---
title: YouTube Digest
url: https://blog.guigpap.com/en/workflows/youtube-digest/
url_md: https://blog.guigpap.com/en/workflows/youtube-digest.md
category: automation
date: '2026-01-31'
maturite: production
techno:
  - n8n
  - telegram
  - claude
  - qdrant
application:
  - ai
  - content
  - knowledge
---

# YouTube Digest

> End-to-end pipeline that turns YouTube videos into actionable AI summaries

## 1. What? — Definition and context

The **YouTube Digest** workflow is a full pipeline that extracts, transcribes and summarises YouTube videos. It supports multiple input sources and generates TLDR summaries with on-demand enrichment via Telegram.

> **Note - Pipeline**
>
> A **pipeline** is a chain of automated steps. Here: fetch a video → transcribe it → summarise it → store the result → notify. Each step can be customised to fit your needs.

### Features

| Feature | Description |
|---------|-------------|
| **3 input sources** | Schedule (playlists), Telegram (URL), Watch Later |
| **Smart transcription** | Prefers YouTube captions (free), Whisper fallback |
| **TLDR summary** | Key points extracted via Claude |
| **Vector storage** | Embeddings in Qdrant for semantic search |
| **On-demand enrichment** | Telegram buttons to dig deeper |

### Input sources

| Source | Trigger | Usage |
|--------|---------|-------|
| **Schedule** | Configured cron | Tech/news playlists |
| **Telegram URL** | Message to the bot | One-off video |
| **Watch Later** | YouTube playlist | Saved videos |

---

## 2. Why? — Stakes and motivations

### Problems solved

| Problem | Without digest | With digest |
|---------|----------------|-------------|
| **Watch time** | Watch the whole video | TLDR in 30 seconds |
| **No search** | Impossible to find a passage | Semantic search via Qdrant |
| **Manual notes** | Take notes while watching | Key points auto-extracted |
| **Video backlog** | Watch Later piles up | Automatic processing |

### Why prioritise YouTube captions?

| Method | Cost | Speed | Quality |
|--------|------|-------|---------|
| **YouTube captions** | Free | Instant | Variable (auto-generated) |
| Whisper API | ~$0.36/h | ~10-30s | Excellent |

> **Tip - Cost optimisation**
>
> By using YouTube captions when available, the average cost per video is ~$0.02 (just Claude + embedding) instead of ~$0.40 with full transcription.

---

## 3. How? — Technical implementation

### Architecture

```mermaid
flowchart TD
  subgraph Sources["Input sources"]
    direction TB
    Cron["Schedule · playlists"]
    TG["Telegram URL"]
    WL["Watch Later · YouTube"]
  end

  Norm["Normalize Input · extract video_id"]
  Check{"Already processed ?"}
  Skip["Skip"]
  Meta["Get Metadata · title, duration"]
  Transcribe["Transcription · captions or Whisper"]
  Summarize["TLDR via cli-ollama"]
  Storage["Storage · Qdrant + Data Table"]
  Notify["Telegram notification + enrichment buttons"]

  Sources --> Norm --> Check
  Check -->|already| Skip
  Check -->|new| Meta --> Transcribe --> Summarize --> Storage --> Notify
```

### Smart transcription

```mermaid
flowchart TD
  Captions["Try YouTube Captions"]
  Found{"Captions found ?"}
  Parse["Parse SRT / VTT"]
  Download["Download Audio · yt-dlp"]
  Whisper["Whisper API"]
  Clean["Clean Transcript"]

  Captions --> Found
  Found -->|Yes| Parse --> Clean
  Found -->|No| Download --> Whisper --> Clean
```

**Long-file handling:**

| Duration | Strategy |
|----------|----------|
| < 25 min | Whisper direct |
| ≥ 25 min | Split into chunks (~25MB limit) |
| > 2h | Skip or confirmation |

### TLDR summary (Claude)

**Prompt:**

```javascript
{
  "model": "claude-haiku-yolo",
  "prompt": `Analyse this YouTube transcript and generate:
1. TLDR (2-3 sentences max)
2. Key points (5 max, bullet points)
3. Topics (3-5 tags)
4. Sentiment (educational/entertaining/promotional/technical)
5. Language (ISO code)

Transcript:
${transcript}`
}
```

**Output:**

```json
{
  "tldr": "2-3 sentence concise summary",
  "key_points": ["point1", "point2", "point3"],
  "topics": ["tech", "ai", "tutorial"],
  "sentiment": "educational",
  "language": "en"
}
```

