gemini-video-understanding（Gemini video understanding 视频）

Gemini Video Understanding Skill

This skill enables comprehensive video analysis using Google's Gemini API, including video summarization, question answering, transcription, timestamp references, and more.

Capabilities

• Video Summarization: Create concise summaries of video content
• Question Answering: Answer specific questions about video content
• Transcription: Transcribe audio with visual descriptions and timestamps
• Timestamp References: Query specific moments in videos (MM:SS format)
• Video Clipping: Process specific segments using start/end offsets
• Multiple Videos: Compare and analyze up to 10 videos (Gemini 2.5+)
• YouTube Support: Analyze YouTube videos directly (preview feature)
• Custom Frame Rate: Adjust FPS sampling for different video types

Supported Formats

• MP4, MPEG, MOV, AVI, FLV, MPG, WebM, WMV, 3GPP

Models Available

Gemini 2.5 Series:

• gemini-2.5-pro - Best quality, 1M context
• gemini-2.5-flash - Balanced quality/speed, 1M context
• gemini-2.5-flash-preview-09-2025 - Preview features, 1M context

Gemini 2.0 Series:

• gemini-2.0-flash - Fast processing
• gemini-2.0-flash-lite - Lightweight option

Context Windows:

• 2M token models: ~2 hours (default) or ~6 hours (low-res)
• 1M token models: ~1 hour (default) or ~3 hours (low-res)

API Key Configuration

The skill supports both Google AI Studio and Vertex AI endpoints.

Option 1: Google AI Studio (Default)

The skill checks for GEMINI_API_KEY in this order:

1. Process environment: process.env.GEMINI_API_KEY or $GEMINI_API_KEY
2. Project root: .env
3. .claude directory: .claude/.env
4. .claude/skills directory: .claude/skills/.env
5. Skill directory: .claude/skills/gemini-video-understanding/.env

Get your API key: https://aistudio.google.com/apikey

To set up:

# Environment variable (recommended)
export GEMINI_API_KEY="your-api-key-here"

# Or in .env file
echo "GEMINI_API_KEY=your-api-key-here" > .env

Option 2: Vertex AI

To use Vertex AI instead:

# Enable Vertex AI
export GEMINI_USE_VERTEX=true
export VERTEX_PROJECT_ID=your-gcp-project-id
export VERTEX_LOCATION=us-central1  # Optional, defaults to us-central1

Or in .env file:

GEMINI_USE_VERTEX=true
VERTEX_PROJECT_ID=your-gcp-project-id
VERTEX_LOCATION=us-central1

Usage Instructions

When to Use This Skill

Use this skill when the user asks to:

• Analyze, summarize, or describe video content
• Answer questions about videos
• Transcribe video audio with visual context
• Extract information from specific timestamps
• Compare multiple videos
• Process YouTube video content
• Create quizzes or educational content from videos

Basic Video Analysis

For video files:

python .claude/skills/gemini-video-understanding/scripts/analyze_video.py \
  --video-path "/path/to/video.mp4" \
  --prompt "Summarize this video in 3 key points"

For YouTube URLs:

python .claude/skills/gemini-video-understanding/scripts/analyze_video.py \
  --youtube-url "https://www.youtube.com/watch?v=VIDEO_ID" \
  --prompt "What are the main topics discussed?"

Advanced Features

Video Clipping (specific time range):

python .claude/skills/gemini-video-understanding/scripts/analyze_video.py \
  --video-path "/path/to/video.mp4" \
  --prompt "Summarize this segment" \
  --start-offset "40s" \
  --end-offset "80s"

Custom Frame Rate:

python .claude/skills/gemini-video-understanding/scripts/analyze_video.py \
  --video-path "/path/to/video.mp4" \
  --prompt "Analyze the rapid movements" \
  --fps 5

Transcription with Timestamps:

python .claude/skills/gemini-video-understanding/scripts/analyze_video.py \
  --video-path "/path/to/video.mp4" \
  --prompt "Transcribe the audio with timestamps and visual descriptions"

Multiple Videos (Gemini 2.5+ only):

python .claude/skills/gemini-video-understanding/scripts/analyze_video.py \
  --video-paths "/path/video1.mp4" "/path/video2.mp4" \
  --prompt "Compare these two videos and highlight the differences"

Model Selection:

python .claude/skills/gemini-video-understanding/scripts/analyze_video.py \
  --video-path "/path/to/video.mp4" \
  --prompt "Detailed analysis" \
  --model "gemini-2.5-pro"

Script Parameters

Required (one of):
  --video-path PATH           Path to local video file
  --youtube-url URL           YouTube video URL
  --video-paths PATH [PATH..] Multiple video paths (Gemini 2.5+)

Required:
  --prompt TEXT              Analysis prompt/question

Optional:
  --model NAME               Model to use (default: gemini-2.5-flash)
  --start-offset TIME        Video clip start (e.g., "40s", "1m30s")
  --end-offset TIME          Video clip end (e.g., "80s", "2m")
  --fps NUMBER               Frame sampling rate (default: 1)
  --output-file PATH         Save response to file
  --verbose                  Show detailed processing info

Common Use Cases

1. Video Summarization

Prompt: "Summarize this video in 3 key points with timestamps"

2. Educational Content

Prompt: "Create a quiz with 5 questions and answer key based on this video"

3. Timestamp-Specific Questions

Prompt: "What happens at 01:15 and how does it relate to the topic at 02:30?"

