In model-based climate sensitivity studies, model errors may grow during continuous long-term integrations in both the “reference” and “perturbed” states and hence the climate sensitivity (defined as the difference between the two). To reduce the errors, we propose a piecewise modeling approach that splits the continuous long-term simulation into subintervals of sequential short-term simulations and updates the modeled states through re-initialization at the end of each subinterval. In the re-initialization processes this approach updates the reference state with analysis data and updates the perturbed states with the sum of analysis data and the difference between the perturbed and the reference states, thereby improving the credibility of the modeled climate sensitivity. We conducted a series of experiments with a shallow water model to evaluate the advantages of the piecewise approach over the conventional continuous modeling approach. We then investigated the impacts of analysis data error and subinterval length used in the piecewise approach on the simulations of the reference and perturbed states as well as the resulting climate sensitivity. The experiments show that the piecewise approach reduces the errors produced by the conventional continuous modeling approach, more effectively when the analysis data error becomes smaller and the subinterval length is shorter. In addition, we applied a nudging assimilation technique to solve possible “spin-up” problems caused by re-initializations using analysis data that contain inconsistent errors between mass and velocity. The nudging technique can effectively diminish the spin-up problem, resulting in a higher modeling skill.