Which works better when making a job assignment in firms, rotation or specialization? To answer this question, we develop a dynamic firm model of matching workers with indivisible jobs as in an overlapping generations (OLG) matching model a la Kurino (2014). First, we build a simple benchmark model without OLG in which all workers are either under-training or fully trained in each period, and then we show that either the rotation or the specialization of jobs for a worker emerges from the firms' profit maximization. We extend the benchmark model to the OLG model in which workers under- and post-training coexist in each period. We show that the profit-maximizing allocation is either a rotation or a specialization in this extended model as well. Hence, in both the benchmark and the extended models, the rotation and specialization schemes are the only variations that can be optimal in terms of profits. Moreover, the rotation scheme is better when the training cost is smaller, the uncertainty about job continuation in the future is larger, or the slope of seniority wages is larger.