Download Transcripts from CSV

Why? Manually downloading transcripts one-by-one is tedious and error-prone. This skill automates bulk transcript downloads with rate limiting, progress tracking, and resume capability.

Quick Start

# Standard usage (recommended)
transcript-download --csv data/channel-name/videos.csv --output-dir data/channel-name/transcripts

# Single video
transcript-download https://www.youtube.com/watch?v=VIDEO_ID --output transcript.md

Workflow

1. Verify Prerequisites

Ensure you have a CSV file with video URLs:

• Created by extract-videos skill, OR
• Manually curated with video_url column

# Check CSV exists and has videos
head -5 data/channel-name/videos.csv

[!TIP] If you don't have a CSV yet, run extract-videos first to build your video list from a YouTube channel.

2. Run Download Command

transcript-download \
  --csv data/<folder-name>/videos.csv \
  --output-dir data/<folder-name>/transcripts \
  --delay 60

[!CAUTION] NEVER set --delay below 30 seconds. YouTube will block your IP if you make requests too quickly. The default of 60 seconds exists for a reason. You WILL get banned and may wait hours before downloading again.

3. Validate Downloads

After the command completes:

# Check how many transcripts downloaded
ls -la data/<folder-name>/transcripts/*.md | wc -l

# Verify CSV status updated
grep -c "True" data/<folder-name>/videos.csv

The command automatically:

• Downloads transcripts as markdown with YAML frontmatter
• Updates CSV transcript_downloaded column with True/False
• Skips already-downloaded videos on re-run

Output Format

Transcripts are saved as markdown with YAML frontmatter:

---
video_id: dQw4w9WgXcQ
video_url: https://www.youtube.com/watch?v=dQw4w9WgXcQ
title: Building Resilient Microservices at Scale
author: Jane Smith
published_date: 2025-03-15
length_minutes: 42.5
views: 15234
---

[00:00] Welcome to today's talk on building resilient microservices...
[00:15] We'll cover three main topics: circuit breakers, retry patterns...

Examples

Overnight Batch Processing

# Large channel - run overnight with safe delay
transcript-download \
  --csv data/aws-reinvent-2025/videos.csv \
  --output-dir data/aws-reinvent-2025/transcripts \
  --delay 90 \
  --verbose

Conference Playlist

# Smaller collection - minimum safe delay
transcript-download \
  --csv data/pycon-2024/videos.csv \
  --output-dir data/pycon-2024/transcripts \
  --delay 30

Non-English Content

# Spanish content with fallback to auto-generated
transcript-download \
  --csv data/spanish-tech/videos.csv \
  --output-dir data/spanish-tech/transcripts \
  --languages es es-MX es-ES auto

Single Video Mode

# Quick single transcript
transcript-download https://www.youtube.com/watch?v=dQw4w9WgXcQ --output talk.md

Troubleshooting

Common Mistakes

1. Reducing delay to "speed up" - Do not set delay below 30 seconds. You will get IP banned. Use overnight runs for large batches.

2. Running without checking CSV first - Always verify your CSV exists and has videos before running. Empty CSVs waste time.

3. Not using resume capability - If interrupted, don't start over. Re-run the same command to resume from where you left off.

4. Wrong language codes - Language codes are specific (e.g., en-US not english). Check YouTube's language for the video.

5. Forgetting auto in languages - Many videos only have auto-generated captions. Add auto to your language list as a fallback.

Quality Checklist

Before running:

• CSV file exists and has video_url column
• Output directory path is correct
• Delay is 30+ seconds (60+ recommended)
• Sufficient time allocated (1 min per video minimum)

After running:

• Transcript count matches expected videos
• CSV transcript_downloaded column updated
• Spot-check 2-3 transcripts for valid content
• No empty or malformed markdown files