watch-youtube（观看 YouTube）

Watch YouTube

Use Google's Gemini API to actually *watch* YouTube videos and answer questions about them.

How It Works

Gemini processes both audio and visual streams of YouTube videos at 1 FPS. You pass a URL + prompt, it returns analysis.

When to Use

• User shares a YouTube URL and wants a summary, transcript, or analysis
• User asks "what did they say about X in this video?"
• User wants timestamps of key moments
• User wants to compare multiple videos (up to 10 per request with Gemini 2.5+)

When NOT to Use

• Private or unlisted videos (only public videos work)
• User just wants the video link or metadata

Setup

Requires GOOGLE_API_KEY environment variable. Get one free at https://aistudio.google.com/apikey

Usage

Run the script:

GOOGLE_API_KEY="$GOOGLE_API_KEY" python3 ~/.openclaw/workspace/skills/watch-youtube/watch.py "<youtube_url>" "<prompt>"

Examples

Summarize:

watch.py "https://www.youtube.com/watch?v=VIDEO_ID" "Summarize this video in 5 bullet points"

Timestamps:

watch.py "https://www.youtube.com/watch?v=VIDEO_ID" "List the key moments with timestamps"

Q&A:

watch.py "https://www.youtube.com/watch?v=VIDEO_ID" "What tools or products did they mention?"

Specific section:

watch.py "https://www.youtube.com/watch?v=VIDEO_ID" "What happens at 05:30?"

Limits

• Free tier: 8 hours of YouTube video per day
• Paid tier: No limit
• Max videos per request: 10 (Gemini 2.5+)
• Max video length: ~1 hour (1M context), ~3 hours (low res)
• ~300 tokens per second of video

Models

• gemini-2.5-flash — fast, cheap, good for most use cases
• gemini-2.5-pro — deeper analysis, longer videos
• gemini-3-flash-preview — latest, best quality

Fallback: Transcript Mode

If Gemini fails (quota exceeded, video too long, API key missing), the script automatically falls back to fetching the YouTube transcript via youtube_transcript_api.

What the fallback does:

1. Extracts the video ID from the URL
2. Fetches the auto-generated captions with timestamps
3. Returns the full transcript as text with your original prompt

What you get back: Raw timestamped transcript — the calling agent should read and answer the prompt from that text. Gemini-level visual/audio analysis is not available in fallback mode.

Limitations of fallback:

• No visual analysis (can't describe what's on screen)
• No answers about non-speech content
• Requires the video to have auto-generated or manual captions
• Private/unlisted videos without captions will fail entirely

When fallback triggers:

• Gemini quota exceeded (429 / RESOURCE_EXHAUSTED)
• Video exceeds context window (>1M tokens / ~1hr)
• GOOGLE_API_KEY not set
• Any other Gemini API error containing: quota, token, context, too long, exceeds