4. Transcription

Prompt: "Transcribe the audio from this video with timestamps for salient events and visual descriptions"

5. Content Comparison

Prompt: "Compare these two product demo videos. Which one explains the features more clearly?"

6. Action Detection

Prompt: "List all the actions performed in this tutorial video with timestamps"

Rate Limits & Quotas

Free Tier (per model):

• 10-15 RPM (requests per minute)
• 1M-4M TPM (tokens per minute)
• 1,500 RPD (requests per day)

YouTube Limitations:

• Free tier: 8 hours of YouTube video per day
• Paid tier: No length-based limits
• Public videos only (no private/unlisted)

Storage (Files API):

• 20GB per project
• 2GB per file
• 48-hour retention period

Token Calculation

Video tokens depend on resolution:

• Default resolution: ~300 tokens per second of video
• Low resolution: ~100 tokens per second of video

Example: A 10-minute video = 600 seconds × 300 tokens = ~180,000 tokens

Error Handling

Common errors and solutions:

Best Practices

1. Use Files API for videos >20MB - More reliable than inline data
2. Wait for file processing - Poll until state is ACTIVE before analysis
3. Optimize FPS - Use lower FPS for static content to save tokens
4. Clip long videos - Process specific segments instead of entire video
5. Cache context - Reuse uploaded files for multiple queries
6. Batch processing - Process multiple short videos in one request (2.5+)
7. Specific prompts - Be precise about what you want to extract

Implementation Notes

For Claude Code:

When a user requests video analysis:

1. Check API key availability first using the helper script
2. Determine video source: local file, YouTube URL, or multiple videos
3. Select appropriate model based on requirements (default: gemini-2.5-flash)
4. Run the analysis script with proper parameters
5. Parse and present results to the user clearly
6. Handle errors gracefully with helpful suggestions

Files API Workflow:

For videos >20MB or reusable content:

1. Upload video using Files API (script handles this automatically)
2. Wait for ACTIVE state (polling included in script)
3. Use file URI for analysis
4. Files auto-delete after 48 hours

Inline Data Workflow:

For videos <20MB:

1. Read video file as bytes
2. Base64 encode for API
3. Send in generateContent request
4. Single-use, no upload needed

Example Workflows

Workflow 1: YouTube Video Summary

# User: "Analyze this YouTube tutorial video"
python .claude/skills/gemini-video-understanding/scripts/analyze_video.py \
  --youtube-url "https://www.youtube.com/watch?v=abc123" \
  --prompt "Create a structured summary with: 1) Main topics, 2) Key takeaways, 3) Recommended audience"

Workflow 2: Interview Transcription

# User: "Transcribe this interview with timestamps"
python .claude/skills/gemini-video-understanding/scripts/analyze_video.py \
  --video-path "interview.mp4" \
  --prompt "Transcribe this interview with speaker labels, timestamps, and visual descriptions of gestures or slides shown"

Workflow 3: Product Comparison

# User: "Compare these two product demo videos"
python .claude/skills/gemini-video-understanding/scripts/analyze_video.py \
  --video-paths "demo1.mp4" "demo2.mp4" \
  --model "gemini-2.5-pro" \
  --prompt "Compare these product demos on: features shown, presentation quality, clarity of explanation, and overall effectiveness"

Troubleshooting

API Key Not Found:

# Check API key detection
python .claude/skills/gemini-video-understanding/scripts/check_api_key.py

Video Too Large:

Error: Request size exceeds 20MB
Solution: Script automatically uses Files API for large videos

Processing Timeout:

Error: File not reaching ACTIVE state
Solution: Check video integrity, try smaller file, or different format

Rate Limit Errors:

Error: 429 Too Many Requests
Solution: Wait before retry, or upgrade to paid tier

Additional Resources

• API Documentation: https://ai.google.dev/gemini-api/docs/video-understanding
• Files API Guide: https://ai.google.dev/gemini-api/docs/vision#uploading-files
• Rate Limits: https://ai.google.dev/gemini-api/docs/rate-limits
• Pricing: https://ai.google.dev/pricing
• Get API Key: https://aistudio.google.com/apikey

Version History

• 1.0.0 (2025-10-26): Initial release with full video understanding capabilities