That was always my assumption of the end game. You have the system prompts, an advertising bias prompt layer over top, then the user prompts.
“Naturally worded” advertising that doesn’t immediately appear to be advertising and searching using natural language always seemed to be the biggest use cases for LLMs to me, considering they can’t be relied on to output accurate info.

















You should have a look at some of the official PowerShell scripts they’ve made to assist with managing, installing, and decommissioning Exchange Servers.
One had the word group misspelled consistently the entire way through it. Another had a wonderful comment noting that an internal function (to manage something with AD permissions or deep internals of the Exchange Server) that looked misspelled was in fact spelled correctly, because the command had a typo in the real name of it.