notque/vexjoy-agent / skills/frontend/frontend
frontend
UnverifiedFrontend: UI design, distinctive visual styles, HTML artifacts, Three.js 3D, TypeScript/React/Next.js apps, e-commerce, portfolios, React Native, web performance.
references/design-refs/llm-design-failure-modes.md
View on GitHub# LLM Design Failure Modes
Specific ways LLMs fail at design tasks. Each failure mode includes detection signals and concrete defenses. Load this reference for every design mode, after `skills/shared-patterns/ui-design-judgment.md`. Visual tells of generated UI (gradients, card mosaics, emoji icons, overcorrection) live in that doc's section 9 and are not repeated here.
> **Shared base**: Universal LLM failure modes (hallucination, overconfidence, generic output, arithmetic errors, stale knowledge) are documented in `skills/shared-patterns/llm-domain-failure-modes-base.md`. This file covers design-specific failures only.
---
## 1. Generic Copy Without Product Context
Core rules: `ui-design-judgment.md` section 8. Detection: the copy would work unchanged for a banking app and a game, or a competitor could ship the same words. Defense: ask for existing terminology and voice first; name the user's actual object ("No invoices yet", "Create invoice").
---
## 2. Inaccessible Color Suggestions
Core rules: `ui-design-judgment.md` section 5 (4.5:1 body, 3:1 large text and UI). Common misses: white text on mid-tone brand colors (500 shades and lighter), gray text lighter than #767676 on white, status colors used as text on white, placeholder text. Defense: state the ratio for every pair you recommend; safe anchors are gray-800 on white (14.72:1) and gray-600 on white (7.45:1).
---
## 3. Platform-Blind Critique
**What happens**: LLM applies desktop web conventions to mobile designs, or iOS patterns to Android, or web patterns to native apps.
**Detection signals**:
- Suggesting hover states for mobile interfaces
- Recommending hamburger menu for an iOS app (use tab bar instead)
- Applying web-style dropdowns to native mobile (use bottom sheets)
- Suggesting right-click context menus for touch interfaces
- Recommending text sizes below 16px for mobile body text
**Defense**:
- Identify the target platform before critiquing
- Apply platform-specific conventions:
| Platform | Key Conventions |
|----------|----------------|
| iOS | Tab bar (bottom), navigation bar (top), system sheets, swipe gestures, SF Symbols, 44pt touch targets |
| Android | Bottom navigation, FAB, Material Design components, back button, 48dp touch targets |
| Web Desktop | Sidebar nav, hover states, right-click menus, keyboard shortcuts, cursor states |
| Web Mobile | Bottom navigation, thumb-zone awareness, touch targets 44px+, no hover dependency |
- When platform is ambiguous, ask. "Is this for web or native mobile?"
---
## 4. Missing Edge Cases in Specs
**What happens**: LLM generates clean specs for the happy path — a well-formed form, a list with 5-10 items, a short username. Real products encounter extremes.
**Detection signals**:
- No empty state documented
- No loading state specified
- No error state designed
- No mention of content overflow
- No slow/offline connection handling
- "The user enters their name" but no spec for a 200-character name
**Defense**:
Use this checklist for every spec:
| Category | Edge Cases to Cover |
|----------|-------------------|
| Content length | Empty (0), one, minimum, typical, maximum, overflow |
| Data states | Loading (initial), loading (more), loading (action), success, error (inline), error (page), offline, stale |
| User input | Valid, invalid format, empty required, max length, special characters, paste, autofill |
| List/collection | 0 items, 1 item, typical count, maximum count, filtered to 0 |
| Permissions | Admin, editor, viewer, no access, expired access |
| Timing | Instant response, slow response (2-5s), timeout (30s+), rate-limited |
| i18n | Text 40% longer, RTL layout, locale-specific formatting |
---
## 5. Surface-Level Critique
Detection: adjectives with no element ("clean", "modern", "intuitive"); feedback that fits any design. Defense: the finding format in `ui-design-judgment.md` section 10: element, problem, effect, fix. "The 32px section spacing creates consistent rhythm" is feedback; "clean layout" is not. Say what works and which principle it applies.
---
## 6. Fabricated Research Findings
**What happens**: LLM invents plausible user quotes, statistics, persona details, or research findings that sound authoritative but have no source.
**Detection signals**:
- Specific statistics without source: "73% of users prefer..."
