In commonsense question answering (CSQA), pre-trained language models are often used to generate background knowledge that provides clues to improve the interpretability and performance of CSQA models. However, these methods are often ineffective for generating clues to distinguish similar options, and they also have limitations in using generative capabilities to solve CSQA tasks. In this paper, we propose a new model for CSQA that can differentiate the options for a question more effectively. This model leverages the generative capabilities of the pre-trained model to better understand the relationship between question and candidate options, creating clues that aid in effectively distinguishing choices for the question. Moreover, a mechanism is introduced to encourage the model to discern distinctions among options. Our proposed model demonstrates superior performance on four CSQA datasets, outperforming comparable models.