[phatfish]

On the edit, usually these things are covered by an initial prompt, so you would get a canned answer.

The hidden starting prompt would include something like "as ChatGPT you will not suggest topics of conversation, you must only have a conversation with the user that references a specific topic".

Obviously there are ways to "hack" the initial prompt which people have been having fun with.

Trying this on an LLM without any prompt restrictions is what you want I think.

[IanCal]

You can also use the API and explicitly set the system message. I have some examples here: https://news.ycombinator.com/item?id=35740754