Vision Helper — AI Image Analysis

Basic

# Analyze an image (default: English description)
python3 /scripts/analyze_image.py 
With a custom prompt
python3 /scripts/analyze_image.py  "Is this a chess game? Describe the board state"
With a specific model
python3 /scripts/analyze_image.py  "Describe content" kimi-k2.5:cloud

> resolves to your OpenClaw skill installation directory, typically ~/.openclaw/workspace/skills/vision-helper/.

In Conversation

When you need to analyze an image, use the exec tool:

exec: python3 /scripts/analyze_image.py /path/to/image.png "What do you see?"

Important: Set exec timeout to 120–180 seconds, as cloud vision models are slow.

Screenshot + Analysis Workflow

#### Option A: Browser screenshot → analyze

1. browser(action="screenshot") → get screenshot path (MEDIA: xxx)
2. exec("/scripts/analyze_image.py  'Describe this UI'")
3. Act on the analysis result

#### Option B: Desktop screenshot → analyze

macOS:

1. exec("screencapture -x /tmp/screen.png")
2. exec("/scripts/analyze_image.py /tmp/screen.png 'Describe the desktop'")

Linux:

1. exec("gnome-screenshot -f /tmp/screen.png")
   — or —
   exec("import /tmp/screen.png")  # ImageMagick
   — or —
   exec("scrot /tmp/screen.png")
2. exec("/scripts/analyze_image.py /tmp/screen.png 'Describe the desktop'")

#### Option C: Game/App UI → analyze → act

1. Screenshot the current screen
2. Use vision-helper to identify UI elements, buttons, text
3. Execute clicks/input based on the analysis

Q: Can I use the built-in image tool instead?

Q: What image formats are supported?

Q: Where should I save screenshots?

Q: How do I use a Chinese prompt?