Imported from janzheng/wod (.claude/skills/wod-imagegen/SKILL.md). Install upstream with npx skills add janzheng/wod --skill wod-imagegen. Copyright stays with the author.

WOD Image Generation

Generate minimalist workout exercise images using Google's Gemini API. Produces consistent coral/peach flat-color illustrations matching the WOD app aesthetic.

When to Use

• Create a new exercise image based on an existing reference
• Update/regenerate exercise images with consistent style
• Batch process multiple images
• User mentions "generate image", "create workout image", "make exercise illustration"

Quick Start

cd /Users/janzheng/conductor/workspaces/wod/san-jose

# Generate from exercise JSON as JPG (recommended)
deno run -A .claude/skills/wod-imagegen/generate.ts --exercise barre-bent-leg-side-lift --no-ref --jpg

# With custom quality (1-100, default 80)
deno run -A .claude/skills/wod-imagegen/generate.ts --exercise bird-dog --no-ref --jpg --quality 60

# Or with custom prompt
deno run -A .claude/skills/wod-imagegen/generate.ts \
  --output static/new-images/fire-hydrant.png \
  --prompt "Fire Hydrant: On all fours, lift bent knee out to the side" \
  --jpg

Prerequisites

1. Get API key from https://aistudio.google.com/apikey
2. Create apps/wod/.env:

   GEMINI_API_KEY=your_api_key_here
   

Usage Examples

Generate Single Image

Always provide a reference image for style consistency:

deno run -A .claude/skills/wod-imagegen/generate.ts \
  --reference static/images/barre-clamshell.png \
  --output static/images/side-lying-leg-raise.png \
  --prompt "person lying on side, lifting top leg straight up"

JPG Output (Recommended)

Use --jpg to convert output to JPG for smaller file sizes (~70-80% reduction):

# Default quality (80)
deno run -A .claude/skills/wod-imagegen/generate.ts --exercise bird-dog --no-ref --jpg

# Lower quality for smaller files
deno run -A .claude/skills/wod-imagegen/generate.ts --exercise bird-dog --no-ref --jpg --quality 60

Uses macOS sips for conversion. PNG is automatically removed after conversion.

Custom Style Override

deno run -A .claude/skills/wod-imagegen/generate.ts \
  --output static/images/new.png \
  --prompt "your custom prompt" \
  --style "photorealistic fitness photo, gym background" \
  --jpg

List Available Reference Images

ls apps/wod/static/images/*.png | head -20

WOD Style

The default style produces warm coral/peach illustrations matching the WOD app aesthetic:

Color Palette (Hex) - Flat Solid Colors

Style Details

• Colors: FLAT SOLID FILLS only - cel-shaded/vector style
• Skin: Uniform single color, NO blotches/gradients/texture
• Hands: Defined fingers (not blobs)
• Hair: Simple strand lines
• Clothes: Sports bra + shorts with fold lines (not color variation)
• Lines: Thin, elegant strokes throughout
• Figure: Faceless (no facial features)
• Background: Pure white, NO floor/ground line, NO shadows

Default prompt (can be overridden with --style):

Minimalist fitness illustration with flat, solid colors (cel-shaded style):

Color palette - use FLAT SOLID FILLS, no gradients or texture:
- Skin: single flat peachy-beige (#F7DCC9) - ONE solid color, NO blotches, NO gradients, NO shading
- Hair: flat muted auburn (#C97B62) with simple highlight streaks (#D4896F)
- Clothes: flat dusty coral (#E9A090) sports bra and shorts - ONE solid color per garment
- Outlines: soft terracotta lines (#D49080) - thin and delicate

CRITICAL style requirements:
- FLAT COLORS ONLY - skin must be one uniform solid color with no variation
- NO gradients, NO shading, NO blotchy textures on skin or clothes
- Clean cel-shaded/vector illustration style
- Hands with defined fingers (not blob hands)
- Hair with simple strand lines
- Clothes with simple fold lines (not color variation)
- Thin, elegant line work throughout
- Faceless figure (no facial features, just head shape)
- Pure white background ONLY - NO floor line, NO ground line, NO shadows
- Figure floats in empty white space

Do NOT include: text, labels, watermarks, ground lines, skin texture, color gradients, or shading.

API Details

• Model: gemini-3-pro-image-preview
• Endpoint: https://generativelanguage.googleapis.com/v1beta/models/gemini-3-pro-image-preview:generateContent
• Auth: API key as query parameter (?key=...)

Workflow for New Exercise

1. Generate - Run with --no-ref --jpg for text-only generation
2. Critique - Review output for common defects (see checklist below)
3. Iterate - If defects found, refine prompt and regenerate
4. Save - Update exercise JSON with working imagePrompt field

Image Critique Checklist

After generating an image, always review for these common defects:

Pose Accuracy

• Correct exercise position - Does the pose match the exercise description?
• Proper joint angles - Are knees/elbows bent correctly (not hyper-extended)?
• Anatomically correct - No impossible body positions?

Arm Issues (Common!)

• No "T-Rex arms" - Arms shouldn't be awkwardly bent with palms up
• Correct arm position for exercise - Ballet exercises need proper positions:
  • First position: rounded in front, like holding a beach ball
  • Second position: extended gracefully to sides
  • Hands on hips: for stability exercises
• Defined hands - Fingers visible, not blob hands

Leg Issues (Common!)

• No accidental splits - Legs bent at knees, not extended into split
• Not sitting on ground - Standing exercises should be STANDING, not seated
• Correct stance width - Wide plié ≠ full splits
• Knees tracking over toes - Proper alignment for squats/pliés

Figure Issues

• Single figure only - Not multiple figures or progression sequences
• Faceless - No facial features, just blank head shape
• Not floating - Figure should feel grounded (even without floor line)

Style Consistency

• Flat colors - No gradients or shading on skin/clothes
• White background - No ground lines, shadows, or floor
• Correct color palette - Coral/peach tones matching WOD style

Prompt Tips for Common Issues

Exercise JSON: imagePrompt Field

For exercises that need custom prompts, add an imagePrompt field:

{
  "id": "barre-wide-plie-lift",
  "name": "Wide Plié Lift",
  "description": "Stand with feet wider than shoulder-width...",
  "imagePrompt": "Single figure doing Wide Plié squat: ONE person standing with feet wider than shoulder-width, toes pointed outward. Deep squat with knees bent tracking over toes. Arms in ballet first position - rounded in front of body at hip level. Viewed from the front."
}

The generate script can be extended to use imagePrompt when available for better results.

Troubleshooting

"API key not configured": Create .env file with GEMINI_API_KEY

Poor output quality: Try a different reference image or more specific prompt

Rate limits: The API has usage limits - wait and retry