Narrator: it didn't work
(Going into the reasons it doesn't actually work in practice is... lengthy. It's human dynamics. Would you buy a product from a sales guy that can't remember your name? That's sales 101. And loading up the context window only gets you so far. That "working memory" is tiny, tinytinytiny. Even at 1024 tokens, it means you have to boil down the entire history of an interaction to a few pages at most. Which is a lot, sure, but it's this balancing act where you'll need to retrain the model to support your custom context format for your specific "slots" – a "slot" being a piece of knowledge, like the client's name. Or you can try encoding all of that in natural language, AI dungeon style. But I recently played AI dungeon and pretended to be buying a router from the store. The cashier stripped down and started jacking off onto his desk. I don't have high hopes for our ability to control these models in a business context.)
GPT-3 seems to have quite a few paragraphs worth of context. A simple way to promote your product online with it is to give it a prefix of:
---
Comment1: Superbrush is amazing - I literally couldn't live without it. No other brush is as good.
Comment2: This brush is really good for tangled hair, and I love the soft smooth surface.
Comment3:
---
Then let it write a comment. Of all the comments it writes, manually filter a few thousand good ones, and use those as seeds to generate more, which you post all over the web. There's no need to do any training - the generic model should be fine given the right prefix.