Gen thumbnail

/gen-thumbnail

Mode: STRICT PRO CREATOR

If ANY rule is violated the output is INVALID. Regenerate.

Core Principle

Thumbnail = Emotion + Curiosity. NOT information. Think like a Click Psychologist, not a Designer.

Architecture

The pipeline uses a 3-layer composite approach:

1. Layer 1 (Background): AI generates ONLY a dark cinematic scene with secondary elements (money, laptop, particles). NO face, NO character, NO text.
2. Layer 2 (Mascot): The real assets/mascot.png is composited onto the left side at ~60-70% of the frame. The mascot IS the face. AI never generates a face.
3. Layer 3 (Typography): Huge, slanted, multi-weight text with heavy drop shadows is rendered by the compositor. AI never generates text.

Hard Rules (Non-Negotiable)

1. TEXT

• MAX 3 words per thumbnail
• First word = white, huge, top area
• Second word = yellow, even bigger, below first
• Optional subline in smaller green
• Slanted ~5-8 degrees for dynamism
• Heavy black stroke + drop shadow
• MUST be readable at 168x94px (mobile search size)
• ONE text block only, right side of frame

If text is cut, small, or too long → FAIL

2. FACE (MANDATORY)

• assets/mascot.png is ALWAYS the main subject
• Composited by code, NEVER generated by AI
• Occupies 60-70% of image, left-dominant
• Subtle glow effect around mascot edges
• AI prompt must say "NO face, NO character, NO person, empty dark left side"

If AI generates a face → FAIL

3. ELEMENT LIMIT

MAX 3 total elements in the AI-generated background:

1. ONE object (money stack / laptop with graph / glowing screen)
2. ONE effect (floating bills / spark particles / green streaks)
3. Dark atmospheric background

If more than 3 → FAIL

4. BACKGROUND

• Simple dark studio / dark gradient ONLY
• Green/cyan neon accents allowed
• NO busy scenes, NO landscapes, NO rooms
• EACH variation must use DIFFERENT background mood

5. SAFE AREA

• All important elements inside center 80%
• NOTHING important near edges
• Bottom-right reserved for YouTube timestamp

6. A/B TEST (CRITICAL)

Generate EXACTLY 2 variations. They MUST be fundamentally different:

Variation A (MINIMAL):

• Fewer background elements
• Clean dark space
• Question-style text ("REPLACED?" / "OBSOLET?")

Variation B (DRAMATIC):

• Explosive particles, flying money
• Intense neon glow
• Statement-style text ("KILLS AI" / "ZERSTÖRT SIE")

If A and B look similar → FAIL

7. FORBIDDEN

• Human faces or generic masked characters in AI generation
• Small text or multiple text blocks
• Paragraphs or explanations
• Technical diagrams or charts
• More than 3 visual elements in background
• Clutter of any kind
• Text generated inside the AI image

Execution

node ~/.config/opencode/skills/gen-thumbnail/scripts/generate_thumbnail_pipeline.js "YOUR TOPIC"

Output

• outputs/thumbnails/thumbnail_A.png (1920x1080)
• outputs/thumbnails/thumbnail_B.png (1920x1080)
• outputs/previews/preview.html (inline base64, works offline)
• outputs/data/last_run.json

Quality Gate (MANDATORY)

• Understood in less than 1 second?
• Only ONE focus (mascot face)?
• Text readable at mobile size?
• 3 or fewer background elements?
• A and B clearly different?
• Real mascot.png used, not AI face?

If ANY = NO → regenerate

Learning Loop

node ~/.config/opencode/skills/gen-thumbnail/scripts/track_ctr.js A 8.2 5.1 \
  --do-more "clean dark background" \
  --do-less "clutter" \
  --brand-token "green-masked mascot" \
  --visual-motif "floating money"

Performance memory stores ONLY short traits (max 5 words each), never full prompts.

Related Skills

• See Skill Catalog for related skills