An easy way of setting up the (utterance, response) pairs: we want to support conversational language for an atypical vocabulary, and we want the responses to be hard-coded by the user. (The internet has terrible meditation advice, we don't want an LM generating what to do.)