抖音视频转报告

抖音视频链接 → 完整图文报告。自动完成：绕过验证抓取视频 → 下载 → 音频提取 → 语音转文字 → 炫酷HTML报告 → 飞书发送。触发词：抖音视频总结、抖音视频分析、这个视频说了什么、总结这个视频

Security Scan

The skill largely matches its stated purpose (fetch Douyin videos, extract audio, transcribe, and analyze frames) but contains mismatches and missing pieces (promised HTML/飞书 send not implemented), undeclared/unknown external tooling (miaoda-* CLI) that may perform network activity, and use of shell execution patterns that risk command injection — review before installing or running on sensitive systems.

Purpose & Capability

Name/description match the included code: Playwright capture, ffmpeg extraction, frame analysis and transcription. However the SKILL.md and description promise extra capabilities (automatic HTML report generation, 飞书 attachment sending) that are not implemented in the provided douyin_pipeline.py (script stops after saving transcript/analysis). The README also lists tools (yt-dlp, miaoda-coding) that the code does not call. These inconsistencies mean the package does not fully implement or document its claimed end-to-end behavior.

Instruction Scope

Runtime instructions and code perform web scraping with Playwright (including anti-detection flags to '绕过 JS 验证'), download remote video URLs, write files under /tmp and /home/gem/workspace/agent/workspace, call external CLIs (miaoda-studio-cli) for transcription and image analysis. The SKILL.md explicitly instructs storing files in specific workspace paths and indicates automatic sending to 飞书 (not implemented). The instructions give broad discretion for bypassing site protections which is scope-creep relative to a simple 'summarize' task and could be abused.

Install Mechanism

There is no install spec (instruction-only skill plus a script). That minimizes supply-chain risk from package downloads. The SKILL.md lists dependencies and install commands for Playwright/ffmpeg/yt-dlp, which are reasonable for the task.

Credentials

The skill declares no required environment variables or credentials, but invokes external tooling (miaoda-studio-cli, miaoda-coding) of unknown provenance which likely perform network I/O and may require credentials or leak data; these are not declared. The code writes into a hard-coded workspace path (/home/gem/...), which may be harmless in a dedicated agent environment but could expose data if run on a shared host. Also the script uses subprocess.run(shell=True) with f-strings that include remote-derived data (video URLs), creating a command-injection risk if inputs are maliciously crafted.

Persistence & Privilege

The skill is not always-enabled and does not modify other skills or global agent configuration. It runs on invocation only. No explicit persistence or elevated privileges are requested in the manifest or code.

What to consider before installing

This skill largely does what it claims (scrape Douyin → extract audio → transcribe → analyze frames), but it has several red flags you should consider before installing or running it: 1) The README promises HTML report generation and automatic Feishu sending, but the included script does not implement those steps — ask the author for the missing implementation or remove any claims you don't want. 2) The script calls unknown external CLIs (miaoda-studio-cli / miaoda-coding); these could perform network calls or require credentials that the skill doesn't declare — verify their provenance and what data they send. 3) The code uses subprocess.run(..., shell=True) with interpolated inputs (URLs/paths) — this is a command-injection risk if inputs are attacker-controlled; run only with trusted inputs or after sanitizing. 4) The script uses Playwright with anti-detection flags to bypass site protections — be mindful of legal/ethical implications of scraping and of running such code in production. Recommended actions: run the script in an isolated sandboxed environment with restricted network egress, audit or replace the miaoda-* tooling, ensure inputs are validated, and ask the publisher to reconcile the documented features with the actual code (especially Feishu sending and report generation).

