tools-only /video-digest
Process a list of YouTube videos using a triage-then-process pattern. Parallel Haiku agents quickly score each video's relevance to your topic, then Sonnet agents deeply analyse only the must-watch videos. Saves time by filtering before investing in deep processing.
When to Use This Skill
- Processing a playlist or list of conference talks
- Filtering a curated reading list of videos by relevance
- Researching a topic across multiple YouTube sources
- Creating a digest of the best content from a set of videos
- Preparing for a conference by pre-screening talk recordings
Usage
/video-digest <topic> --urls <url1> <url2> ... [--threshold 7]
Parameters
| Parameter | Description | Required |
|---|---|---|
topic |
Topic or question to evaluate relevance against | Yes |
--urls |
List of YouTube URLs to process | Yes |
--threshold |
Minimum relevance score for deep processing (1-10, default: 7) | No |
Instructions
Phase 1: Collect Video List
Gather the list of videos from the user:
- Accept URLs directly as arguments
- Accept a file path containing URLs (one per line)
- Accept a YouTube playlist URL (extract individual video URLs)
Phase 2: Triage — Agent Team (Parallel Haiku Agents)
Use the Triage + Selective Processing pattern. Launch parallel Haiku agents, each processing 3-5 videos.
Agent 1-N: Video Triage Agent (Haiku) Task: Quick-assess relevance of assigned videos
- Fetch video title, description, and duration (via MCP or WebFetch)
- If transcript is quickly available, scan the first 5 minutes
- Score relevance to the user's topic on a 1-10 scale:
- 8-10 (Must Watch): Directly addresses the topic with depth
- 5-7 (Worth Processing): Partially relevant, some useful content
- 1-4 (Skip): Not relevant enough to warrant deep processing
- Categorise: must-watch, worth-processing, skip
- Write a one-line reason for the score
Return: List of
{ videoUrl, title, duration, score, category, reason }
Phase 3: Deep Processing — Agent Team (Selective Sonnet Agents)
Only for videos scoring at or above the threshold (default: 7):
Agent 1-N: Deep Analysis Agent (Sonnet) Task: Fully analyse one must-watch video
- Fetch full transcript with timestamps
- Generate detailed summary (3-5 paragraphs)
- Extract key frameworks, concepts, and methodologies
- Identify actionable takeaways
- Note tools and resources mentioned
- Create timestamped topic index Return: Full analysis document for one video
Phase 4: Synthesise Digest
Combine all results:
- Triage summary table — All videos with scores
- Deep analyses — Full notes for must-watch videos
- Cross-video themes — Common themes across analysed videos
- Recommended viewing order — Prioritised by value and logical flow
Output Format
# Video Digest: <Topic>
**Videos assessed:** X | **Must-watch:** X | **Processed:** X
## Triage Summary
| # | Video Title | Duration | Score | Category | Reason |
|---|--------------------------|----------|-------|----------------|-----------------------|
| 1 | <Title> | 45:00 | 9/10 | Must Watch | <One-line reason> |
| 2 | <Title> | 30:00 | 8/10 | Must Watch | <One-line reason> |
| 3 | <Title> | 20:00 | 5/10 | Worth Processing| <One-line reason> |
| 4 | <Title> | 60:00 | 3/10 | Skip | <One-line reason> |
## Recommended Viewing Order
1. **<Title>** (Score: 9) — Start here because...
2. **<Title>** (Score: 8) — Builds on concepts from #1...
## Cross-Video Themes
### Theme 1: <Common theme across videos>
- Discussed in: Video 1 (15:30), Video 2 (08:45)
- Key insight: <synthesis>
### Theme 2: <Another theme>
...
---
## Detailed Analysis
### Video 1: <Title> (Score: 9/10)
> **URL:** <url> | **Duration:** 45:00 | **Channel:** <name>
#### Summary
<Detailed summary>
#### Key Takeaways
1. <Takeaway>
2. <Takeaway>
#### Frameworks and Concepts
- **<Framework>:** <Description>
#### Timestamped Topics
| Timestamp | Topic | Key Point |
|-----------|-------------|-----------|
| 00:00 | <Topic> | <Point> |
---
### Video 2: <Title> (Score: 8/10)
...
Examples
Example 1: Conference Talk Digest
/video-digest "event-driven architecture patterns" --urls https://youtube.com/watch?v=abc https://youtube.com/watch?v=def https://youtube.com/watch?v=ghi
Triages 3 conference talks about event-driven architecture, then deeply processes the most relevant ones.
Example 2: Lenient Threshold
/video-digest "kubernetes security" --urls <list> --threshold 5
Processes all videos scoring 5 or above (more inclusive).
Example 3: Research Topic
/video-digest "AI in enterprise architecture" --urls <list-of-10-videos>
Triages 10 videos in parallel, deeply processes the top-scoring ones, and synthesises cross-video themes.
Invoke with: /video-digest <topic> --urls <url1> <url2> ... to triage and deeply process the most valuable videos