douyin-to-obsidian

抖音视频文案自动提取工具，一键将抖音视频转为结构化 Obsidian 笔记。支持绕过风控、本地 Whisper 语音识别、长视频分段处理。

Security Scan

The skill's code, dependencies, and runtime instructions are consistent with a Douyin (TikTok China) scraping + local Whisper transcription tool, but it performs anti-bot workarounds and downloads binaries at runtime—review and sandbox before running.

Purpose & Capability

Name/description (extract Douyin video text → Obsidian) align with the included code: Playwright-based page scraping, SSR data parsing, audio download, local Whisper transcription, Markdown output. Features like short-link resolution, modal-based redirection, custom user-agent and navigator modifications are coherent with a scraper aiming to avoid anti-bot protections.

Instruction Scope

SKILL.md and run_extract.py only instruct installing Python deps, running Playwright, and executing the extractor. The extractor code modifies browser fingerprinting properties, follows redirects, writes failed_page.html and other diagnostic files, downloads audio and writes Markdown notes to a filesystem path—all consistent with its purpose. Note: it explicitly includes logic to '绕过风控' (anti-bot/captcha workarounds), which is expected for scraping but has policy/ethical implications outside security scope.

Install Mechanism

Skill is instruction-only (no packaged installer), but the Python code will auto-download and extract an FFmpeg zip at runtime from https://www.gyan.dev. Runtime download-and-extract of an executable is higher-risk than pure Python deps; gyan.dev is a commonly used FFmpeg build host but pulling and extracting binaries at runtime should be reviewed. Playwright also requires downloading browser binaries via the documented install step.

Credentials

No environment variables, secrets, or external credentials are requested. The requirements.txt includes 'openai' which the visible code does not use (the transcription uses local openai-whisper). No credentials or unrelated tokens are requested.

Persistence & Privilege

The skill is not always-enabled and does not request elevated agent privileges. It stores a cached ffmpeg in a per-user path (~/.douyin_skill/ffmpeg on non-Windows or in user home on Windows) and writes output files to a default Obsidian path (hardcoded Windows path by default). It does not modify other skills or global agent configuration.

Assessment

This skill appears to do what it claims (scrape Douyin pages and transcribe audio locally), but there are a few things to consider before installing or running: - Anti-bot behavior: The code contains explicit fingerprinting changes and modal redirect techniques to avoid Douyin's protections. That can violate site terms of service and may be legally or ethically problematic depending on your jurisdiction and use. - Runtime binary download: The extractor will download and extract an FFmpeg ZIP from https://www.gyan.dev at first run. Although gyan.dev is a known FFmpeg build provider, downloading and executing binaries at runtime increases risk—verify the URL and checksum or supply your own ffmpeg binary instead. - Default save location: The script defaults to a hardcoded Windows iCloud/Obsidian path (E:\icloud\...), which may cause files to sync to cloud storage unexpectedly. Change the base_dir in scripts/run_extract.py to a location you control before running. - Dependencies: requirements.txt includes 'openai' but the visible code uses local openai-whisper; consider trimming unused packages. Playwright requires installing browser binaries separately (playwright install chromium). - Sandbox/testing: Run initially in an isolated environment (VM or disposable account/directory) and inspect created files (failed_page.html, failed_ssr_data.json, cache directories). Consider reviewing the remainder of extractor.py (the provided file was truncated in the manifest) to confirm there are no hidden network endpoints or exfiltration logic. - If you intend to use this at scale or for others' content, check legal/terms-of-service implications for scraping and CAPTCHA circumvention. If you accept these tradeoffs, change the save path, optionally provide your own ffmpeg binary, and run in a controlled environment.

🎯 抖音视频文案自动提取工具

一键将抖音视频转为结构化 Obsidian 笔记，支持绕过风控、本地 Whisper 语音识别、长视频分段处理。

核心功能

• ✅ 智能识别 - 支持抖音短链/分享口令自动解析
• ✅ 绕过风控 - 自动绕过抖音滑动验证码和 SSR 数据封锁
• ✅ 本地识别 - 使用 Whisper AI 本地语音识别，零 API 成本
• ✅ 结构化输出 - 自动生成带 YAML Frontmatter 的 Markdown 笔记
• ✅ 长视频支持 - 自动分段处理，支持任意时长视频
• ✅ 自动安装 - FFmpeg 自动下载，无需手动配置

使用方式

快速开始

# 1. 安装依赖
pip install -r scripts/requirements.txt

# 2. 安装浏览器自动化
playwright install chromium

# 3. 运行提取
python scripts/run_extract.py "https://v.douyin.com/xxxxx/"

输入格式

支持多种输入格式：

• 完整 URL: https://v.douyin.com/xxxxx/
• 分享口令：3.58 KJi:/ [e-mail] 06/21 通达信自选股+Python#qmt#量化交易 https://v.douyin.com/xxxxx/
• 短链接：https://v.douyin.com/xxxxx/

输出示例

笔记自动保存到 Obsidian 目录（默认：E:\icloud\iCloudDrive\iCloud~md~obsidian\myobsidian）

---
标题：视频标题
作者：作者名
链接：https://v.douyin.com/xxx
时长：7 分 42 秒
提取时间：2026-03-22 20:15:00
标签：#AI #杨立昆 #AGI
---

# 视频标题

完整的语音识别文案内容...

技术特点

反爬策略

• 精选模态链接绕过风控
• 自定义 User-Agent 和浏览器指纹
• 智能短链解析和重定向处理

语音识别

• OpenAI Whisper base 模型（中文优化）
• 支持长视频自动分段（默认 10 分钟/段）
• 自动标点符号和文本规范化

架构设计

• 基于 asyncio 的异步处理
• FFmpeg 自动下载和管理
• 临时文件自动清理

系统要求

• Python: 3.8+
• 磁盘空间: 首次运行约 200MB（Whisper 模型 + FFmpeg）
• 操作系统: Windows（macOS/Linux 需手动安装 FFmpeg）

依赖项

requests>=2.31.0
aiohttp>=3.9.0
ffmpeg-python>=0.2.0
openai-whisper>=20231117
openai>=1.0.0
playwright>=1.40.0

常见问题

Q: 运行时提示找不到 ffmpeg？

A: 首次运行时会自动下载 FFmpeg（约 100MB），请确保网络连接正常。Windows 用户无需手动安装。

Q: 语音识别很慢？

A: 首次运行需下载 Whisper 模型（base 模型约 150MB），后续会使用缓存。识别速度取决于 CPU 性能。

Q: 如何修改保存路径？

A: 编辑 scripts/run_extract.py 第 50 行的 base_dir 变量。

Q: 提取失败怎么办？

A: 检查网络连接，确保能访问抖音。如仍失败，查看生成的 failed_page.html 和 failed_ssr_data.json 文件。

文件结构

douyin-to-obsidian/
├── SKILL.md                 # 技能描述
├── README.md                # 详细说明
├── scripts/
│   ├── run_extract.py       # 主入口
│   ├── extractor.py         # 核心引擎
│   └── requirements.txt     # 依赖列表
└── reference/
    └── architecture.md      # 架构文档

更新日志

v1.0.0 (2026-03-22)

• 🎉 初始发布
• ✅ 抖音短视频文案提取
• ✅ 本地 Whisper 语音识别
• ✅ FFmpeg 自动下载
• ✅ 结构化 Obsidian 输出

License

MIT-0