Douyin Video → Report Pipeline

将抖音视频链接自动转换为可阅读的 HTML 报告，适合需要提炼视频内容的场景。

完整工作流程

抖音短链 (v.douyin.com/xxx)
    │
    ▼
1. Playwright (headless Chrome)
   - 绕过 JS 验证 / 滑动验证码
   - 提取视频直链
    │
    ▼
2. curl 下载视频 (.mp4)
    │
    ▼
3. ffmpeg 提取音频 (.mp3)
    │
    ▼
4. miaoda-studio-cli speech-to-text
   - 中文普通话识别
   - 100% 转录
    │
    ▼
5. 生成炫酷 HTML 报告
   - 深色渐变主题
   - 核心数据卡片
   - 洞察总结
   - 完整转录（可展开）
   - Ctrl+P 可打印成 PDF
    │
    ▼
6. 飞书附件发送给用户

核心脚本

脚本位于：/home/gem/workspace/agent/skills/douyin-video-to-report/douyin_pipeline.py

使用方式

# 方式一：直接运行（传入抖音URL）
python3 /home/gem/workspace/agent/skills/douyin-video-to-report/douyin_pipeline.py "https://v.douyin.com/xxxxx"

# 方式二：分步执行（见下方分步说明）

分步执行命令

Step 1：提取视频直链

# douyin_fetch.py - Playwright 提取视频 URL
import asyncio
from playwright.async_api import async_playwright

async def get_video_url(douyin_url):
    async with async_playwright() as p:
        browser = await p.chromium.launch(headless=True, args=[
            '--no-sandbox', '--disable-setuid-sandbox',
            '--disable-blink-features=AutomationControlled',
        ])
        context = await browser.new_context(
            user_agent='Mozilla/5.0 (iPhone; CPU iPhone OS 17_0 like Mac OS X) AppleWebKit/605.1.15 (KHTML, like Gecko) Version/17.0 Mobile/15E148 Safari/604.1',
            viewport={'width': 390, 'height': 844},
            locale='zh-CN',
        )
        page = await context.new_page()
        await page.goto(douyin_url, timeout=30000, wait_until="networkidle")
        await asyncio.sleep(3)
        # 从 video 元素获取直链
        video_info = await page.evaluate("""
            () => {
                const v = document.querySelector('video');
                if (v && v.src) return v.src;
                return null;
            }
        """)
        await browser.close()
        return video_info

Step 2：下载视频

curl -L -o /tmp/douyin_video.mp4 "视频直链" \
  -H "User-Agent: Mozilla/5.0 (iPhone; CPU iPhone OS 17_0 like Mac OS X) AppleWebKit/605.1.15 (KHTML, like Gecko) Version/17.0 Mobile/15E148 Safari/604.1" \
  --max-time 60 -s

Step 3：提取音频

ffmpeg -i /tmp/douyin_video.mp4 -vn -acodec libmp3lame -q:a 2 /tmp/douyin_audio.mp3 -y

Step 4：语音转文字

miaoda-studio-cli speech-to-text --file /tmp/douyin_audio.mp3 --lang zh --output text

Step 5：生成报告

使用 miaoda_coding 或直接使用 report_generator.py（见下方）

依赖环境

依赖	安装命令
Playwright	pip install playwright && python3 -m playwright install chromium
ffmpeg	系统自带或 apt install ffmpeg
yt-dlp	pip install yt-dlp
miaoda-studio-cli	系统自带

关键经验

1. 必须用 Playwright headless Chrome，不能用 requests/httpx，抖音会拦截
2. 移动端 UA 更容易绕过，用 iPhone UA 而非桌面 Chrome
3. 视频直链从 document.querySelector('video').src 获取，不要从 network 请求里找
4. 下载视频时必须加 User-Agent header，否则 403
5. 飞书附件发送：文件必须放在 /home/gem/workspace/agent/workspace/ 或 /tmp/openclaw/ 目录

报告模板结构

• Hero：标题 + 视频元信息
• 数据卡片：关键数字高亮
• 洞察区：6个核心发现卡片
• 对比表：进门 vs 留存因素
• 案例区：引用高亮
• 行动指南：5项建议
• 完整转录：可展开/收起

相关工具

• miaoda-studio-cli speech-to-text：音频转文字
• miaoda-studio-cli doc-parse：文档解析
• miaoda-coding：生成可视化报告