Video Transcribe Skill

Download a video from any source and transcribe it with mlx_whisper on Apple Silicon.

Required Tools

• yt-dlp — video downloading (streaming platforms)
• ffmpeg — audio extraction
• mlx_whisper — Apple Silicon GPU transcription
• node — for Google Drive protected video download
• drive-video-download skill — for canDownload:false Drive files

Step 1: Identify the video source

Step 2: Generate a timestamp-based filename

Always derive the output filename from the current timestamp, not from the video title or date. This ensures multiple transcriptions on the same day never overwrite each other.

FILENAME=$(date +%Y%m%d-%H%M%S)
echo "Using filename stem: $FILENAME"

Use $FILENAME as the stem for all audio, transcript, and summary files below.

Step 3: Download audio

Streaming platforms (Vimeo, YouTube, Loom, etc.)

CRITICAL: Always prefer HTTP single-file format over fragmented DASH/HLS streams. DASH and HLS formats split audio into hundreds of tiny fragments, resulting in ~150 KB/s downloads due to per-fragment HTTP overhead. HTTP formats download as a single file at ~8-14 MB/s.

# List available formats
yt-dlp -F "VIDEO_URL"

# Download using HTTP format (fastest)
# For Vimeo: http-240p (lowest quality with audio, ~120MB/hr, downloads in ~20s)
# For YouTube: pick a format with protocol "https" (not "m3u8" or "dash")
yt-dlp -f "http-240p" -o "/tmp/video-$FILENAME.%(ext)s" "VIDEO_URL"

# Extract audio
ffmpeg -y -i /tmp/video-$FILENAME.mp4 -vn -acodec libmp3lame -q:a 2 /tmp/video-$FILENAME.mp3

Format selection priority:

1. http-* formats (protocol: https) — fastest, single-file
2. hls-audio-* formats (protocol: m3u8) — fallback, audio-only fragmented
3. dash-audio-* formats (protocol: dash) — AVOID, slowest

Fallback if no HTTP format:

yt-dlp -x --audio-format mp3 -o "/tmp/video-%(id)s.%(ext)s" "VIDEO_URL"

Google Drive videos — canDownload: false (corporate recordings, Meet recordings)

yt-dlp and the Drive API both return 403. Use the drive-video-download skill:

Prerequisites: Brave running, user logged in, remote debugging enabled (brave://inspect/#devices → toggle).

node ~/.claude/skills/drive-video-download/scripts/download.mjs <file-id-or-url> /tmp/video-$FILENAME.m4a

The script intercepts the workspacevideo-pa streaming manifest via CDP and downloads with exact browser headers. No conversion needed — .m4a is accepted directly by mlx_whisper.

Google Drive videos — canDownload: true

mcp__gws__drive_files_get with params: {"fileId": "...", "alt": "media", "supportsAllDrives": true}

Save to /tmp/video-$FILENAME.mp4, then extract audio with ffmpeg if needed.

Step 4: Transcribe

IMPORTANT: Use the local HuggingFace snapshot path, NOT the model name string. The short form (mlx-community/whisper-medium-mlx) fails with FileNotFoundError because Hub path resolution is broken in this environment.

The language is taken from the 3rd skill argument (e.g. uk, de). Default to en if not provided.

# Resolve local snapshot path
SNAPSHOT_DIR=~/.cache/huggingface/hub/models--mlx-community--whisper-large-v3-turbo/snapshots
SNAPSHOT=$(ls "$SNAPSHOT_DIR" | head -1)
MODEL_PATH="$SNAPSHOT_DIR/$SNAPSHOT"

# LANGUAGE = 3rd skill argument, default "en"
LANGUAGE="${3:-en}"

# Run in background for long videos (NEVER pipe — SIGPIPE kills mlx_whisper silently)
nohup mlx_whisper /tmp/video-$FILENAME.m4a \
  --model "$MODEL_PATH" \
  --output-name "$FILENAME" \
  --language "$LANGUAGE" \
  --output-format all \
  --output-dir /tmp/transcripts \
  > /tmp/whisper-$FILENAME.log 2>&1 &

# Monitor progress
tail -f /tmp/whisper-$FILENAME.log

Critical rules:

• Always use mlx_whisper — runs on Apple Silicon GPU (~3 min/hour of video on M4 Pro)
• Never use whisper (openai-whisper) — falls back to CPU, 10-20x slower
• Never use --device mps with openai-whisper — produces NaN errors on Apple Silicon
• Always run as background job (nohup ... &) for videos over 10 minutes
• Never pipe (| head, | tee) — SIGPIPE silently kills the process
• .m4a input works directly — no need to convert to .mp3 first
• If model not downloaded yet: mlx_whisper --model mlx-community/whisper-large-v3-turbo /dev/null downloads it (ignore the error)

Step 5: Summarize (optional)

If the 4th skill argument is summarize (or the user requested a summary), read the transcript and produce a structured summary.

cat /tmp/transcripts/$FILENAME.txt

Then write a markdown summary to /tmp/transcripts/$FILENAME-summary.md covering:

• Date / meeting title (from filename or transcript context)
• Participants mentioned
• Key topics discussed (bullet points)
• Decisions made
• Action items (with owner if mentioned)

Save it alongside the transcript and show it to the user inline.

Output

• Audio file: /tmp/video-$FILENAME.m4a (or .mp3)
• Transcript files: /tmp/transcripts/$FILENAME.{txt,vtt,srt,json,tsv}
• Summary (if requested): /tmp/transcripts/$FILENAME-summary.md
• Copy transcript/summary to project directory as needed, preserving the $FILENAME timestamp stem — never rename to the meeting date or video title

Usage Examples

# Transcribe a Vimeo video
/video-transcribe https://vimeo.com/123456789

# Transcribe a protected Google Drive recording
/video-transcribe https://drive.google.com/file/d/1GJqYdf.../view

# Specify language (default: en)
/video-transcribe https://vimeo.com/123456789 /tmp/transcripts uk

# Transcribe and summarize
/video-transcribe https://drive.google.com/file/d/1GJqYdf.../view /tmp/transcripts uk summarize