Abstract: A CNDO/2 geometry optimization program is written in the BD-200 algorithmic language for DJS-18 type computers with some study of algorithmic improvement. The optimization method proposed by Davidon is adopted with the parameter β variable versus the p-value instead of a fixed β. The density matrix of the old cycle in the optimization iteration is retained as an initial density matrix instead of a new Huckel guess for the new cycle. Variable energy tolerances are used in self-consis-tcnt-field iterations to provide increassing precision of the density matrices. These improvements result in high reliability and comparatively low computation time.