Parallel kinematic machines (PKMs) have the advantages of a compact structure, high stiffness, a low moving inertia, and a high load/weight ratio. PKMs have been intensively studied since the 1980s, and are still attracting much attention. Compared with extensive researches focus on their type/dimensional synthesis, kinematic/dynamic analyses, the error modeling and separation issues in PKMs are not studied adequately, which is one of the most important obstacles in its commercial applications widely. Taking a 3-(P)RS parallel manipulator as an example, this paper presents a separation method of source errors for 3-DOF parallel manipulator into the compensable and non-compensable errors effectively. The kinematic analysis of 3-PRS parallel manipulator leads to its six-dimension Jacobian matrix, which can be mapped into the Jacobian matrix of actuations and constraints, and then the compensable and non-compensable errors can be separated accordingly. The compensable errors can be compensated by the kinematic calibration, while the non-compensable errors may be adjusted by the manufacturing and assembling process. Followed by the influence of the latter, i.e., the non-compensable errors, on the pose error of the moving platform through the sensitivity analysis with the aid of the Monte-Carlo method, meanwhile, the configurations of the manipulator are sought as the pose errors of the moving platform approaching their maximum. The compensable and non-compensable errors in limited-DOF parallel manipulators can be separated effectively by means of the Jacobian matrix of actuations and constraints, providing designers with an informative guideline to taking proper measures for enhancing the pose accuracy via component tolerancing and/or kinematic calibration, which can lay the foundation for the error distinguishment and compensation.