Anything you can repeat can be improved. If success is in the realm of possibility, you can go from clueless to reliably getting results. Cause and effect is a marvel of the universe and humans can tap into it.

Customer support agents are incentivized to solve tickets quickly. If it's an edge case? Terrible experience. But if they're incentivized to discover the best knowledge instead? Edge cases get treated well + more tickets are solved fast too, as better knowledgebase = better AI.

Could ChatGPT write its messages to get a certain response from you? If every possible followup message you write in a thread were a token, could it try to elicit a specific token from you? Reverse prompt engineering.