Image Editor Skill

Core Rules (NEVER BREAK)

1. Preserve original resolution — never resize, scale, or change dimensions
2. Verify every output with vision model — use the image tool to check the result before delivering
3. Use PIL directly — Python PIL/Pillow for pixel-level manipulation, not external tools

Workflow

Step 1: Analyze the original image

python3 -c "
from PIL import Image
img = Image.open('PATH')
print(f'Size: {img.size}, Mode: {img.mode}')
"

Step 2: Identify target region

• Open the image with read tool (supports jpg/png)
• Use image tool with prompt "identify the exact pixel coordinates of [target area]"
• Record: x_start, y_start, x_end, y_end

Step 3: Edit with PIL

from PIL import Image, ImageDraw, ImageFont
img = Image.open('INPUT')
draw = ImageDraw.Draw(img)

# For text replacement:
# 1. Fill background color over old text
# 2. Draw new text with matching font/color
# macOS Chinese fonts: /System/Library/Fonts/STHeiti Light.ttc

font = ImageFont.truetype('/System/Library/Fonts/STHeiti Light.ttc', size=PIXEL_SIZE)
draw.text((x, y), 'NEW TEXT', fill=(R, G, B), font=font)
img.save('OUTPUT.jpg', quality=95)

Step 4: Verify output (MANDATORY)

Use the image tool on the output file with prompt: "Compare this edited image with the original at ORIGINAL_PATH. Is the text replacement correct? Is resolution preserved? Any artifacts?"

Step 5: Only deliver after verification passes

Common mistakes (from real failures)

• ❌ Using web_fetch to "download" an image (it resizes)
• ❌ Skipping vision verification
• ❌ Calling external APIs that don't support the format
• ❌ Saving with wrong quality/format
• ✅ PIL direct pixel manipulation
• ✅ Vision verification of every output