Base editors (BEs) hold great potential for gene therapy. However, high precision base editing requires BEs that can discriminate between the target base and multiple bystander bases within a narrow active window (4 - 10 nucleotides). To assist in the design of these optimized editors, we propose a discrete-state stochastic approach to build an analytical model that describes the probabilities of editing the target base and bystanders. Combined with all-atom molecular dynamic simulations, our model well reproduces the experimental data of A3A-BE3 and its variants for target and bystander editing. Building upon this model, we propose several general principles that can guide the design of BEs with a reduced bystander effect. We used these principles to improve the A3G-BEs with high precision and verified their base-editing activities experimentally. In summary, our study provides a computational-aided platform to assist in designing BEs with reduced bystander effects.