YouTube Transcript to Markdown for Claude / ChatGPT: The 2026 Workflow

YouTube is the largest single audio-knowledge corpus on the open web. Long-form interviews (Lex Fridman, Tim Ferriss, Acquired), conference talks, tutorial deep dives, lecture series — all of it is searchable, free, and almost entirely unused in serious AI research workflows. The bottleneck isn't the model; it's getting clean transcript text out of YouTube and into your AI of choice.

This post is the workflow that turns a 90-minute talk into Markdown Claude or GPT-5.5 can actually reason over.

Why YouTube transcripts are hard to use directly

If you click "Show transcript" on a YouTube video and copy the result, you get something like:

0:00 hey everyone welcome back to the show today we're talking about
0:03 transformer architectures and how attention scales with
0:06 the input sequence length so basically the
0:09 main thing you need to understand is that

Three problems for LLM input:

1. Token waste: every 3-second timestamp is 2-4 tokens. A 60-minute video accumulates ~3,000 tokens of pure timestamp noise.
2. No semantic structure: no paragraphs, no sections, no speaker labels. Claude has to infer topic shifts from prose alone.
3. Punctuation is missing: auto-generated transcripts are unpunctuated continuous text. Sentence boundaries are inferred not given.

The result: token-inefficient, harder to reason over, and useless for citation ("at what timestamp did the host say X?").

What clean YouTube Markdown looks like

After running through a YouTube-aware extractor:

# Transformers Explained Simply
**Channel**: 3Blue1Brown · **Duration**: 45:12 · **Published**: 2026-04-15
**Source**: https://www.youtube.com/watch?v=abc123

## 00:00 — Introduction and motivation

Hey everyone, welcome back to the show. Today we're talking about transformer
architectures and how attention scales with the input sequence length. The main
thing you need to understand is...

## 08:42 — Self-attention mechanism

[Continues with proper paragraphs and section breaks]

## 23:15 — Multi-head attention

...

## 38:50 — Practical implementation

...

## Top Comments

- **@user1234** (👍 847): "The diagram at 12:30 finally made me understand what query/key/value
  vectors actually mean — thank you!"
- **@user5678** (👍 412): "Small correction: at 19:30 the multiplication should be QK^T not Q*K..."

Roughly 40% smaller than the raw transcript. Timestamps preserved as section anchors so you can cite specific moments. Top comments included for corrections and added context. Claude reads this and produces accurate quotes with timestamp-precise citations.

The workflow

Three paths, depending on your setup:

Path 1: Web2MD's YouTube extractor (easiest)

Open the YouTube video in Chrome. Click Web2MD. The extractor pulls:

• Title, channel, duration, publish date, description
• Full transcript with auto-detected section breaks
• Timestamps preserved as ## HH:MM — section heading anchors
• Top comments by upvote count
• Formatted as clean Markdown ready to paste into Claude or ChatGPT

End-to-end: about 8 seconds per video. Free tier handles 3 videos/day; Pro is unlimited.

Path 2: YouTube Transcript API + custom script

For developers who want batch processing:

from youtube_transcript_api import YouTubeTranscriptApi
import re

def youtube_to_markdown(video_id):
    transcript = YouTubeTranscriptApi.get_transcript(video_id)

    # Group into ~5-minute sections
    sections = []
    current_section = {"start": 0, "text": []}
    for entry in transcript:
        if entry["start"] - current_section["start"] > 300:  # 5 min
            sections.append(current_section)
            current_section = {"start": entry["start"], "text": []}
        current_section["text"].append(entry["text"])
    sections.append(current_section)

    md = []
    for s in sections:
        mins = int(s["start"] // 60)
        secs = int(s["start"] % 60)
        md.append(f"## {mins:02d}:{secs:02d}")
        md.append(" ".join(s["text"]).replace("\n", " "))
        md.append("")
    return "\n".join(md)

Works for batch jobs (100+ videos for a corpus). Misses comments and metadata — add YouTube Data API for those if needed.

Path 3: Whisper for videos without captions

For uploaded videos missing auto-captions:

yt-dlp -x --audio-format mp3 <video_url>
whisper.cpp -m models/ggml-large-v3.bin -f audio.mp3 -of transcript -otxt

Then run the same Markdown-cleaning pass on Whisper's output. Costs ~$0.36 per hour via OpenAI's hosted API, or free via local Whisper.cpp on an M-series Mac.

A real use case: Multi-podcast research synthesis

Last month I wanted to compare how three different AI podcasts (Latent Space, Cognitive Revolution, No Priors) had covered a specific architectural choice across 6 months of episodes.

• Identified 15 relevant episodes via search.
• Web2MD batch-export of transcripts: about 12 minutes.
• Result: 180-page Markdown corpus, ~140k tokens.
• Pasted into Claude Opus 4.7 with the prompt: "These are 15 podcast transcripts. Identify how each host approached the topic of [X]. Show evolution over time with timestamped quotes."
• Output: a chronological comparison with verified citations to specific podcast minutes.

Total workflow time: ~80 minutes including the actual listening I had already done. The manual-only version would have been an entire weekend.

What this is not

Honest about the limits:

• Not a substitute for watching the video. For demos, code walk-throughs, or anything where visual content matters, the transcript loses the show-not-tell. Use this for talk-heavy content (interviews, lectures, podcasts).
• Not for live streams. Snapshot workflow. Use the transcript only after the stream concludes.
• Not for music or non-speech audio. Whisper is good but designed for speech.
• Not commercial training data. YouTube's terms restrict bulk extraction for model training. Personal research and individual AI prompts are fine; building a 10M-video training corpus is not.

Pairing with other workflows

This workflow composes well with:

• Reddit-to-Claude pipeline: Reddit discussions about the podcast + transcript = full discourse
• Fill Claude's 1M context window: 12 podcast transcripts is roughly 200k tokens — fits comfortably
• DeepSeek R2 + Chinese content pipeline: Chinese podcasts on Bilibili use the same workflow with Bilibili-specific extractors
• Reduce LLM token costs: clean transcripts cost 40% less than raw

Quick wins

If you already use Web2MD, open any YouTube video right now and click the extension. The result is what this post describes. The free tier handles 3 videos/day; Pro unlocks bulk queue for multi-episode research sessions.

For dev workflows, the YouTube Transcript API + 25 lines of Python (above) gets you 90% of the way.

Related

• Why Claude can't read Reddit (and how to fix it)
• How to fill Claude's 1M context window
• Reddit to Claude 1M context: research pipeline
• How to reduce LLM token costs (practical)
• Convert YouTube to Markdown — supported sites page

Install

Web2MD on the Chrome Web Store →

Free tier: 3 conversions/day. Pro at $9/mo unlocks unlimited + queue + bulk export + dedicated YouTube extractor with timestamp anchors.