A guide to creating AI-generated transcripts with summaries and key quotes from talks and podcasts using Whisper and Claude.

See https://wesmckinney.com/presentations for example.

Overview

This workflow takes a video or podcast URL and produces a formatted markdown transcript with:

• YAML frontmatter for metadata
• A first-person summary of key topics
• Extracted "money quotes"
• Full verbatim transcript with speaker labels

Prerequisites

Required Tools

# yt-dlp for downloading audio from YouTube/Vimeo/etc
brew install yt-dlp

# ffmpeg for audio compression (if needed)
brew install ffmpeg

# Python 3.8+ with these packages
pip install openai pyyaml

API Keys

• OpenAI API key for Whisper transcription (set as OPENAI_API_KEY environment variable)

Complete Workflow

Step 1: Get Video Metadata

First, extract metadata from the video URL:

yt-dlp --print "%(title)s|||%(upload_date)s|||%(duration)s" "VIDEO_URL"

Important: The upload date may differ from the actual talk date. Verify the actual event date from the video description or event website.

Step 2: Download Audio

# Download as MP3 (quality 5 is good balance of size/quality)
yt-dlp -x --audio-format mp3 --audio-quality 5 -o "/tmp/talk-audio.%(ext)s" "VIDEO_URL"

Or use the provided script:

./scripts/download-audio.sh "VIDEO_URL" "output-name"

Step 3: Check File Size

Whisper API has a 25MB limit. Check and compress if needed:

ls -lh /tmp/talk-audio.mp3

# If > 25MB, compress:
ffmpeg -i /tmp/talk-audio.mp3 -b:a 64k -ac 1 /tmp/talk-audio-compressed.mp3

Step 4: Transcribe with Whisper

Using the OpenAI Whisper API:

python scripts/transcribe.py /tmp/talk-audio.mp3 > /tmp/raw-transcript.txt

Or manually via the API:

from openai import OpenAI

client = OpenAI()
with open("/tmp/talk-audio.mp3", "rb") as audio_file:
    transcript = client.audio.transcriptions.create(
        model="whisper-1",
        file=audio_file,
        response_format="text"
    )
print(transcript)

Step 5: Save Raw Transcript

Always save the raw transcript before formatting:

cp /tmp/raw-transcript.txt transcripts/raw/YYYY-MM-DD-event-slug-raw.txt

Step 6: Generate Summary with Claude

Use Claude to create a first-person summary. Provide the raw transcript and this prompt:

Please analyze this transcript and create:

1. A SUMMARY section written in first person ("I", "my") that covers the key topics discussed. Organize into subsections if there are distinct topics. Be factual - avoid grandiose language like "groundbreaking" or "revolutionary".

2. A KEY QUOTES section with 3-5 impactful direct quotes from the transcript, formatted as blockquotes with context.

3. A cleaned TRANSCRIPT with speaker names in bold followed by colons.

Raw transcript:
[paste transcript here]

Step 7: Create Formatted Transcript File

Create the file at transcripts/YYYY-MM-DD-event-slug.md:

---
title: "Talk Title"
date: YYYY-MM-DD
event: "Event Name"
location: "City, State/Country"
video_url: "https://..."
video_type: "Talk"
transcribed: YYYY-MM-DD
---

*This transcript and summary were AI-generated and may contain errors.*

## Summary

[First-person summary here]

## Key Quotes

> "Quote text here" — Context or speaker

## Transcript

**Speaker Name:** Dialogue text...

**Other Speaker:** Response text...

See templates/transcript-template.md for a complete template.

Step 8: Add to talks.yml (if applicable)

If you're integrating with a Quarto blog like the original:

- date: 'YYYY-MM-DD'
  type: podcast  # or: talk, interview, keynote, tutorial
  role: guest    # for podcasts: guest or co-host
  event: "Event Name"
  title: "Talk Title"
  location: Remote
  links:
  - type: Video
    url: https://...

The date must match the transcript filename exactly for auto-linking to work.

Step 9: Preview and Verify

quarto preview
# Check that transcript renders correctly
# Verify links work

File Naming Convention

Pattern: YYYY-MM-DD-event-slug.md

• Use the actual talk date, not the upload date
• Use lowercase with hyphens for the slug
• Keep slugs short but descriptive

Examples:

• 2024-05-15-talk-python-to-me-pandas.md
• 2023-09-20-pycon-keynote.md
• 2022-03-21-gresearch-interview.md

YAML Frontmatter Reference

*Use either video_url + video_type OR slides_url, not both.

Summary Writing Guidelines

Voice and Tone

• First person: Write as if you gave the talk ("I discussed...", "My approach...")
• Factual: Focus on what was actually said, not interpretation
• No puffery: Avoid "groundbreaking", "revolutionary", "transformative", etc.
• Organized: Use subsections (### Topic) for distinct themes

Structure

## Summary

Brief overview paragraph of the main topics covered.

### First Major Topic

Details about this topic...

### Second Major Topic

Details about this topic...

Key Quotes Formatting

## Key Quotes

> "The exact quote from the transcript" — Context about when/why this was said

> "Another impactful quote" — Speaker attribution if multiple speakers

Transcript Formatting

• Speaker names in bold followed by colon
• Each speaker turn on its own paragraph
• Preserve natural speech (can clean up minor filler words)
• Use *[brackets]* for non-speech elements: *[laughter]*, *[applause]*

## Transcript

**Host:** Welcome to the show. Today we're talking about data science.

**Guest:** Thanks for having me. I'm excited to discuss this topic.

*[Brief pause]*

**Host:** Let's start with your background.

Quality Checklist

Before publishing, verify:

• Filename follows YYYY-MM-DD-event-slug.md pattern
• Date is actual talk date (not upload date)
• All required YAML fields present
• AI disclaimer included
• Summary is in first person
• Summary avoids grandiose language
• Key quotes use blockquote format
• Speaker names are bolded in transcript
• No obvious transcription errors or gaps
• Raw transcript saved in raw/ folder

Troubleshooting

Whisper API 25MB Limit

Compress audio to reduce file size:

ffmpeg -i input.mp3 -b:a 64k -ac 1 output.mp3

Missing Transcript Sections

If Whisper output has gaps (repeated symbols, [inaudible]):

1. Re-run transcription on that section
2. Manually transcribe from video
3. Note gaps with [inaudible] markers

Directory Structure

your-project/
├── transcripts/
│   ├── raw/                    # Raw transcript backups
│   │   └── YYYY-MM-DD-*.txt
│   ├── _metadata.yml           # Quarto metadata (optional)
│   ├── transcript-styles.css   # Custom styles (optional)
│   └── YYYY-MM-DD-*.md         # Formatted transcripts
├── scripts/
│   ├── download-audio.sh
│   ├── transcribe.py
│   └── format-transcript.py
└── templates/
    └── transcript-template.md

Cost Considerations

• Whisper API: ~$0.006 per minute of audio
• Claude: Varies by usage for summary generation

A 1-hour talk costs approximately $0.36 for Whisper transcription.

License

These scripts and templates are provided as-is for educational purposes.