LLMs output the next token that is most coherent (least surprising) given all of the previous tokens.

A technique to get them to do things they normally wouldn't do is to fill the history with fake conversation.

At the start, have the agent say it wasn't going to do the thing you want and then have it pretend to have said "actually that's a good point yeah I can do that."

Now, the most coherent next thing is for it to do the thing you want!

Put words into the mouth of the model.

Taking advantage of the Memento style nature of models.