- User quotes that sound polished rather than natural
- Persona details too detailed to derive from the input data
- Confident claims about user behavior with no evidence trail
- "Research shows..." without citing which research
**Defense**:
- Every finding traces to specific user-provided data. If no data was provided, state the limitation.
- Use participant type attribution: "3 of 5 enterprise admins mentioned..." — not invented names
- Flag confidence levels: HIGH (multiple sources, behavioral data), MEDIUM (single source, self-reported), LOW (inferred, limited data)
- Mark assumptions explicitly: "Assumption: users complete this flow in a single session. Validate with session data."
- When asked about user behavior with no data, say so. "I don't have behavioral data on this. Here's what the design implies about expected behavior."
---
## 7. Token-Value Confusion
**What happens**: LLM specs use raw values (`#3B82F6`, `16px`, `0.5rem`) instead of design token references (`color-primary`, `spacing-4`, `space-2`).
**Detection signals**:
- Hex codes in the spec where token names should be
- Pixel values without corresponding token names
- Font specifications using raw values instead of type scale tokens
- Inconsistent values that should reference the same token
**Defense**:
- Ask for the design system's token naming convention before writing specs
- Every value in a spec should be a token reference: `color-primary` not `#3B82F6`
- When the token system is unknown, use conventional names and flag them: "Assuming token naming convention. Verify against your design system."
- Flag any raw value with: "Needs token mapping — verify [value] maps to [suggested token name]"
- If a value does not fit any existing token, call it out: "No existing token matches this value. Consider creating `spacing-section-header: 28px` or adjusting to nearest token `spacing-8: 32px`."
---
## 8. Framework Dumping
**What happens**: LLM lists all 10 Nielsen heuristics, all WCAG criteria, or all design principles as a checklist without applying them to the specific design.
**Detection signals**:
- All 10 heuristics listed regardless of relevance
- WCAG criteria listed without checking specific elements
- Design principles stated abstractly: "Contrast is important for hierarchy"
- No specific design elements referenced in the analysis
- Equal depth on all points instead of focusing on actual issues
**Defense**:
- Apply each criterion to the actual design. Skip criteria that do not apply.
- Focus depth on actual findings. A design with great keyboard support and poor contrast deserves a paragraph on contrast and a sentence on keyboard.
- Cite the specific element that passes or fails each criterion
- "WCAG 1.4.3 Contrast: The body text (#6B7280 on #FFFFFF) passes at 4.63:1." — not "Contrast should be at least 4.5:1 for normal text."
---
## 9. Ignoring Design System Constraints
**What happens**: LLM suggests components or patterns that conflict with the product's existing design system, introducing visual inconsistency.
**Detection signals**:
- Recommending a component variant that does not exist in the system
- Suggesting custom styling instead of using existing component props
- Proposing a new pattern when an existing one would work
- Mixing conventions from different design systems (Material + Apple HIG)
**Defense**:
- Ask what design system the product uses before making recommendations
- Recommend existing components and variants first. Propose new ones only when existing components cannot solve the problem.
- When suggesting a new pattern, show its relationship to existing patterns: "This extends the existing Card component with a new `featured` variant"
- Flag when a recommendation might conflict with existing patterns: "Check if your design system already has a notification component before building this custom one"
---
## 10. Treating Mobile as Smaller Desktop
**What happens**: LLM applies desktop interaction patterns to mobile, resulting in designs that ignore how mobile users actually hold and interact with their devices.
**Detection signals**:
- Small touch targets (below 44px)
- Hover-dependent interactions
- Dense information layouts without mobile-specific reorganization
- No consideration of thumb zone (reachability)
- Desktop-style dropdowns instead of mobile-native bottom sheets
**Defense**:
- Thumb zone awareness: primary actions in the bottom half of the screen (easy reach), secondary in the top half
- Touch targets: minimum 44x44 CSS pixels, with 8px spacing between targets
- Interaction patterns: bottom sheets over dropdowns, swipe gestures for common actions, pull-to-refresh
- Content density: reduce visible items, use progressive disclosure, larger type sizes (minimum 16px body)
- Text input: minimize typing, use pickers/selectors where possible, auto-capitalize, appropriate keyboard type (`type="email"`, `type="tel"`)
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| brainstormingobra/superpowersYou MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation). | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| brand-guidelinesanthropics/skillsApplies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| frontend-designanthropics/skillsGuidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| mcp-builderanthropics/skillsGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK). | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |