Agent-Reach/skills/video/skill.md

# Video & Podcast Digest Skill

> Send a video/podcast link → get full transcript + structured summary

## Supported Platforms

| Platform | Type | Subtitles | Whisper Transcription |
|----------|------|-----------|----------------------|
| YouTube | Video | ✅ | ✅ |
| Bilibili | Video | ✅ | ✅ |
| X/Twitter | Video | ❌ | ✅ |
| Xiaoyuzhou (小宇宙) | Podcast | ❌ | ✅ |
| Apple Podcasts | Podcast | ❌ | ✅ |
| Direct links (mp3/mp4/m3u8) | Any | ❌ | ✅ |

## Trigger

Auto-triggered when a media URL is detected:
- YouTube: `youtube.com`, `youtu.be`
- Bilibili: `bilibili.com`, `b23.tv`
- X/Twitter: `x.com`, `twitter.com` (tweets with video)
- Xiaoyuzhou: `xiaoyuzhoufm.com`
- Apple Podcasts: `podcasts.apple.com`
- Direct: `.mp3`, `.mp4`, `.m3u8`, `.m4a`, `.webm`

## Pipeline

### Step 0: Detect Media Type

| URL Pattern | Type | Pipeline |
|-------------|------|----------|
| `xiaoyuzhoufm.com/episode/` | Podcast | → Step 1b (Xiaoyuzhou) |
| `podcasts.apple.com` | Podcast | → Step 1c (Apple) |
| `bilibili.com`, `b23.tv` | Video | → Step 1d (Bilibili API) |
| `.mp3`, `.m4a` direct link | Audio | → Step 2b (direct download) |
| Other | Video | → Step 1a (subtitle extraction) |

### Step 1a: Video — Extract Subtitles

```bash
# Clean up temp files
rm -f /tmp/media_sub*.vtt /tmp/media_audio.mp3 /tmp/media_transcript*.json /tmp/media_segment_*.mp3 2>/dev/null || true

# YouTube (prefer English, fallback Chinese)
yt-dlp --skip-download --write-auto-sub --sub-lang "en,zh-Hans" -o "/tmp/media_sub" "VIDEO_URL"

# Bilibili
yt-dlp --skip-download --write-auto-sub --sub-lang "zh-Hans,zh" -o "/tmp/media_sub" "VIDEO_URL"
```

Check for subtitles:
```bash
ls /tmp/media_sub*.vtt 2>/dev/null
```
- **Has subtitles** → Read VTT content, skip to Step 3
- **No subtitles** → Step 2a (download audio)

### Step 1b: Xiaoyuzhou (小宇宙) — Extract Audio URL

```bash
# Extract CDN direct link from __NEXT_DATA__
# Xiaoyuzhou is a Next.js SPA, but initial HTML contains audio URL in __NEXT_DATA__
AUDIO_URL=$(curl -sL -H "User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36" \
  "EPISODE_URL" \
  | grep -oE 'https://media\.xyzcdn\.net/[^"]+\.(m4a|mp3)' \
  | head -1)

echo "Audio URL: $AUDIO_URL"

# Download audio
curl -L -o /tmp/media_audio.mp3 "$AUDIO_URL"
```

> If curl extraction is empty (rare), fallback: use Puppeteer/browser to get rendered page and extract.

→ Step 2b (check size & transcribe)

### Step 1c: Apple Podcasts — via yt-dlp

```bash
yt-dlp -f "ba[ext=m4a]/ba/b" --extract-audio --audio-format mp3 --audio-quality 5 \
  -o "/tmp/media_audio.%(ext)s" "APPLE_PODCAST_URL"
```

→ Step 2b (check size & transcribe)

### Step 1d: Bilibili — API Direct Audio Stream

yt-dlp returns 412 for Bilibili even with cookies. Use Bilibili's API instead:

```bash
# 1. Extract BV number from URL
BV="BV1xxxxx"  # Replace with actual BV number

# 2. Get video info (title, duration, CID)
curl -s "https://api.bilibili.com/x/web-interface/view?bvid=$BV" \
  -H "User-Agent: Mozilla/5.0" -H "Referer: https://www.bilibili.com/" \
  | python3 -c "import json,sys; d=json.load(sys.stdin)['data']; print(f\"Title: {d['title']}\nDuration: {d['duration']}s\nCID: {d['cid']}\")"

# 3. Get audio stream URL
CID=<CID from previous step>
AUDIO_URL=$(curl -s "https://api.bilibili.com/x/player/playurl?bvid=$BV&cid=$CID&fnval=16&qn=64" \
  -H "User-Agent: Mozilla/5.0" -H "Referer: https://www.bilibili.com/" \
  | python3 -c "import json,sys; print(json.load(sys.stdin)['data']['dash']['audio'][0]['baseUrl'])")

# 4. Download audio (Referer header required, otherwise 403)
curl -L -o /tmp/media_audio.m4s \
  -H "User-Agent: Mozilla/5.0" -H "Referer: https://www.bilibili.com/" "$AUDIO_URL"

# 5. Convert to mp3
ffmpeg -y -i /tmp/media_audio.m4s -acodec libmp3lame -q:a 5 /tmp/media_audio.mp3
```

→ Step 2b (check size & transcribe)

### Step 2a: Video — Download Audio (when no subtitles)

```bash
# YouTube may need --cookies-from-browser chrome to bypass bot detection
yt-dlp --cookies-from-browser chrome -f "ba[ext=m4a]/ba/b" --extract-audio --audio-format mp3 --audio-quality 5 \
  -o "/tmp/media_audio.%(ext)s" "VIDEO_URL"
```

