creative-review

Meta-review skill for validating generated creative assets (videos, images, designs) against a reference style brief. Extracts frames, compares against style criteria, scores adherence, and produces a

SKILL.md

GitHub

Creative Review — Style Adherence & Feedback Loop

Validate generated creative assets against a reference style brief. Score adherence, produce actionable feedback, and feed improvements back into the generation pipeline.

Compounding Skills

Skill Role
/agent-browser Watch generated videos, take screenshots, visual comparison
/launch-video Style brief and quality checklist to validate against
/content-creation Reference extraction, visual analysis with Gemini
/blog-post Distribution quality gates

Review Pipeline

REFERENCE → EXTRACT BRIEF → GENERATE ASSET → REVIEW → SCORE → FEEDBACK → ITERATE

Phase 1: Reference Extraction

When given a reference video or image:

  1. Download the reference (yt-dlp for URLs, direct path for local files)
  2. Extract frames at 1fps: ffmpeg -i ref.mp4 -vf "fps=1" frames/frame_%03d.png
  3. Analyze each frame for:
    • Color palette (dominant colors, background, accent)
    • Typography (font style, size, placement, weight)
    • Composition (layout, perspective, depth)
    • Motion style (cuts, transitions, pacing)
    • Material treatment (glass, shadow, glow, reflection)
  4. Produce a style brief — structured document capturing all visual attributes

Style Brief Format

## Style Brief: {Reference Name}

### Palette
- Background: {hex}
- Panel fill: {rgba}
- Text primary: {hex}
- Accent: {hex}

### Typography
- Font: {family}
- Weight: {bold/regular}
- Placement: {center/left/overlay}
- Max words per card: {N}

### Composition
- Panel perspective: {degrees}
- Panel material: {glass/solid/wireframe}
- Depth technique: {shadow/glow/parallax}
- Background treatment: {void/gradient/particles}

### Motion
- Entrance style: {spring/fade/slide}
- Scene duration: {N-N seconds}
- Transition type: {cut/crossfade/spring}
- Pacing: {fast/confident/slow}

### Audio
- Style: {ambient/narration/music}
- Sync points: {beat drops/scene changes}

Phase 2: Asset Review

When given a generated asset to review:

Video Review Process

  1. Extract frames at 1fps from the generated video
  2. Compare frame-by-frame against the style brief criteria
  3. Score each dimension (0-10):
Dimension What to Check Weight
Color adherence Does the palette match the brief? Dark void bg? Accent colors? 15%
Typography Font style, size, placement, word count per card 10%
3D perspective Are panels tilted? Proper perspective depth? 15%
Glass material Border glow, shadow, rounded corners, semi-transparency 15%
Motion quality Spring animations? Organic movement? No CSS transitions? 15%
Pacing Scene duration 3-5s? No rapid cuts? Confident rhythm? 10%
Particle/depth Background particles? Depth layering? 5%
Hook effectiveness Does the first 3s grab attention? 10%
Overall polish Does it feel professional? Would you stop scrolling? 5%
  1. Overall score: Weighted average (0-100)

Scoring Bands

Score Rating Action
90-100 Excellent Ship it
75-89 Good Minor tweaks, optional iteration
50-74 Needs work Specific feedback, iterate before shipping
0-49 Redo Major issues, regenerate with updated prompts

Phase 3: Feedback Generation

For each dimension scoring below 8/10, generate specific, actionable feedback:

Feedback Format

## Creative Review: {Asset Name}

**Overall Score**: {N}/100 ({Rating})
**Reference**: {Brief Name}

### Scores
| Dimension | Score | Notes |
|-----------|-------|-------|
| Color adherence | {N}/10 | {specific note} |
| Typography | {N}/10 | {specific note} |
| ... | ... | ... |

### Must Fix (score < 6)
1. {Specific issue} → {Specific fix with code/prompt change}
2. ...

### Should Fix (score 6-7)
1. {Issue} → {Fix}

### Nice to Have (score 8-9)
1. {Polish suggestion}

### What Worked Well
- {Positive observation}
- {Pattern to keep}

Phase 4: Self-Evolution

After Each Review Cycle

  1. If the fix worked → Add the technique to the style brief as a confirmed pattern
  2. If the fix didn't work → Document why and what was tried, update guidance
  3. If a new technique emerged → Capture it and add to the relevant skill's references

Style Brief Evolution

The style brief is a living document. After each review cycle:

Initial brief (from reference extraction)
  → Review #1 feedback applied
    → Review #2: new patterns discovered
      → Brief updated with confirmed patterns
        → Next generation starts from improved brief

Cross-Skill Feedback

When review findings affect other skills, propagate:

Finding Update Target
"Veo prompts produce static shots" /launch-video Veo prompt patterns
"Subtitles are unreadable on mobile" /blog-post reel-production.md
"Hook doesn't grab in 3 seconds" /blog-post reel-production.md hook formulas
"Glass panels look flat" /launch-video GlassPanel component
"Pacing too fast" /launch-video scene duration guidance

Agent Behavior

On Review Invocation

  1. Identify the asset — video file path, URL, or generated content package
  2. Identify the reference — style brief, reference URL, or skill defaults (e.g., /launch-video checklist)
  3. Extract frames from both (if video)
  4. Score each dimension against the brief
  5. Generate feedback with specific fixes
  6. Report — overall score, must-fix items, what worked

Using /agent-browser for Video Review

When the asset is a deployed video (hosted URL):

# Install if needed
npm install -g @anthropic-ai/agent-browser

# Open and screenshot for review
agent-browser open "https://broomva.tech/images/writing/slug/reel.mp4"
agent-browser screenshot --output review-frame.png

For local files, use ffmpeg frame extraction instead.

Quick Review Command

/creative-review /path/to/video.mp4 --against launch-video

This automatically:

  1. Extracts frames from the video
  2. Loads the /launch-video quality checklist
  3. Scores each dimension
  4. Reports findings with specific fixes

Quality Gate Integration

The creative review can be wired as a gate in the /blog-post pipeline:

Phase 6 (Media) → Generate video
  → /creative-review scores the output
    → Score ≥ 75? → Proceed to Phase 7
    → Score < 75? → Iterate with feedback → Re-generate → Re-review

This creates an automatic quality loop — no substandard creative ships.