Communitygithub.com

heyuxuan0209/watch-anything

Watch endless videos and let AI understand them for you: X/YouTube/Bilibili videos → Chinese card interpretation (sister skill of read-anything)

What is watch-anything?

This skill (watch-anything) enables AI agents to automatically understand and summarize videos from platforms like X (Twitter), YouTube, and Bilibili (B站). Given a video link, the agent reads a central PLAYBOOK.md as its behavioral specification—defining routing rules, transcription fallback chains, card output format, honesty guidelines, and downgrade scripts. For X tweets, it first fetches text and media via `scripts/fetch-x.mjs`; if a video is found, it transcribes via `scripts/transcribe-video.mjs`. YouTube and Bilibili links go directly to the transcription script, which prefers subtitles and falls back to Groq cloud whisper (needs `GROQ_API_KEY`) or local faster-whisper. The skill outputs formatted cards using templates (`templates/card.md` by default), and can optionally push results to Feishu (Lark) via webhook or bot if environment variables (`FEISHU_WEBHOOK` / `FEISHU_APP_ID`) are configured. All scripts run on Node 18+ with zero npm dependencies, and a strict honesty rule forbids hallucinating content beyond what was actually transcribed. This skill is designed for Claude Code and similar agentic development tools.

Works with✓Claude Code~Codex CLI~Cursor
npx skills add heyuxuan0209/watch-anything

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

看懂视频(watch-anything)

按本目录 PLAYBOOK.md 执行,它是唯一的行为规范(路由表、转写降级链、卡片格式、 诚实守则、降级话术都在里面)。本文件只是 Claude Code 的触发入口。

要点提醒:

  1. 先读 PLAYBOOK.md,按链接类型走路由:X 推文先 scripts/fetch-x.mjs 拿正文+媒体清单, 带视频再 scripts/transcribe-video.mjs 转写;YouTube/B站 直接 transcribe-video.mjs (字幕优先,无字幕才转写)。脚本零 npm 依赖、Node 18+、JSON 进出。
  2. 转写降级链:视频字幕 > Groq 云 whisper(需 GROQ_API_KEY)> 本地 faster-whisper。 缺哪级降哪级,长视频转写前告知用户「需等几分钟」。
  3. 解读模板在 templates/,默认 card.md(卡片格式);用户一句话可覆盖, 自定义模板 = 目录里的新 .md 文件。 输出终点在 integrations/(可选):环境里有 FEISHU_WEBHOOK / FEISHU_APP_ID 时, 卡片出完可问一句要不要推飞书,跑 scripts/push-feishu.mjs;没配就当它不存在,别催用户配。
  4. 诚实守则优先级最高:抓不到的推文、转不出的音频、未解析的画面,如实降级并显式声明, 绝不脑补视频/图片里的内容。

Related Skills