watch-video

name	watch-video
description	Use ONLY when the user explicitly asks to "Watch Video", "watch-video", or run the Watch Video skill. Builds a multimodal evidence bundle from a video URL or local media file by extracting transcript/subtitles plus ffmpeg screenshots at timed intervals so the AI can inspect what was visible on screen.
metadata	{"short-description":"Watch videos with transcript plus timed screenshots"}

Watch Video

Use this skill only when the user explicitly says Watch Video or names watch-video.

This is different from a transcript job. It creates a bundle with:

• source media from a URL or local video file
• subtitles/transcript, using local WhisperX/Whisper if captions are unavailable
• screenshots sampled with ffmpeg
• watch_context.md, pairing each sampled frame timestamp with nearby transcript text
• manifest.json, the machine-readable version of the same context

Command

From the installed skill folder:

python3 "$HOME/.claude/skills/watch-video/scripts/watch_video.py" "VIDEO_URL_OR_FILE" --output-dir "DEST_DIR" --rate 1s

Default rate is 1s, one frame every second.

Supported presets:

4fps
2fps
1fps
1s
3s
5s
10s

Numeric seconds also work:

--rate 15
--rate 0.5

Rate Selection

• Default to 1s.
• Use 4fps or 2fps only when the user needs detailed UI/action analysis or the clip is short.
• Use 3s, 5s, or 10s for long videos, lectures, podcasts, or low-motion screen recordings.
• If the user asks for a specific cadence, use it directly.

Workflow

1. Run the script with the requested source and rate.
2. Open watch_context.md and inspect relevant frame links plus transcript context.
3. For detailed visual questions, use local image inspection on specific frames rather than relying only on the markdown.
4. Answer from both sources: what was said and what was visible.

Dependencies

The script expects:

• yt-dlp
• ffmpeg
• local whisperx or whisper for caption fallback

It checks PATH first, then common local install paths:

$HOME/.local/bin/whisperx
$HOME/.local/bin/whisper
$HOME/.buttercut/venv/bin/whisperx
$HOME/.perfect-cuts/venv/bin/whisperx
$HOME/GitHub/Marketing Assets and Skills/.venv-whisperx/bin/whisperx
$HOME/Library/Python/3.9/bin/whisperx
$HOME/Library/Python/3.9/bin/whisper

Optional overrides:

YTDLP_BIN=/path/to/yt-dlp
WHISPERX_BIN=/path/to/whisperx
WHISPER_BIN=/path/to/whisper
WHISPER_MODEL=base
WHISPER_DEVICE=cpu
WHISPER_COMPUTE_TYPE=int8

The script detects yt-dlp in this order:

$YTDLP_BIN
python3 -m yt_dlp
yt-dlp on PATH