video-digest

/video-digest

Intelligently triage new videos from subscribed YouTube channels, process selectively based on relevance, and generate a daily digest with watch recommendations.

Usage

/video-digest                    # Full triage and processing
/video-digest --triage-only      # Score videos without processing
/video-digest --report           # Show current queue status
/video-digest --watched <video>  # Mark video as watched (removes from queue)

Examples

/video-digest
/video-digest --triage-only
/video-digest --watched "YouTube - AI Security Patterns"

Model Strategy

Phase	Model	Rationale
RSS fetch	Haiku	Simple data extraction
Title/desc triage	Haiku	Fast relevance scoring
Worth Processing	Sonnet	Full transcript analysis
Quick Capture	Haiku	Brief summary only
Digest generation	Sonnet	Quality recommendations

Instructions

Phase 1: Gather New Videos

Model: Haiku (parallel subagents)

1. Read subscriptions from .claude/subscriptions.yaml

2. For each channel (in parallel), fetch RSS feed:

   https://www.youtube.com/feeds/videos.xml?channel_id={{channel_id}}
   

   Use WebFetch with prompt: "Extract video entries: title, video_id, published_date, description (first 500 chars)"

3. Identify new videos:

   • Compare video IDs against last_video_id
   • Select videos published since last check
   • If first check, take only most recent video

4. Collect candidates with basic metadata:

   • video_id, title, description, duration (if available)
   • channel_name, channel_trust_score, always_watch

Phase 2: Rapid Triage

Model: Haiku

For each candidate video, compute triage score:

Scoring Algorithm

TRParentCorpE_SCORE = (
    PROJECT_RELEVANCE * 0.35 +    # 0-100
    DOMAIN_RELEVANCE * 0.25 +     # 0-100
    FRESHNESS_BOOST * 0.20 +      # 0-100
    CHANNEL_TRUST * 0.20          # 0-100 (from trust_score * 5)
)