Prompt engineering and training are often the first solutions we reach for to improve language model behavior, but they're not the only way.