视频摘要器

Overview

Use this skill whenever you need to keep Richard (or any configured subscriber) up to date on new Nate B Jones videos. The workflow:

1. Detect a new upload on https://www.youtube.com/@NateBJones.
2. Retrieve the transcript (official captions first, Whisper fallback if missing).
3. Summarize the video into an abstract, bullet highlights, and a "References & Links" list.
4. Publish according to the installation's config: email, Control UI/Telegram chat, Google Doc, Markdown file, etc.

All runtime options live in references/config-example.yml. Copy that file, rename it (e.g. config.yml), fill in your preferences, and point the workflow to it.

1. Configure

1. Copy references/config-example.yml to config.yml (or any path you prefer).
2. Fill in:
   • channel_url or channel_id (the example already targets @NateBJones).
   • poll_cron (default daily at 09:00 local).
   • outputs.email.to, outputs.chat.targets, outputs.doc.type/path.
   • API credentials: YouTube Data API key (for upload polling), Gmail/Google Docs auth handled via gog skill.
3. Store the config path somewhere easy to reference (e.g. skills/nate-b-jones-digest/config.yml).

2. Poll for new videos

• Preferred: use the YouTube Data API playlistItems endpoint for the channel's uploads playlist. Example:
  bash

  curl "https://www.googleapis.com/youtube/v3/playlistItems?part=snippet,contentDetails&maxResults=5&playlistId=UPLOADS_PLAYLIST_ID&key=$YOUTUBE_API_KEY"
  

• Lightweight alternative: use yt-dlp to check the latest upload ID without downloading video:
  bash

  yt-dlp --flat-playlist --dump-json "https://www.youtube.com/@NateBJones/videos" | head -n 1 > latest.json
  jq -r '.id' latest.json
  

• Compare the discovered video ID with the last processed ID stored in your run logs (e.g., a simple last_video.txt or a Notion/Sheets tracker). Only proceed if it's new.

3. Fetch transcripts

1. Try official captions via youtube_transcript_api:
   python

   from youtube_transcript_api import YouTubeTranscriptApi
   transcript = YouTubeTranscriptApi.get_transcript(video_id, languages=['en'])
   text = '\n'.join([chunk['text'] for chunk in transcript])
   

2. If captions are unavailable, download audio and run Whisper:
   bash

   yt-dlp -f 140 -o audio.m4a "https://www.youtube.com/watch?v=$VIDEO_ID"
   whisper audio.m4a --model medium --language en --task transcribe --output_format txt
   

3. Save the raw transcript alongside metadata (title, URL, publish date, duration). Keep it in your logs for traceability but do not distribute it by default.

4. Summarize

Produce:

• Abstract (2–3 sentences) summarizing the thesis of the video.
• Highlights – 4–6 bullets (verb-led). Mention timestamps where possible (e.g., [05:42] Key insight).
• References & Links – always include the YouTube URL and any external resources the video mentions.

Template:

# Nate B Jones Daily Digest — {{DATE}}

**Video:** {{TITLE}} ({{DURATION}}) → {{URL}}
**Abstract:** ...

## Highlights
- ...

## References & Links
- {{URL}}
- ...

5. Publish per config

Email (uses gog skill)

Do not attach the transcript unless someone explicitly asks for it—email only the digest body linked above.

GOG_KEYRING_PASSWORD=... gog gmail send \
  --to "{{config.outputs.email.to}}" \
  --subject "Nate B Jones Digest — {{DATE}}" \
  --body-file summary.txt \
  --body-html summary.html

Chat

• Control UI / Telegram: paste the summary or use the relevant messaging command (e.g., message action=send ...).
• Respect config.outputs.chat.targets (list of surfaces).

Document archive

• Google Docs:
  bash

  gog docs create "Nate B Jones Digest {{DATE}}" --body summary.md
  gog docs share <docId> --email {{config.outputs.doc.share_with}}
  

• Markdown on disk: write to the specified path in outputs.doc.path.

6. Automate (optional)

• Create a cron job or OpenClaw cron entry using poll_cron from config. Each run should:
  1. Poll for new video.
  2. If found, fetch transcript, summarize, publish, log the video ID.
• Keep lightweight audit logs (CSV or JSON) so you can prove what was sent and avoid duplicate emails.

References

• references/config-example.yml — copy/edit this to match each installation.
• youtube_transcript_api docs: https://pypi.org/project/youtube-transcript-api/
• Whisper CLI: https://github.com/openai/whisper

Stick to the playbook format every time so downstream consumers get consistent digests, and always fall back to Whisper if captions are missing.