A large number of energy points add great difficulty to construct reactive potential energy surfaces (PES). To alleviate this, exemplar-based clustering is applied to partition the configuration space into several smaller parts. The PES of each part can be constructed easily and the global PES is obtained by connecting all of the PESs of small parts. This divide and conquer strategy is first demonstrated in the fitting of PES for OH3 with Gaussian process regression (GPR) and further applied to construct PESs for CH5 and O +CH4 with artificial neural networks (NN). The accuracy of PESs is tested by fitting errors and direct comparisons with previous PESs in dynamically important regions. As for OH3 and CH5, quantum scattering calculations further validate the global accuracy of newly fitted PESs. The results suggest that partitioning the configuration space by clustering provides a simple and useful method for the construction of PESs for systems that require a large number of energy points.