visual-verdict

<Use_When>

• The task includes visual fidelity requirements (layout, spacing, typography, component styles)
• You have one generated screenshot and at least one reference image
• Before continuing to revise, you need deterministic pass/fail guidance </Use_When>

<Output_Contract> Return only JSON, and strictly conform to the following structure:

{
  "score": 0,
  "verdict": "revise",
  "category_match": false,
  "differences": ["..."],
  "suggestions": ["..."],
  "reasoning": "short explanation"
}

Rules:

• score: integer 0-100
• verdict: short status (pass、revise or fail）
• category_match: when the generated screenshot matches the expected UI category / style, it is true
• differences[]: specific visual differences (layout, spacing, typography, color, hierarchy)
• suggestions[]: actionable next-step revisions corresponding to the differences
• reasoning: a 1-2 sentence summary </Output_Contract>

<Threshold_And_Loop>

• The target pass threshold is 90+。
• If score < 90, continue revising and rerun before any further visual review /oh-my-claudecode:visual-verdict。
• Do not consider the visual task complete until the next round of screenshots reaches the threshold. </Threshold_And_Loop>

<Debug_Visualization> When difference diagnosis is difficult:

1. Use $visual-verdict as the authoritative verdict.
2. Use pixel-level diff tools (pixel diff / pixelmatch overlay) as anauxiliary debugging aidto locate hotspots.
3. Convert pixel diff hotspots into specific differences[] and suggestions[] updates. </Debug_Visualization>

Task: {{ARGUMENTS}}