AgentSkillsCN

analyze-scene-changes

适用于分析生成故事中场景自评系统所作的改动，对比每页的初稿与最终输出

SKILL.md
--- frontmatter
name: analyze-scene-changes
description: Use when analyzing what the scene self-critique system changed in a generated story, comparing draft vs output for each page

Analyze Scene Changes

Overview

Analyzes what the scene description self-critique system changed between draft and output for each page of a generated story. Shows issues found, corrections applied, and field-by-field differences.

Usage

code
/analyze-scene-changes <job-id> [page-number]

Examples:

  • •/analyze-scene-changes job_1769373235483_1cazuipuh - Analyze all pages
  • •/analyze-scene-changes job_1769373235483_1cazuipuh 5 - Analyze only page 5

Running the Script

bash
node scripts/analyze-scene-changes.js <job-id> [page-number]

Requires DATABASE_URL in .env pointing to Railway PostgreSQL public proxy.

Output Format

For each page with changes:

  1. •Issues Found - Problems the self-critique identified
  2. •Corrections Applied - How each issue was fixed
  3. •Field Changes - Exact draft → output differences:
    • •imageSummary - Scene description text
    • •setting - location, camera, lighting, depthLayers
    • •characters - added/removed, position, pose, action, expression, holding
    • •objects - position changes

Pages with no issues show "Draft was correct".

Example Output

markdown
## Page 5

**Issues Found:** 3
1. Critical Rule #3 violation: Scene requires 3 children but only Lukas included
2. Story text fidelity: Manuel 'verschränkte die Arme' - missing character
3. Holding inventory conflict: Franziska holding with both hands

**Corrections Applied:**
1. Add Manuel to character array with crossed arms pose
2. Add Sophie to character array displaying her golden shield
3. Fix Franziska's holding: L:holding Roger's hand, R:resting on table

**Field Changes (Draft → Output):**

`characters` (added/removed):
  + Added: Manuel, Sophie

`Lukas`:
  - pose: "Standing, body turned three-quarter" → "Standing, weight on right leg"

Use Cases

  • •Debug scene quality - See what self-critique is catching/missing
  • •Prompt tuning - Identify common issues to add new checks for
  • •Validate fixes - Confirm prompt changes reduce specific error types
  • •Track improvements - Compare draft accuracy across story generations

Summary Statistics

At the end of a full analysis, shows:

  • •Pages with issues: X/Y (Z% draft accuracy)
  • •Total issues found: N