YouTube Transcribe Skill

Tool Discovery

Before running, the agent checks for available tools and sets PATH:

# Find tools dynamically — don't hardcode paths
export PATH="/tmp/miniforge/bin:$(python3 -m site --user-base)/bin:$PATH"

YTDLP=$(command -v yt-dlp 2>/dev/null || echo "yt-dlp")
FFMPEG=$(command -v ffmpeg 2>/dev/null || echo "ffmpeg")
WHISPER=$(command -v whisper 2>/dev/null || echo "whisper")

# Verify tools exist
for TOOL in "$YTDLP" "$FFMPEG" "$WHISPER"; do
  [ -x "$TOOL" ] || echo "[WARN] Tool not found or not executable: $TOOL"
done

Tool requirements:

Environment PATH

export PATH="/tmp/miniforge/bin:$(python3 -m site --user-base)/bin:$PATH"

Workflow

Step 1 — Parse YouTube URL

URL="https://www.youtube.com/watch?v=Q5kYrmzNhcU"
VIDEO_ID=$(echo "$URL" | grep -oE 'v=[^&]+' | cut -d= -f2 | head -1)
# Handles: https://youtu.be/ID, https://www.youtube.com/watch?v=ID&t=..., https://youtube.com/embed/ID

Step 2 — Get Video Metadata

TITLE=$($YTDLP --extractor-args "youtube:player_client=android" \
  --print title --no-warnings "https://www.youtube.com/watch?v=${VIDEO_ID}" 2>/dev/null)
CHANNEL=$($YTDLP --extractor-args "youtube:player_client=android" \
  --print channel --no-warnings "https://www.youtube.com/watch?v=${VIDEO_ID}" 2>/dev/null)
DURATION=$($YTDLP --extractor-args "youtube:player_client=android" \
  --print duration_string --no-warnings "https://www.youtube.com/watch?v=${VIDEO_ID}" 2>/dev/null)

Step 3 — Download Audio (with Fallback Chain)

mkdir -p /tmp/yt_audio

# Strategy: try android client first → if GVS PO Token error, fall back to web client
# (web client may 403 on some videos; android client needs PO token for high-quality formats
# but usually succeeds with format 18 even without PO token)

# Attempt 1: android client (works without PO token for format 18)
$YTDLP -x --audio-format mp3 --audio-quality 0 \
  --extractor-args "youtube:player_client=android" \
  -f "best[ext=mp4]/best" \
  -o "/tmp/yt_audio/${VIDEO_ID}.%(ext)s" \
  "https://www.youtube.com/watch?v=${VIDEO_ID}" 2>&1 | grep -v "^Deprecated\|^NotOpenSSL\|^Warning:"

# If android fails (GVS PO Token required), fall back to web
if [ ! -f "/tmp/yt_audio/${VIDEO_ID}.mp4" ] && [ ! -f "/tmp/yt_audio/${VIDEO_ID}.mp3" ]; then
  echo "[*] Android client failed, trying web client..."
  $YTDLP -x --audio-format mp3 --audio-quality 0 \
    -o "/tmp/yt_audio/${VIDEO_ID}.%(ext)s" \
    "https://www.youtube.com/watch?v=${VIDEO_ID}" 2>&1 | grep -v "^Deprecated\|^NotOpenSSL"
fi

Why --extractor-args "youtube:player_client=android": Web client returns 403 for many videos; android client returns format 18 (mp4, ~480p) without requiring a GVS PO Token, which is sufficient for transcription.

Step 4 — Convert to MP3 (if needed)

# If yt-dlp downloaded .mp4 instead of .mp3
if [ -f "/tmp/yt_audio/${VIDEO_ID}.mp4" ]; then
  $FFMPEG -i "/tmp/yt_audio/${VIDEO_ID}.mp4" \
    -vn -acodec libmp3lame -q:a 2 \
    "/tmp/yt_audio/${VIDEO_ID}.mp3" -y 2>/dev/null
  rm -f "/tmp/yt_audio/${VIDEO_ID}.mp4"
fi

Step 5 — Transcribe

$WHISPER "/tmp/yt_audio/${VIDEO_ID}.mp3" \
  --model tiny \
  --language en \
  --output_dir /tmp/yt_audio \
  --output_format txt 2>&1 | grep -v "^Deprecated\|^UserWarning"

# Whisper outputs to {output_dir}/{filename}.txt
# Rename if needed
[ -f "/tmp/yt_audio/${VIDEO_ID}.txt" ] && \
  mv "/tmp/yt_audio/${VIDEO_ID}.txt" "/tmp/yt_audio/${VIDEO_ID}_transcript.txt"

Model choice: tiny is fastest for English. Use base or small for better accuracy if time permits.

Step 6 — Save to Memory

Append to memory/YYYY-MM-DD.md:

## YouTube 转录: <Video Title>

- **URL**: https://www.youtube.com/watch?v=<video_id>
- **频道**: <channel_name>
- **时长**: <duration>
- **日期**: YYYY-MM-DD

### 摘要
<3-5 sentence summary>

### 关键引用
> "<notable quote>"

### 核心洞察
<1-3 insights>

Step 7 — Post to Feishu (optional)

If user requests it, send a Feishu message with the summary and key quotes.

Error Handling

Cleanup

rm -f /tmp/yt_audio/${VIDEO_ID}.*

When NOT to Use This Skill

• Video has accurate YouTube captions → Use web_fetch with transcript extraction instead (faster, more accurate, preserves speaker labels)
• User only wants a summary → Ask if full transcript is needed before running (5+ min transcription vs instant captions)
• Video is very long (>30 min) → Whisper inference takes significant time on CPU; warn user before starting
• Non-English video → Specify language with --language <code> (e.g., --language zh for Chinese); tiny model quality degrades significantly for non-English

One-Time Installation

# yt-dlp
pip3 install yt-dlp

# Miniforge (ffmpeg + whisper dependencies)
curl -sL "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-MacOSX-arm64.sh" -o /tmp/miniforge.sh
chmod +x /tmp/miniforge.sh
/bin/bash /tmp/miniforge.sh -b -p /tmp/miniforge
/tmp/miniforge/bin/conda install -y ffmpeg -c conda-forge
/tmp/miniforge/bin/pip install openai-whisper

# Add to ~/.zshrc
echo 'export PATH="/tmp/miniforge/bin:$(python3 -m site --user-base)/bin:$PATH"' >> ~/.zshrc