This paper describes the difficulties in the process of using the trial-and-error SYSTEM90 program to determine ab initio the structures of two small proteins App [Woolfson & Yao (1990). Acta Cryst. A46, 409–413] and rubredoxin [Sheldrick et al. (1993). Acta Cryst. D49, 18–23] with high-resolution data. Some strategies for overcoming the difficulties are discussed and the upgraded SYSTEM95 program was used successfully to determine the two structures. The most characteristic feature of this structure-determination process is that the two proteins are treated as unknown structures with only their chemical compositions and high-resolution data sets known. A new figure of merit R(sc), replacing the old figure of merit, XDFOM, is quite effective in picking out a good set of phases in the multi-solution stage when the phases are overconsistent. Controlling the Fourier recycling technique and the residuals can separate the mixture of structures and the enantiomorph and finally give one absolute structure. The results are compared with known structures to verify their reliability.