Claude kann keine Videos schauen. Mit diesem Prompt schon.

Set up a "watch" skill so you can watch videos for me. Do it step by step and tell me in plain English what you are doing.

1. Check which computer I am on (Mac, Windows or Linux).

2. Install these three free tools if they are missing:
   - yt-dlp (downloads videos and subtitles)
   - ffmpeg (cuts videos into pictures and sound)
   - openai-whisper (turns speech into text, runs on my computer)
   Use these ways, because plain "pip install" is blocked or missing on many computers:
   - Mac: install Homebrew first if it is missing, then: brew install yt-dlp ffmpeg openai-whisper
   - Windows: winget install yt-dlp.yt-dlp Gyan.FFmpeg astral-sh.uv, then: uv tool install openai-whisper --python 3.12
   - Linux: yt-dlp and ffmpeg from the package manager, then install uv and run: uv tool install openai-whisper --python 3.12
   Whisper does not work on the newest Python yet, which is why 3.12 is named.
   If something fails, explain why in one sentence and try the next best way. Check at the end that all three start.

3. Create the file ~/.claude/skills/watch/SKILL.md with this content (if you ask me for permission to write it, I will say yes):

```
---
name: watch
description: Watch a video from a link (YouTube, Instagram, TikTok and more) or a local file. Reads what is said and looks at what is shown, then answers questions about it. Use when the user says "/watch <link>", "watch this video", or pastes a video link and asks about it.
---

# watch

For every video, make a new folder in ~/watch/<short-name>/ and work in it.

## 1. Words
- First try the subtitles, without downloading the video:
  yt-dlp --skip-download --write-subs --write-auto-subs --sub-langs "en,en-orig,de,de-orig" -o "subs" "<link>"
- If no subtitle file appears, download only the sound and transcribe it:
  yt-dlp -x --audio-format mp3 -o "audio.%(ext)s" "<link>"
  whisper audio.mp3 --model small --output_format srt
- If you get "429 Too Many Requests", wait a minute and retry once, then use the whisper route.
- Clean the subtitle file (.vtt or .srt) into transcript.txt, one line per caption, keeping the timestamp in front, and removing repeated lines.

## 2. Pictures
- Download a small version of the video:
  yt-dlp -f "bv*[height<=720]+ba/b[height<=720]/b" -o "video.%(ext)s" "<link>"
- Save a picture every time the scene changes, with its timestamp:
  ffmpeg -i video.* -vf "select='gt(scene,0.3)',showinfo,scale=1280:-1" -fps_mode vfr frames/%04d.jpg 2> frames.log
  (Older ffmpeg that says "Unrecognized option fps_mode"? Use -vsync vfr instead.)
  Read the pts_time values from frames.log and write frames.txt as "0001.jpg  00:01:12".
- Fewer than 8 pictures (a static talking video)? Instead take evenly spaced pictures: one every 3 seconds for videos under 2 minutes, one every 10 seconds for longer ones:
  ffmpeg -i video.* -vf "fps=1/3,scale=1280:-1" frames/%04d.jpg   (use fps=1/10 for long videos)
- More than 60 pictures? Keep 60, spread evenly across the video.

## 3. Watch
- Read transcript.txt.
- Open every picture in frames/ and look at it. Read any text, code, slides or charts on screen.
- Match words and pictures by timestamp.
- Then answer the user's question. If there is no question, give a short summary: what the video is about, the key steps or ideas with timestamps, and anything that was only shown on screen and never said out loud.
```

4. Test it on a short public YouTube video of your choice and show me the summary.

5. Tell me in one sentence how to use it from now on: type /watch and paste a link.