Agent Skills: Auto-Validator

Auto-Validator

Purpose

Automate pre-manifest visual validation for generated KR Solidarity assets. Input an asset image and receive a deterministic score, violations list, and correction prompt.

When to Use

• After Gemini/DALL-E asset generation attempts.
• Before promoting assets into manifest/registry workflows.
• During rapid iteration loops where manual review is too slow.

Decision Gate

• PACKAGE when score >= 90 and no critical violations.
• REGENERATE when score < 90 or critical violations exist.

Validation Scorecard (100)

1. Semantic Token Alignment (25)

• Palette intent aligns to canonical KR Solidarity tokens.
• Reject hardcoded off-canon color behavior in UI-oriented outputs.

2. M3 Expressive Typography (20)

• Expressive hierarchy and contrast.
• Approved families only: Work Sans, Fraunces, JetBrains Mono, Libre Bodoni, Caveat, Nabla (restricted decorative use).
• Banned defaults: Inter, Roboto, Arial, Sora, Plus Jakarta Sans.

3. Color Palette Compliance (20)

• Canonical references include:
  • --sys-color-asphaltBlack-base
  • --sys-color-paperWhite-base
  • --sys-color-solidarityRed-base
  • --sys-color-inkGold-base
  • --sys-color-stencilYellow-base
  • --sys-color-kr-activistSmokeGreen-base
  • --sys-color-worker-ash-base
  • --sys-color-concreteGrey-base
• Reject purple-generic drift and white-first expressive surfaces.

4. Asymmetric Composition (15)

• Shape language aligns to expressive asymmetry.
• Reject rigid symmetry and border-radius: 50% dependency for core forms.

5. Accessibility (10)

• WCAG 2.2 AA contrast targets (4.5:1 text, 3:1 UI elements).
• Focus visibility and touch target expectations.

6. Identity & Social Framing (10)

• Urban solidarity framing and culturally respectful representation.
• Reject bureaucratic/corporate sterility and appropriation patterns.

Workflow

Input:

• image path
• asset id
• target score (default 90)

Process:

1. Extract palette and composition signals.
2. Infer typography and hierarchy quality.
3. Evaluate contrast and accessibility signals.
4. Score each dimension.
5. Generate violations and correction prompt.

Output:

{
  "asset_id": "KR-SOLID-HERO-001",
  "overall_score": 87,
  "decision": "REGENERATE|PACKAGE",
  "dimensions": {
    "semantic_token_alignment": {"score": 20, "violations": []},
    "m3_expressive_typography": {"score": 18, "violations": []},
    "color_palette_compliance": {"score": 19, "violations": []},
    "asymmetric_composition": {"score": 14, "violations": []},
    "accessibility": {"score": 8, "violations": []},
    "identity_social_framing": {"score": 8, "violations": []}
  },
  "critical_violations": [],
  "correction_prompt": "string",
  "iteration_priority": "high|medium|low"
}

Integration

• With prompt-composer: apply correction prompt for next attempt.
• With asset-packager: pass through only when decision=PACKAGE.
• With vision-scorer-mcp: final post-manifest validation after integration.

Troubleshooting

Low-resolution image

• Return needs_more_evidence note and avoid critical typography judgments.

OCR/font ambiguity

• Downgrade confidence and avoid hard fails unless corroborated by multiple signals.

Contrast calculation uncertainty

• Re-run with cleaner screenshot and explicit foreground/background surfaces.

False positives on palette drift

• Compare against canonical token intent in frontend/src/design/styles/design-tokens.css.

Missing model or vision service

• Return deterministic failure object with service_unavailable and no scoring.

Important Limits

• Vision-based checks are inference-driven and should be human-reviewed for final approvals.
• This skill does not validate manifest integrity or registry linkage.

Critical Difference vs vision-scorer-mcp

• auto-validator: pre-manifest generation loop optimizer.
• vision-scorer-mcp: post-manifest deterministic system gate.

Last Updated: 2026-03-08 | Version: 1.1.0