### Telegram notification

```
📺 NEW DIGEST

**How AI is Changing Software Development**
Channel: Fireship • 12:34

**TLDR:**
AI tools like Copilot and Cursor are transforming dev.
Focus on productivity gains and risks.

**Key points:**
• GitHub Copilot boosts velocity by 40%
• Risk: junior developers less autonomous
• Cursor = promising AI-first IDE
• Trend: "vibe coding" vs "intentional coding"

Tags: #ai #development #tools

[📜 Full transcript] [🧠 Atomic notes] [📎 Link to Odoo] [❓ Q&A]
```

### On-demand enrichment

| Button | Action |
|--------|--------|
| 📜 Full transcript | Full structured transcription with chapters |
| 🧠 Atomic notes | Zettelkasten notes (one per concept) |
| 📎 Link to Odoo | Link to a CRM contact/project |
| ❓ Q&A | Generate flashcard-style Q&A |

### Qdrant storage

```javascript
{
  "collection": "youtube_digests",
  "vector": embedding,
  "payload": {
    "video_id": "dQw4w9WgXcQ",
    "title": "...",
    "tldr": "...",
    "key_points": [...],
    "topics": [...]
  }
}
```

### N8N Data Table

| Column | Type | Description |
|--------|------|-------------|
| video_id | string | YouTube video ID |
| title | string | Video title |
| channel | string | Channel name |
| duration | number | Duration in seconds |
| processed_at | datetime | Processing date |
| tldr | text | Short summary |
| key_points | json | Key points list |
| topics | json | Extracted tags |
| qdrant_id | string | Qdrant vector ID |

---

## 4. What if? — Outlook and limits

### Cost estimation

| Operation | 10-min video | 60-min video |
|-----------|--------------|--------------|
| YouTube captions | Free | Free |
| Whisper (if needed) | ~$0.06 | ~$0.36 |
| Claude TLDR | ~$0.01 | ~$0.05 |
| Embedding | ~$0.0001 | ~$0.001 |
| **Total** | **~$0.01-0.07** | **~$0.05-0.41** |

### Current limits

| Limit | Impact | Mitigation |
|-------|--------|------------|
| **Videos > 2h** | Possible timeout | Manual confirmation |
| **Missing captions** | Whisper cost | Acceptable for important videos |
| **Multi-language** | Summary in source language | Optional translation |

### Evolution scenarios

**If multi-language is needed**:
- Auto-detect language
- Optional summary translation
- Multilingual embeddings

**If video volume grows**:
- Filter by relevance score
- Grouped daily digest
- Priority queue

**If advanced search is needed**:
- Web interface for Qdrant search
- Filters by topic/channel
- Similar video suggestions

### Useful commands

```bash
# Test video_id extraction
echo "https://www.youtube.com/watch?v=dQw4w9WgXcQ" | \
  grep -oP '(?<=v=)[^&]+'

# Check available captions
yt-dlp --list-subs "https://www.youtube.com/watch?v=VIDEO_ID"

# Download audio for testing
yt-dlp -x -f bestaudio "https://www.youtube.com/watch?v=VIDEO_ID"

# Semantic search in Qdrant
curl -X POST "http://qdrant:6333/collections/youtube_digests/points/search" \
  -H "Content-Type: application/json" \
  -d '{"vector": [...], "limit": 5}'
```

### Troubleshooting

| Problem | Check |
|---------|-------|
| No captions | Video without subtitles → Whisper fallback |
| Whisper timeout | File too long? Increase timeout |
| Claude error | Quota exhausted? Check claude-ollama usage |
| Qdrant 404 | Collection exists? Create if needed |

---

## Related pages

### Workflows
- [Telegram Orchestrator](/en/workflows/telegram-orchestrator/) — URL submission
- [Notification Hub](/en/workflows/notification-hub/) — Notification routing

### Infrastructure
- [AI Stack](/en/infrastructure/ai-stack/) — Claude Ollama + Qdrant
- [N8N Queue Mode](/en/infrastructure/n8n-queue-mode/) — Backend workers

## Metadonnees agent

- Cet article est issu du blog GuiGPaP Lab.
- Contexte global du blog: https://blog.guigpap.com/llms.txt
- Contact auteur: https://odoo.guigpap.com/mon-cv
- Licence: CC-BY-SA 4.0