media-ingest

Media Ingest

Convert YouTube videos and podcast episodes into markdown notes with YAML frontmatter and full transcripts. Output directory is configurable via the --folder and vault path arguments.

Default vault path: /Users/adam/Documents/Obsidian Vault

Workflow

1. Determine content type from the URL or user request
2. Run the appropriate ingest script
3. Confirm to user with file path and summary

YouTube Videos

Run scripts/ingest_youtube.py:

python3 ~/.claude/skills/media-ingest/scripts/ingest_youtube.py "<youtube_url>" "/Users/adam/Documents/Obsidian Vault" --folder YouTube

• Fetches metadata via yt-dlp (title, channel, date, duration, chapters, tags)
• Fetches clean transcript via youtube-transcript-api, falls back to yt-dlp subtitles
• Output: <vault>/YouTube/<video-title>.md

Podcast Episodes

Run scripts/ingest_podcast.py:

# Latest episode from Apple Podcasts URL
python3 ~/.claude/skills/media-ingest/scripts/ingest_podcast.py "https://podcasts.apple.com/us/podcast/<name>/id<id>" "/Users/adam/Documents/Obsidian Vault"

# List episodes first
python3 ~/.claude/skills/media-ingest/scripts/ingest_podcast.py "<url>" "/Users/adam/Documents/Obsidian Vault" --list

# Specific episode by number (1=latest)
python3 ~/.claude/skills/media-ingest/scripts/ingest_podcast.py "<url>" "/Users/adam/Documents/Obsidian Vault" --episode 3

# Search by title
python3 ~/.claude/skills/media-ingest/scripts/ingest_podcast.py "<url>" "/Users/adam/Documents/Obsidian Vault" --search "keyword"

# Skip transcription (metadata only, fast)
python3 ~/.claude/skills/media-ingest/scripts/ingest_podcast.py "<url>" "/Users/adam/Documents/Obsidian Vault" --skip-transcribe

• Resolves Apple Podcasts URLs to RSS feeds automatically via iTunes Lookup API
• Downloads audio, transcribes locally with Whisper (small model by default)
• Output: <vault>/Podcasts/<Show Name>/<Episode Title>.md
• Use --model tiny for faster transcription, --model medium for higher accuracy
• Transcription takes ~15 min per hour of audio on M1 Pro with small model

Granola Meetings

Reads the local Granola cache (~/Library/Application Support/Granola/cache-v4.json) directly — no API needed.

# Export all meetings
python3 ~/.claude/skills/media-ingest/scripts/ingest_granola.py --vault "/Users/adam/Documents/Obsidian Vault"

# List meetings (last 30 days)
python3 ~/.claude/skills/media-ingest/scripts/ingest_granola.py --list --days 30

# Only meetings with transcripts
python3 ~/.claude/skills/media-ingest/scripts/ingest_granola.py --vault "/Users/adam/Documents/Obsidian Vault" --transcripts

• Output: <vault>/Granola-Meetings/<date>-<title>.md
• Includes attendees, duration, calendar link, notes, and full transcript with speaker separation
• Only 6 of 187 meetings have transcripts cached locally (Granola caches recent transcripts)
• QMD collection: meetings

Dependency Check

python3 ~/.claude/skills/media-ingest/scripts/check_deps.py

Required: yt-dlp, youtube-transcript-api, feedparser Required for podcasts: openai-whisper

pip3 install youtube-transcript-api feedparser openai-whisper

Markdown Output Format

Both scripts produce notes with YAML frontmatter containing: title, date, ingested, duration, type (youtube|podcast), url, and source-specific fields (channel for YouTube, show/author for podcasts). The body contains a description/summary section and the full transcript.