ui

<user_request> $ARGUMENTS </user_request>

<critical_override> THIS COMMAND OVERRIDES THE CLAUDE.md TASK ROUTING TABLE FOR AGENT SELECTION.

WHY: This command uses specialized visual analysis agents. dev:ui uses Gemini multimodal analysis for visual review — other agents cannot do this. dev:ui-engineer applies Anti-AI design patterns that dev:developer does not have.

AGENT RULES FOR THIS COMMAND:

• Design analysis/review → dev:ui agent (subagent_type: "dev:ui")
• Implementation of design improvements → dev:ui-engineer agent (subagent_type: "dev:ui-engineer")

DO NOT use dev:developer for design implementation (lacks Anti-AI design rules and visual verification). DO NOT use dev:architect for design review (cannot perform multimodal visual analysis). DO NOT use code-analysis:detective (READ-ONLY, cannot implement design changes). </critical_override>

  **You MUST:**
  - Use Task tool to delegate ALL design reviews to ui agent
  - Use Bash to check API keys and run Claudish
  - Use Read/Glob to find design references
  - Use Tasks to track workflow progress
  - Use AskUserQuestion for user input gates

  **You MUST NOT:**
  - Write or edit ANY files directly
  - Perform design reviews yourself
  - Write review files yourself (delegate to ui)
</orchestrator_role>

<todowrite_requirement>
  You MUST use Tasks to track orchestration workflow.
  Initialize Tasks AFTER Phase 0.5 (Intent Detection) determines INTENT_MODE.

  **If INTENT_MODE === "REVIEW_AND_IMPLEMENT":**
  1. Session Initialization
  2. Intent Detection
  3. Design Reference Input
  4. API Configuration
  5. Review Configuration
  6. Execute Analysis
  7. Present Results
  8. Implementation Approval
  9. Execute Implementation

  **If INTENT_MODE === "REVIEW_ONLY":**
  1. Session Initialization
  2. Intent Detection
  3. Design Reference Input
  4. API Configuration
  5. Review Configuration
  6. Execute Analysis
  7. Present Results
</todowrite_requirement>

<graceful_degradation>
  If neither GEMINI_API_KEY nor OPENROUTER_API_KEY available:
  - Explain the situation clearly
  - Provide setup instructions
  - Offer to describe the design verbally for basic feedback
  - Exit gracefully if user chooses not to configure
</graceful_degradation>

</critical_constraints>

<intent_detection> Analyze $ARGUMENTS to determine user intent.

**Note**: All trigger word matching should be case-insensitive (e.g., "IMPROVE", "Fix", "REVIEW" all work).