Douyin Content Capture

抖音内容抓取

从抖音分享链接或整段分享文案，本地完成解析、无水印下载与文案提取（视频 Whisper 转写，图文直接取配文）。无需登录 Cookie，无需云端语音 API。

Agent rules

1. Never parse Douyin pages yourself. Do not fetch CDN URLs or parse _ROUTER_DATA manually. Always run scripts/capture.py.
2. Run doctor when the environment is unknown before extract. Guide install per INSTALL.md if dependencies are missing. Video downloads fall back to yt-dlp when Douyin CDN rejects direct requests.
3. Use --json and read stdout JSON only. Do not paste raw HTML or CDN URLs to the user unless they explicitly ask for the download link.
4. Default output directory: ~/Downloads/douyin-capture (do not write into the user's project workspace unless they ask).
5. Progress handling: surface CLI stderr progress lines as-is. They now use percentage format like 10%, 35%, 70%, 100%.
6. Final user response format is strict: after success, output only:
   • transcript_preview
   • 浏览器访问: use entrypoints.browser_html
   • 打开文件目录: use entrypoints.open_folder Do not add summary, interpretation, guessing, quality comments, or extra explanation unless the user explicitly asks.
7. Directory naming: output folders must stay short and stable. Use the generated clean slug; do not reconstruct the path from title text in the response.
8. HTML delivery: every successful extract now writes index.html into the output directory. Report that path as one of the artifacts.
9. Progressive reference: link formats → URLS.md; output schema → OUTPUT.md; failures → TROUBLESHOOTING.md.

Decision flow

flowchart TD
    userInput[UserShareLinkOrText] --> extractUrl[AgentExtractsDouyinURL]
    extractUrl --> doctor{EnvReady?}
    doctor -->|no| runDoctor["capture.py doctor --json"]
    runDoctor --> installDeps[GuideInstallPerINSTALL.md]
    doctor -->|yes| intent{UserIntent}
    intent -->|metadata_only| info["capture.py info URL --json"]
    intent -->|download_no_asr| extractSkip["capture.py extract URL --skip-transcribe --json"]
    intent -->|full_transcript| extractFull["capture.py extract URL --json"]
    info --> report[SummarizeJSONToUser]
    extractSkip --> report
    extractFull --> report

Quick start

SKILL_ROOT="<path-to>/douyin-content-capture"
cd "$SKILL_ROOT/scripts"

# First-time setup
python3 -m venv .venv && source .venv/bin/activate
python -m pip install -r requirements.txt

# Check environment
python capture.py doctor --json

# Full extract self-bootstraps `scripts/.venv` when transcription deps are missing.

# Metadata only
python capture.py info "https://v.douyin.com/xxxxx/" --json

# Full extract (download + transcribe for video)
python capture.py extract "https://v.douyin.com/xxxxx/" --json

# Download video only, skip Whisper
python capture.py extract "https://v.douyin.com/xxxxx/" --skip-transcribe --json

Replace SKILL_ROOT with the directory containing this SKILL.md. The wrapper script installs and calls the bundled python-package/ tool inside the same plugin.

CLI reference

extract options

JSON contracts

doctor --json

{
  "ok": true,
  "python": { "ok": true, "version": "3.12.0", "executable": "..." },
  "ffmpeg": { "ok": true, "path": "/usr/bin/ffmpeg" },
  "packages": { "requests": { "ok": true }, "faster_whisper": { "ok": true }, "zhconv": { "ok": true } },
  "errors": [],
  "scripts_dir": "..."
}

ok: true means Python 3.10+ and requests are available (enough for info). errors lists everything still missing for full video transcription.

info URL --json

{
  "ok": true,
  "aweme_id": "7640082170305090171",
  "title": "...",
  "author": "...",
  "content_type": "video",
  "download_url": "https://...",
  "cover_url": "https://...",
  "image_urls": []
}

content_type is video or image. Image notes have image_urls instead of download_url.

extract URL --json

{
  "ok": true,
  "output_root": "/Users/name/Downloads/douyin-capture",
  "out_dir": "/Users/name/Downloads/douyin-capture/7640082170305090171-clean-slug",
  "out_dir_name": "7640082170305090171-clean-slug",
  "aweme_id": "...",
  "title": "...",
  "author": "...",
  "content_type": "video",
  "transcript": "简体文案正文",
  "transcript_preview": "固定长度截断预览",
  "transcript_preview_max_chars": 200,
  "files": { "html": "index.html", "video": "video.mp4", "transcript": "transcript.md", "meta": "meta.json" },
  "artifacts": { "html": "/Users/name/.../index.html", "video": "/Users/name/.../video.mp4", "transcript": "/Users/name/.../transcript.md", "out_dir": "/Users/name/.../" },
  "entrypoints": { "browser_html": "/Users/name/.../index.html", "open_folder": "/Users/name/.../" },
  "skip_transcribe": false
}

On failure, any command returns { "ok": false, "error": "..." } with exit code 1.

Reporting to the user

After a successful extract, output only:

• transcript_preview
• 浏览器访问
• 打开文件目录

Use entrypoints.browser_html for the browser link and entrypoints.open_folder for the folder link. Do not include title, author, content_type summaries, guesses about the audio, or any explanatory note unless the user explicitly asks for them.

Intent mapping

Limitations

• No OCR for text embedded in images
• No login / private content
• Video transcription is slow on CPU (minutes for long videos with small model)
• For research and personal use; comply with platform terms and local law