You have to tailor your inputs to be closer to the machine's language, rather than expecting it to answer to a natural language input by making the right inferences of intent, context, code switches etc.
Recognize it for what it is.
A symbolic manipulator engine.
Using natural language causes linguistic drift to natural language spaces.
If you don't want to do all that, simply be more specific using priming tokens such as context, intent, seed(s) (token(s) that control token domain focus).