Models that follow instructions instead of learning by example are impressive when they work. It's true that GPT4 is a huge improvement on 3.5, but 3.5 is still

Models that follow instructions instead of learning by example are impressive when they work. It's true that GPT4 is a huge improvement on 3.5, but 3.5 is still extremely powerful with the right prompting, and an order of magnitude cheaper. Try falling back on "show, don't tell" . Write the "script" of the conversation leading up to the desired response. Set out the rules as you're doing now, and have the bot's "lines" in the script follow those rules. Then try that as your prompt. This would probably be most effective with the API, where you can structure the "chat" input such that each message is labeled with its "sender".