ChatGPT Chunker

This was the first thing I ever coded.

March 15, 2023. GPT-4 had launched the day before. I had a full podcast transcript and I wanted ChatGPT to write the show notes, the title, the description. I pasted the whole thing in. It came back with a red error: the message you submitted was too long.

The context window was bigger than what you were allowed to send in one message. Around 4,000 tokens per paste. You could keep a longer conversation going, you just weren't allowed to dump a document in a single go. People were splitting books by hand and watching the model forget the first half the moment they sent the second.

I spent about twenty minutes doing that. Then I got annoyed enough to build a page.

That's the whole product. You write a bit of context so the model knows what's coming. You set a token size — 3,000 was the safe number for GPT-3.5 and GPT-4. You paste the text. It hands you a stack of prompts, each one telling ChatGPT to just confirm receipt and wait. You paste them in order. I ran a twenty-chunk transcript through it the same morning and it still had the whole conversation when I asked for the title.

I put it on conturata.com/ai/chunker because that was the site I already had. I recorded a video the same week walking through the error. I called it my first ever app, because it was.

Then Google started sending people.

Not because I ran ads. Because everyone who hit that red error searched for it, and the page was sitting there. I don't have the Analytics tab open and I honestly forgot this thing existed until now, so I don't remember the exact number. Somewhere between half a million and a million people over the life of it, all organic.

Then the copycats showed up. Same layout, same FAQ, different domain.

Then the problem went away. Context windows went to 32k, then 128k, then a million tokens, and you can paste a book into a chat now without thinking about it. The page is still up. It still works. It just doesn't get the traffic, which is the right ending for a tool that existed to paper over a limit that was never supposed to last.

I taught it at Mindvalley's AI summit. It was the first thing I pointed people at when they asked how to actually use ChatGPT on real work instead of toy prompts.