with no product meaning or state change

These elements create visual noise that detracts from usability while signaling automated generation.

Enforcement via ANTI-AI-TELLS and Clarity Checks

The taste-skill framework enforces these rules through dedicated ANTI-AI-TELLS sections and mandatory quality gates. Every skill concludes with a Clarity Check that explicitly asks: "Is it free of obvious AI tells?"

This verification appears across multiple skill files:

• skills/imagegen-frontend-web/SKILL.md applies the Clarity Check to web designs
• skills/image-to-code-skill/SKILL.md repeats the check for code-generation workflows
• skills/imagegen-frontend-mobile/SKILL.md embeds the check within mobile-specific guidelines

Unless the user explicitly overrides the rule—such as specifically requesting a "purple-blue fintech gradient"—the model must reject prompts that would produce these forbidden patterns.

Implementation Example

The following Python pseudo-code demonstrates how a taste-skill implementation might programmatically detect prohibited patterns before generation:

FORBIDDEN_VISUAL = [
    "purple-blue fintech gradient",
    "glass card",
    "ambient blob",
    "fake neon",
    "dribbble-style widget",
    "oversized corner radius",
    "glossy surface",
]

def contains_forbidden_visual(prompt: str) -> bool:
    lowered = prompt.lower()
    return any(term in lowered for term in FORBIDDEN_VISUAL)

if contains_forbidden_visual(user_prompt):
    # Respect the rule: either ask for clarification or refuse the specific element

    raise ValueError("Prompt contains a prohibited AI-tell visual pattern.")

Similar validation logic applies to layout, copy, and UI-clutter categories, with the final Clarity Check re-scanning generated output to ensure no prohibited patterns slipped through.

Summary

• taste-skill maintains explicit prohibitions against AI tells across four categories: visual aesthetics, layout structure, marketing copy, and UI clutter.
• Forbidden patterns include purple-blue gradients, "elevate your life" copy, fake dashboard widgets, and generic brand names like NovaCore.
• Enforcement occurs through ANTI-AI-TELLS sections in skills/imagegen-frontend-mobile/SKILL.md and mandatory Clarity Checks that verify outputs contain no obvious AI tells.
• The system rejects generation requests containing banned patterns unless the user explicitly overrides the restriction.
• Similar validation logic appears in web and image-to-code skill files, ensuring consistent quality across all output types.

Frequently Asked Questions

What constitutes an "AI tell" in design?

An AI tell is a recognizable shortcut, cliché, or pattern that signals a design was generated by an LLM rather than crafted by a human designer. These include overused visual effects like glass cards, generic marketing copy such as "unlock your potential," and structural patterns like fake dashboard spam that prioritize aesthetic noise over functional clarity.

Where are the forbidden patterns documented in taste-skill?

The primary definitions reside in skills/imagegen-frontend-mobile/SKILL.md, which contains dedicated sections for Visual AI tells, Layout AI tells, Copy AI tells, and UI-clutter tells. The enforcement mechanism appears in this file as well as in skills/imagegen-frontend-web/SKILL.md and skills/image-to-code-skill/SKILL.md through their respective Clarity Check sections.

Can users override the AI tell restrictions?

Yes, but only through explicit request. The taste-skill rules state that the model should reject requests producing forbidden elements unless the prompt explicitly overrides the rule—for example, asking specifically for a "purple-blue fintech gradient" or "Acme branding." Without such explicit instruction, the system enforces the prohibition.

How does the Clarity Check prevent AI tells from appearing in outputs?

The Clarity Check serves as a final quality gate that asks, "Is it free of obvious AI tells?" This verification step rescans the generated design or code before emission, catching any prohibited patterns that may have slipped through initial filtering. It represents the last line of defense in taste-skill's multi-layered enforcement strategy.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}