A brief review is made of empirical methods for prediction of three aspects of protein conformation: (i) secondary structures, (ii) packing of secondary structures and (iii) surface and interior of globular proteins. The physical assumptions implicit in these empirical methods are analyzed. To explain the effectiveness of these empirical methods, "a consistency principle" is proposed which asserts that the various conformational energy terms contributing to the stability of native conformations are consistent with each other and therefore individually consistent with the native conformations. The low resolving power of the short-range interactions necessitates a prediction algorithm with a feedback mechanism by which the long-range interactions are involved in the determination of local structures. Such an algorithm should appear to simulate a folding process in which the intrinsic instability of small local structures serves as a built-in mechanism to overcome the low resolving power of the short-range interactions.