AgentSkillsCN

image-analysis

GLM-4.6v 图像分析子代理,专用于游戏截图与精灵图质量评估。

SKILL.md
--- frontmatter
name: image-analysis
description: GLM-4.6v image analysis subagent for game screenshot and sprite quality assessment

Image Analysis Skill

Purpose: Delegate image analysis tasks to GLM-4.6v subagent for visual quality assessment of game screenshots and sprites.

When to Use

Invoke this skill when:

  • New sprites are generated (via glm-image-gen)
  • Screenshots captured during playtesting
  • Visual quality validation needed before commits
  • Checking for blocky backgrounds or transparency issues
  • Comparing assets against Harvest Moon SNES style

Direct vs Delegated Analysis

Claude Opus can analyze images directly via the Read tool. Choose the right approach:

ScenarioApproachWhy
1-3 imagesClaude DirectNo overhead, faster
Quick quality checkClaude DirectImmediate feedback
Batch analysis (10+)MiniMax ParallelToken efficiency
Extensive visual QAMiniMax Parallel50x cheaper

Claude Direct Pattern

code
Read image file → Analyze in-context → Continue workflow

MiniMax/GLM Delegation Pattern (for batches)

code
Spawn parallel agents → Each analyzes subset → Aggregate results

Default: Use Claude Direct for small tasks, delegate batches to save tokens.

Quick Start

Analyze Screenshot

bash
mcp__4_5v_mcp__analyze_image \
  --imageSource "temp/screenshots/[filename].png" \
  --prompt "Analyze this game screenshot for visual quality. Check for: blocky backgrounds, tiling seams, sprite transparency issues, pixel art scaling problems, overall Harvest Moon SNES readability. Provide specific, actionable feedback."

Analyze Sprite Asset

bash
mcp__4_5v_mcp__analyze_image \
  --imageSource "assets/sprites/placeholders/[filename].png" \
  --prompt "Analyze this pixel art sprite for Harvest Moon SNES style quality. Check: transparency (clean edges), silhouette (strong/readable), proportions (compact not stretched), color palette (dusk-friendly), pixel detail (minimal noise). Reference: assets/sprites/PLACEHOLDER_ASSET_SPEC.txt"

Analyze NPC Sprite

bash
mcp__4_5v_mcp__analyze_image \
  --imageSource "assets/sprites/placeholders/npc_[name].png" \
  --prompt "Analyze this 64x64 NPC sprite. Target proportions: ~48h x 32w centered. Check: strong silhouette, Mediterranean theme alignment, clean transparency, no elongation. Harvest Moon SNES style reference."

Standard Prompts

Full Quality Assessment

code
Analyze this image for Circe's Garden game quality standards:

Visual Style:
- Harvest Moon SNES readability (clean pixel art, strong silhouettes)
- Stardew Valley influence (but more mythic/moody)
- Mediterranean ancient Greek island theme
- Dusk-friendly palette (muted, not oversaturated)
- 32px tiles, 2x camera zoom

Check For:
1. Background issues: blocky patterns, tiling seams, color banding
2. Transparency: clean edges (no halos), proper alpha channel
3. Pixel art: minimal noise, clear details, proper scaling
4. Proportions: compact bodies, not stretched/elongated
5. Colors: natural earth tones, muted jewel accents
6. Lighting: soft ambient, no harsh specular
7. Silhouette: strong, recognizable at small size

Provide specific feedback: what works, what needs fixing, file references.

Transparency Check (Focused)

code
Check this sprite for transparency issues:
- Halo effect (light/dark ring around edges)
- Color bleeding (background showing through)
- Harsh edges (no smooth curves on diagonals)
- Partial transparency (should be opaque or transparent, not in-between)

Report specific pixel locations and severity.

Style Comparison

code
Compare this asset to Harvest Moon SNES style:
- Clean pixel art with minimal noise
- Strong silhouette (recognizable at 32x32)
- Soft, natural color palette
- Proper alpha blending
- No anti-aliasing on edges
- Cozy, approachable aesthetic

Report what matches style and what deviates.

Output Format

Expected response should include:

markdown
## Analysis Summary
[One-line assessment: PASS / NEEDS WORK / FAIL]

### Issues Found
**Critical (P0):**
- [Specific issue with location]

**Important (P1):**
- [Issue description]

**Minor (P2):**
- [Minor detail]

### What Works
- [Positive aspects to preserve]

### Recommendations
1. [Actionable fix 1]
2. [Actionable fix 2]

Common Issues

IssueDetectionFix
Blocky backgroundVisible pixel grid, tiling patternsRegenerate with better seam handling
Halo effectLight/dark ring around spriteFix alpha channel, clean edges
Color bleedingBackground color in transparent pixelsRe-export with proper transparency
Stretched proportionsElongated body/limbsRegenerate with compact proportions
Oversaturated colorsToo bright/vibrantAdjust palette to dusk-friendly tones
Poor silhouetteNot recognizable at small sizeStrengthen outline, simplify details

File Locations

Input:

  • Screenshots: temp/screenshots/
  • Sprites: assets/sprites/
  • Placeholders: assets/sprites/placeholders/

Reference:

  • Asset specs: assets/sprites/PLACEHOLDER_ASSET_SPEC.txt
  • Style guide: assets/sprites/PLACEHOLDER_README.md

Integration

With glm-image-gen

bash
# After generating new sprite
./generate-image.sh '{"subject":"..."}' 1024x1024 standard
./process-image.sh "$TEMP" "32x32" "assets/sprites/output.png"

# Then analyze quality
mcp__4_5v_mcp__analyze_image --imageSource "assets/sprites/output.png" --prompt "[standard prompt]"

With HPV (Playtesting)

bash
# During screenshot review
ls temp/screenshots/
mcp__4_5v_mcp__analyze_image --imageSource "temp/screenshots/world_map.png" --prompt "[screenshot prompt]"

Pre-commit Check

bash
# Before committing asset changes
git status
mcp__4_5v_mcp__analyze_image --imageSource "assets/sprites/new_asset.png" --prompt "[validation prompt]"

Subagent Role

Full subagent configuration: .claude/roles/image-analysis-subagent.md

Model: GLM-4.6v (vision-capable) Tool: mcp__4_5v_mcp__analyze_image Purpose: Specialized visual quality assessment

Success Criteria

Analysis is effective when:

  • Issues match human findings (80%+ accuracy)
  • False positives < 20%
  • Actionable feedback provided
  • Response time < 30 seconds
  • Consistent report format

Related:

  • .claude/skills/glm-image-gen/SKILL.md (sprite generation)
  • assets/sprites/PLACEHOLDER_ASSET_SPEC.txt (asset specifications)
  • .claude/roles/image-analysis-subagent.md (full subagent docs)

[Kimi Code CLI - 2026-01-26]