This won’t help in zero shot but if you have in-context examples try splitting each example response into two half-messages, so it appears assistant responses are expected come in pairs with a semantically meaningless split.

→ View original post on X — @goodside