watch-anything · 看不完的视频，让 AI 替你看懂

丢一个 X / YouTube / B站 视频链接进来，几分钟后拿到一张一眼看完的中文卡片 （摘要 + 要点 + 金句中英对照）+ 完整解读 + 全文中译，还能就着它追问。

Drop in an X / YouTube / Bilibili video link, and minutes later get a glanceable card (summary + key points + bilingual highlights) + a full breakdown + a full Chinese translation, all of which you can then ask questions about.

它是 read-anything 的姊妹： read 管读文字，watch 管看视频。

⚠️ About the GIF above: that browser side panel is my own knowledge workbench, and it is not open source. What's open source is watch-anything, a skill — you drop a link in Claude Code / Codex, it fetches, transcribes and produces a Chinese card; the output is markdown (see the real example under "How to use"). So what you get by starring: fetching + caption-less transcription + Chinese card breakdown + follow-up Q&A + optional Feishu push. What you don't get: this side-panel UI, the inspiration library. The GIF shows: an X tweet with a 28-minute video (Anthropic's official prompt workshop) → audio downloaded and transcribed → Chinese card.

English · 简体中文

---

What is this

Your bookmarks are probably full of videos you "meant to watch someday" — a forty-minute English tech talk, a tweet with a video, a podcast conversation, a caption-less long video on Bilibili... and they just sit there collecting dust, too long to finish, too much to get through.

watch-anything is a Claude Code / Agent skill: drop in a video link, it first fetches the content, transcribes the audio track into text (if there are no captions), then has your AI produce a Chinese card breakdown, with the output landing as markdown in your own knowledge base. The whole thing happens inside your agent — no switching back and forth between a dozen web tools.

The specific pain it solves

• Bookmarked it but no time to watch → in minutes you get a card that lets you "get it without watching the original"
• English / caption-less videos, where other summary tools just give up → auto-downloads the audio track and transcribes, then produces a Chinese breakdown
• Want to dig into details → chat over the transcript, answers grounded only in the source, no fabrication
• Watch and forget, scattered everywhere → the breakdown lands as markdown, accumulating into your own system

The killer feature: understand even without captions

Most "video summary" tools rely on the platform's existing captions, and choke on videos without them. The core of watch-anything is a complete transcription fallback chain, where each level automatically drops down to the next on failure:

Video captions (fast, accurate)  →  Groq cloud whisper (tens of seconds, nearly free)  →  local faster-whisper (zero API cost, never leaves your machine)

If there are captions, use them; if not, download the audio track and transcribe. With GROQ_API_KEY configured it goes through the cloud; without it, transcribe locally (audio never leaves your machine). This is also its real value relative to read-anything: read currently only grabs X as text, and does not download and transcribe video; watch fills in exactly that gap.

A browser extension? There is one — but only as a trigger