### Step 2b: Check Audio Size & Segment

```bash
FILE_SIZE=$(stat -f%z /tmp/media_audio.* 2>/dev/null || stat -c%s /tmp/media_audio.* 2>/dev/null)
echo "File size: $FILE_SIZE bytes"
```

- **≤ 25MB (25000000)** → Step 2c (transcribe directly)
- **> 25MB** → Split first, then transcribe each segment

**Splitting large audio (>25MB)**:
```bash
# Get total duration
DURATION=$(ffprobe -v error -show_entries format=duration -of csv=p=0 /tmp/media_audio.* | head -1)

# Split into 10-minute segments (keeps each under 25MB)
SEGMENT_SEC=600
SEGMENTS=$(python3 -c "import math; print(math.ceil(float('$DURATION')/$SEGMENT_SEC))")

# Cut segments
for i in $(seq 0 $((SEGMENTS-1))); do
  START=$((i * SEGMENT_SEC))
  ffmpeg -y -i /tmp/media_audio.* -ss $START -t $SEGMENT_SEC -acodec libmp3lame -q:a 5 \
    "/tmp/media_segment_${i}.mp3" 2>/dev/null
done
```

→ Call Step 2c for each segment **sequentially** (parallel triggers Groq 524 timeout), concatenate results

### Step 2c: Whisper Transcription

**Prerequisite**: `GROQ_API_KEY` environment variable

```bash
# Check API key
if [ -z "$GROQ_API_KEY" ]; then
  echo "❌ GROQ_API_KEY not set. Get one at: https://console.groq.com/keys"
  exit 1
fi

# Transcribe single file (replace AUDIO_FILE with actual path)
curl -s -X POST "https://api.groq.com/openai/v1/audio/transcriptions" \
  -H "Authorization: Bearer $GROQ_API_KEY" \
  -H "Content-Type: multipart/form-data" \
  -F "file=@AUDIO_FILE" \
  -F "model=whisper-large-v3-turbo" \
  -F "response_format=verbose_json" \
  -F "language=zh" \
  > /tmp/media_transcript.json

# Extract plain text
python3 -c "import json; print(json.load(open('/tmp/media_transcript.json'))['text'])"
```

**Whisper model options**:
| Model | Speed | Accuracy | Use Case |
|-------|-------|----------|----------|
| `whisper-large-v3-turbo` | 10x realtime | High | Default choice |
| `whisper-large-v3` | 5x realtime | Highest | Professional/noisy content |

**Language parameter**:
- Chinese: `language=zh`
- English: `language=en`
- Auto-detect: omit language parameter

### Step 3: Structured Summary

Choose output format based on media type:

**Video (≤20 min)**:
1. **Overview** (1-2 sentences)
2. **Key Points** (3-5 bullet points)
3. **Notable Quotes** (if any)
4. **Action Items** (if applicable)

**Podcast (>20 min)**:
1. **Overview** (2-3 sentences: who discussed what)
2. **Chapter Summary** (segmented by topic, 2-3 sentences each)
3. **Key Points** (5-8 bullet points)
4. **Notable Quotes**
5. **Action Items** (if applicable)

## Output Format

### Video

```
## 📺 Video Digest

**Title**: [Video Title]
**Duration**: [x minutes]
**Language**: [Chinese/English]

### Overview
[1-2 sentence summary]

### Key Points
1. [Point 1]
2. [Point 2]
...

### Notable Quotes
> "xxx" — [timestamp]

### Action Items
- [if applicable]
```

### Podcast

```
## 🎙️ Podcast Digest

**Show**: [Podcast Name]
**Episode**: [Episode Title]
**Duration**: [x minutes]
**Guests**: [if any]

### Overview
[2-3 sentences: who discussed what, core conclusions]

### Chapter Summary
#### 1. [Topic] (~xx:xx-xx:xx)
[2-3 sentences of core content]

#### 2. [Topic] (~xx:xx-xx:xx)
[2-3 sentences of core content]
...

### Key Points
1. [Point 1]
2. [Point 2]
...

### Notable Quotes
> "xxx"

### Action Items
- [if applicable]
```

## Error Handling

| Situation | Action |
|-----------|--------|
| No subtitles + no GROQ_API_KEY | Prompt user to set API key |
| No subtitles + has API key | Auto Whisper transcription |
| Xiaoyuzhou curl extraction empty | Use Puppeteer/browser to get rendered HTML |
| Audio >25MB | ffmpeg segment (10min/segment), transcribe sequentially |
| Podcast >2 hours | Warn user about duration, confirm before proceeding |
| Groq 524 timeout | Do NOT parallelize — transcribe sequentially, sleep 5-8s between segments |
| Groq 429 rate limit | 7200s/hour limit, wait for retry-after header, then retry |
| yt-dlp Bilibili 412 | Use Bilibili API instead (Step 1d) |
| yt-dlp YouTube bot detection | Add `--cookies-from-browser chrome` |
| Network timeout | Retry once |
| Spotify links | Inform user: not supported (DRM protected) |

## Groq Whisper Limits

- Max 25MB per request
- Free tier: 7200 seconds of audio/hour (rolling window), ~20 hours/day
- Supported formats: mp3, mp4, mpeg, mpga, m4a, wav, webm

## Dependencies

- `yt-dlp`: video download + subtitle extraction
- `ffmpeg`: audio conversion + segmentation
- `curl`: Xiaoyuzhou audio download, Bilibili API
- `GROQ_API_KEY`: Whisper transcription API (free at https://console.groq.com/keys)