Always give good examples in the prompt, preferably a few (2-4). Still not 100% reliable though, especially with models below GPT-4.