douyin-video-summary

SKILL.md

Douyin Video Summary

Summarize Douyin videos: extract audio → transcribe locally → AI summary.

Prerequisites

Install these tools (macOS example):

brew install whisper-cpp ffmpeg
# Download whisper.cpp GGML model (small recommended for speed/quality balance)
curl -L -o models/ggml-small.bin "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.bin"

Workflow

When a Douyin link is received:

Step 1: Extract Video ID
Parse the Douyin URL to get the video ID. Douyin share links come in two formats:

Short link: https://v.douyin.com/xxxxx/ → follow redirect to get video ID

Direct link: https://www.douyin.com/video/7604713801732365681

# Follow redirect to get final URL, extract numeric video ID
curl -sL -o /dev/null -w '%{url_effective}' 'https://v.douyin.com/xxxxx/' | grep -oE '[0-9]{15,}'

Step 2: Get Audio via Browser

Douyin blocks direct downloads (yt-dlp, aria2c all get 403). Use the browser to intercept the audio URL:

Open the Douyin video page in the browser

Inject JS to intercept network requests before navigation:

window.__audioUrls = [];
const origOpen = XMLHttpRequest.prototype.open;
XMLHttpRequest.prototype.open = function(method, url) {
  if (url && (url.includes('.mp3') || url.includes('.m4a') || url.includes('mime_type=audio'))) {
    window.__audioUrls.push(url);
  }
  return origOpen.apply(this, arguments);
};

Navigate to the video page, click play to trigger audio loading

Retrieve intercepted URLs: window.__audioUrls

Download with curl (Referer header required):

curl -H "Referer: https://www.douyin.com/" -o audio.mp4 "<audio_url>"

Important: aria2c will 403 on Douyin CDN URLs. Always use curl with the Referer header.

Step 3: Convert to WAV

ffmpeg -i audio.mp4 -ar 16000 -ac 1 -c:a pcm_s16le audio.wav

Step 4: Transcribe with whisper.cpp

whisper-cli -m /path/to/ggml-small.bin -l zh -f audio.wav -otxt -of output

Use -l zh for Chinese content (auto-detect if unsure)

Apple Silicon GPU acceleration is automatic (Metal)

Performance: ~20s for 5min audio on M4

Step 5: Generate Summary

Read the transcription text and produce a structured summary:

📹 **[Video Title] | [Author]**
时长：X分X秒 | 发布：YYYY-MM-DD

🎯 **核心观点：[one-line core message]**

**1. [Point 1 title]**
• [detail]
• [detail]

**2. [Point 2 title]**
• [detail]

💬 **一句话总结：[concise takeaway]**

Step 6 (Optional): Sync to Feishu Doc

If Feishu integration is configured, append the summary to a Feishu document using the Feishu Open API. See references/feishu-sync.md for the API details.

Tips

For short videos (<1min), the summary may be very brief — that's fine

If browser interception fails, retry once; Douyin pages sometimes need a second load

Clean up downloaded audio/wav files after processing to save disk space

whisper.cpp small model is the best speed/quality tradeoff; medium may OOM on 8GB machines