Agent Skills: Gemini Vision Skill

Gemini Vision Skill

Generate AI images and videos by invoking Gemini CLI's vision extension. This skill provides access to:

• Nano Banana (gemini-2.5-flash-image) - Image generation and transformation
• Veo 3 (veo-3.0-generate-001) - Video generation from images
• Webcam capture - Live frame capture for AI processing

Prerequisites

1. Gemini CLI: Must be installed and configured
2. Vision Extension: Install via:

   gemini extensions install vision
   

3. API Key: Set GEMINI_API_KEY environment variable

When to Use This Skill

Use this skill when the user asks to:

• Generate images from text prompts
• Transform or reimagine existing images
• Create AI-generated videos from images
• Capture webcam frames for AI processing
• Create "nano banana" style images
• Generate Veo videos

Available Operations

1. Image Generation (Nano Banana)

Generate images from text prompts or transform existing images.

Command Pattern:

gemini -p "/vision:banana prompt=\"Your creative prompt here\" n=1 out_dir=./output"

Parameters: | Parameter | Default | Description | |-----------|---------|-------------| | prompt | Required | Creative description of desired image | | n | 1 | Number of images to generate | | out_dir | "." | Output directory for images | | model | gemini-2.5-flash-image | Image generation model |

Models Available:

• gemini-2.5-flash-image (default, recommended)
• gemini-3-pro-image-preview (newer, experimental)

2. Video Generation (Veo 3)

Generate short videos from images or prompts.

Command Pattern:

gemini -p "/vision:veo prompt=\"Animate this scene\" aspect_ratio=16:9 out_dir=./output"

Parameters: | Parameter | Default | Description | |-----------|---------|-------------| | prompt | Required | Animation/motion description | | aspect_ratio | "16:9" | Video aspect ratio (16:9 or 9:16) | | resolution | auto | Video resolution (e.g., "1080p") | | negative_prompt | "" | What to avoid in video | | veo_model | veo-3.0-generate-001 | Video model |

3. Webcam Capture + AI

Capture from webcam and process with AI.

# Start camera
gemini -p "/vision:start"

# Capture and transform
gemini -p "/vision:banana prompt=\"Transform into oil painting\""

# Stop camera
gemini -p "/vision:stop"

Instructions for Claude

When the user requests image or video generation:

1. Determine the operation type:

   • Text-to-image → Use /vision:banana
   • Image transformation → Use /vision:banana with input image
   • Image-to-video → Use /vision:veo
   • Webcam capture → Use /vision:capture or /vision:banana

2. Construct the Gemini CLI command:

   gemini -p "/vision:<command> prompt=\"<user prompt>\" <params>"
   

3. Execute via Bash tool:

   • Run the command
   • Capture the output paths
   • Report success and file locations to user

4. Handle output:

   • Images saved as banana_*.png or banana_*.jpg
   • Videos saved as veo_*.mp4
   • Return the file paths to the user

Example Workflows

Generate a Single Image

User: "Generate an image of a cyberpunk city at sunset"

Action:

gemini -p "/vision:banana prompt=\"A sprawling cyberpunk city at sunset, neon lights reflecting off wet streets, flying cars in the distance, highly detailed, cinematic\" n=1 out_dir=."

Transform an Image

User: "Make this photo look like a Studio Ghibli scene" (with image attached)

Action:

1. Save the attached image to a temp location
2. Run:

gemini -p "/vision:banana prompt=\"Transform into Studio Ghibli animation style, soft colors, whimsical atmosphere\" input_paths=['/path/to/image.jpg']"

Generate a Video

User: "Create a video of ocean waves"

Action:

gemini -p "/vision:veo prompt=\"Calm ocean waves gently rolling onto a sandy beach, golden hour lighting, peaceful atmosphere\" aspect_ratio=16:9"

Webcam to Art

User: "Take a photo of me and make it look like a Renaissance painting"

Action:

# Capture and transform in one step
gemini -p "/vision:banana prompt=\"Transform into a Renaissance oil painting, dramatic lighting, classical composition\""

Output Format

Always report results in this format:

## Generated Content

**Type:** Image/Video
**Files:**
- `/path/to/banana_20251227_123456_000.png`

**Prompt Used:** [the prompt]
**Model:** gemini-2.5-flash-image

To view: Open the file path above or use `open /path/to/file`

Error Handling

Common issues and solutions:

| Error | Solution | |-------|----------| | "Camera not found" | Run /vision:devices to list cameras | | "GEMINI_API_KEY not set" | Export the API key in environment | | "Model not available" | Check model ID spelling | | "Generation failed" | Try simpler prompt or different model |

Script Usage (Alternative)

For programmatic access, use the helper script:

python ~/.claude/skills/gemini-vision/scripts/gemini_vision.py \
  --operation banana \
  --prompt "Your prompt here" \
  --output-dir ./output \
  --count 1

Options:

• --operation: banana, veo, capture, devices
• --prompt: The generation prompt
• --output-dir: Where to save files
• --count: Number of images (for banana)
• --aspect-ratio: For veo (16:9 or 9:16)
• --model: Override default model