Common Mistake ❌

You ask your LLM what a YouTube educator teaches about a topic, and it answers from training data: outdated, incomplete, sometimes invented.

The real answer sits in the creator's channel, in 200 videos the model has never read.

So people copy paste transcripts into the chat, one video at a time, and give up around video number six.

Problems Addressed 😔

Some of the best domain knowledge online exists only as speech, locked in a format agents can't touch.

A channel with a few years of uploads is 300+ hours of video but only ~500k tokens of text, and text is something an agent can actually work with.

Naive bulk approaches dump everything into one response and blow up the context window mid conversation.

Scraping transcripts yourself at scale runs into YouTube throttling and IP blocks.

How to Do It 🛠️

Get every video's transcript as plain text, one document per video, with title and upload date attached, because temporal questions need that metadata.

Connect the transcripts through an MCP server instead of manual uploads, so the agent fetches exactly what a question needs, at request time. Disclosure: I built one at ytbulktranscript.app (single videos free, bulk pay per use), but the pattern works with any transcript source you wrap in MCP.

Paginate bulk responses with cursors and tell the agent, inside the tool description, to fetch incrementally and summarize as it goes. Tool descriptions are prompts; the model does what they say, including the dumb things.

Prefer creator uploaded captions over auto generated ones when they exist. Auto captions mangle niche jargon and skip punctuation on older videos.

Benefits 🎯

Answers cite actual videos by title and date instead of paraphrasing training data.

The whole catalog becomes queryable: core concepts, abandoned topics, contradictions between videos, how explanations changed over the years.

The same pipeline feeds research agents, repurposing workflows, and dataset building without a browser tab in sight.

Prompt Reference 📝

Bad Prompt 🚫

What does [creator] teach about risk management?

Good Prompt 👉

Use the transcript tools. List the videos of channel X,
fetch the relevant transcripts page by page, then summarize
what the creator teaches about risk management.
Cite video title and upload date for every claim.

Considerations ⚠️

Transcripts capture words, not visuals. Anything shown on screen instead of said out loud is invisible to the agent.

Use this for research, learning and analysis. Republishing a creator's content wholesale is a different thing entirely.

Large channels still need pagination or retrieval; no context window swallows 200 videos in one gulp.