I Fed an Entire YouTube Channel Into an LLM (200 Videos, ~550k Tokens) 🥊

One YouTube channel. 200 videos. Three AI setups. Let's see who chokes first.

🧐 What is this, and why should you care?

I wanted to answer one slightly silly question: can an LLM actually digest a whole YouTube channel, or does that only work in Twitter demos?

Not one video. Not a playlist. A full channel with three years of uploads. The kind of channel where the creator sells a $1,500 course while the same knowledge sits in the back catalog. For free.

So I grabbed the transcripts of the entire channel and force fed them to three different setups:

• Brute force: paste everything into one big context window
• NotebookLM: Google's "throw sources at it and ask" tool
• Claude Projects: a knowledge base with retrieval on top

If you build with LLMs, learn from YouTube, or ever wondered "could I just ask an AI about this creator's whole channel", this one is for you.

Spoiler: all three survived. None of them behaved the same. 😅

✅ Prerequisites

Nothing scary. You're good if you have:

• Some way to get YouTube transcripts (three options below)
• Access to one large context LLM
• A free NotebookLM or Claude account
• Copy paste skills. That's genuinely it.

🤔 The Problem (and why it pokes everyone)

Quick story.

Every niche has that one educator. Good teacher, big free channel, paid program on top. The knowledge is public. It's just trapped in a format you can't search, quote, or query. 200 videos, 15 to 30 minutes each, no index. That one brilliant explanation from 2022? Good luck.

Text fixes this. A 15 minute video is roughly 2,000 spoken words, call it ~2,700 tokens. Suddenly a channel is not "300 hours of watching". It's a dataset. Mine came out around 550k tokens.

And 550k tokens is the whole problem, because no chat window on earth swallows that in one gulp.

🧪 The Experiment

1. Getting the text 🎲

Goal: transcripts for every video, as plain text.

Three ways to do it: YouTube's own transcript button (fine for 3 videos, torture for 200), yt-dlp with --write-auto-subs --skip-download (free, but YouTube throttles and blocks IPs when you hammer a full channel, ask me how I know), or a hosted tool. Disclosure: I built one of those, ytbulktranscript.app, precisely because of the previous sentence. Single videos are free, bulk costs a few dollars. The experiment doesn't care which route you take.

One tip that saved me later: one file per video, title and upload date in the filename. You'll want that metadata. Promise.

2. Contender one: brute force 🐘

Goal: concatenate everything with --- VIDEO: title (date) --- separators and paste it into a 200k context model.

Result: quick math ruins the party. At ~2,700 tokens per video, a 200k window fits roughly 70 average videos. My 200 video channel? Dead on arrival. But for channels under ~50 videos this is honestly the best setup there is. Everything in context, no retrieval deciding what the model gets to see.

3. Contender two: NotebookLM 🟡

Goal: upload everything, let Google worry about it.

The catch: 50 sources per notebook. 200 transcripts is not 50 sources... until you merge 4 videos per file. 50 files, 200 videos, one notebook. It ingested the whole thing without complaints and gives you citations pointing at the exact source snippet.

4. Contender three: Claude Projects 🔵

Goal: upload the files to a project knowledge base and let retrieval pull what's relevant per question.

Works with the full channel, no merging tricks needed. The trade: retrieval decides what the model sees, and for broad "across the whole channel" questions that selection sometimes misses things you know are in there.

📊 The Results

So no clean winner. Sound familiar? Small channel: brute force. Big channel and you want sources you can check: NotebookLM. Already living in Claude: Projects.

🌶️ The prompts (this is where it gets spicy)

Generic "summarize this channel" produces generic mush on every setup. These earned their keep:

1. "List the 10 to 20 core concepts this creator repeats across videos. For each, cite the videos that explain it best."
2. "Build a topic map: what gets covered often, what was mentioned once and abandoned?"
3. "Compare how [concept] is explained in the oldest vs. newest videos. What changed?"
4. "Find contradictions between videos."
5. "Extract every framework, checklist and step by step process into one reference document."

Prompt 3 is why the upload dates in the filenames matter. Without them, the model has no idea what "older" means.

🎯 So what did I actually learn?

Channel size picks your architecture, not preference. Under 50 videos, skip all tooling and paste. Above that, you need NotebookLM or retrieval, and you accept their quirks.

The creator repeated ~15 core ideas across 200 videos. Different examples, same frameworks. Watching hides this completely. Querying across everything makes it embarrassingly obvious, including how precisely the free content funnels into the paid course.

Auto captions are messier than you think. No punctuation on older videos, niche jargon gets mangled. The models coped better than expected, but verify anything you plan to quote.

Metadata is the quiet MVP. Titles and dates in filenames turned "a pile of text" into something you can ask temporal questions about. Cost: zero effort at collection time.

🙌 Wrapping up

That was a fun one. One channel, ~550k tokens, three setups, and a pretty clear picture of who's good at what.

If you learn from YouTube educators: their channel is a dataset now. Treat it like one.

If you try this on a channel and something interesting falls out, tell me in the comments. And if people are interested, I might write a follow up on automating the whole pipeline.

Thanks for reading :)