It's actually reacting to what you say. Your action is appended onto the previous story and fed into the model to generate the next part of the result. Sometimes it does fail to tie the result close enough to the action you had, perhaps with more training data it could do better on that front.