When generating a reasoning, limit to a couple of sentences that are highly specified rather than our model's current interpretation which could be one or ten sentences.