youtube2text

A Claude Code skill that turns a YouTube video into a clean transcript and a structured conspectus of its key points.

What it does

• Downloads the video's captions with yt-dlp — manual subtitles if available, otherwise auto-generated.
• Cleans the transcript: removes the rolling-window duplication YouTube auto-captions produce, then joins and re-paragraphs the text.
• Writes the full transcript and a conspectus (theses, not a retelling) side by side.

Why

Auto-captions are noisy and full of repetition. This gives you a readable transcript and a digest you can act on, with deduplication done deterministically by a script rather than guessed at by the model.

Requirements

• yt-dlp: pip install yt-dlp
• Claude Code

Install

Copy this folder into your skills directory:

• Project-level: .claude/skills/youtube2text/
• User-level: ~/.claude/skills/youtube2text/

Then invoke it in Claude Code:

/youtube2text https://www.youtube.com/watch?v=...

What you get

digested/<slug>/
  conspectus.md              ← key points, grouped by theme
  attachments/transcript.md  ← full cleaned transcript

The skill proposes a short slug from the title and, on your confirmation, produces both files.

Output directory

Defaults to ./digested. Point it anywhere that fits your project — tell the skill where, or edit OUT_DIR in SKILL.md.

clean_transcript.py

The cleaning step is a standalone script — usable without Claude:

python clean_transcript.py \
  --srt video.ru.srt \
  --out transcript.md \
  --title "Title" --source "url" --channel "Channel" --duration "12 min"

Run python clean_transcript.py --help for all options.

Limitations

• Quality is bounded by YouTube's auto-captions. Proper names, numbers, and terms may be misrecognized (e.g. "Бафет" instead of "Баффет"). The script cleans structure; it cannot fix recognition errors.
• If a video has no captions at all, the skill cannot transcribe it — it does not run speech-to-text on the audio.

License

MIT