youtube-decoderlisted
Install: claude install-skill umerdarazkha-debug/youtube-decoder
# YouTube Decoder
Claude can't play videos, and pasting a YouTube link alone gives it almost nothing to work with. This skill reads
the video's captions instead: every spoken word, with timestamps. It cannot see the visuals, so say so when the
answer depends on what is shown rather than said.
## 1. Get the transcript
Run the bundled script with the link the user gave (captions only, the video itself is never downloaded):
```bash
python3 scripts/fetch_transcript.py "<youtube url>" --out transcript.md
```
- It prefers human-made captions and falls back to YouTube's auto-captions (de-duplicated).
- It needs `yt-dlp` (`pip install yt-dlp`) or `uv` (it then runs `uvx yt-dlp`). If neither exists, install yt-dlp with the user's permission.
- Other languages: `--lang es`, `--lang de`, ...
- If it reports no captions, tell the user plainly: the creator turned captions off, so there is nothing to read.
Read `transcript.md` in full before answering. Long podcasts are fine; read them in chunks.
## 2. The breakdown
Reply in this order, short and scannable. Every claim must point at a timestamp from the transcript, never at a guess.
1. **The hook**: the line that earns the next 30 seconds, with its timestamp, and why it works (curiosity, stakes, contrast, promise).
2. **The structure**: the 3-7 beats the video actually follows, each as `[m:ss] beat name: one line`. Name the pattern if it has one (e.g. "three stories", problem -> agitation -> solution, list, before/after).
3. **Mom