Zeolites are an important class of tetrahedrally coordinated inorganic materials that have been widely used in many chemical industries as catalysts, adsorbents, and ion-exchangers. To date, over 200 types of zeolite framework have been discovered. Predicting not-yet-discovered zeolite frameworks is of great importance not only for zeolite structure determination but also for the identification of promising synthetic candidates with desirable functions. However, owing to the complexity and diversity of zeolite framework topologies, zeolite structure prediction has been a challenging task for several decades. Many efforts have been devoted towards this end, among which the computer-aided assembly of zeolite framework constituent atoms (T atoms) in predefined Wyckoff positions (WPs) is a promising approach because of its high efficiency in configuration space searching. However, this approach suffers from high computational overheads caused by the large number of possible WP combinations. On the basis of the analysis of known zeolite structures, we find that the site symmetries of many WPs are incompatible with the tetrahedral coordination of T atoms. Moreover, to avoid the formation of chemically unfeasible distorted tetrahedral coordination, T atoms cannot be too "crowded" in some WPs. We define, for the first time, the T atom distribution (TAD) densities for special site symmetries as the numbers of T atoms per special point, per unit length of rotation axes, or per unit area of mirror planes, respectively. By restricting the number of T atoms in every WP so as not to exceed the highest allowed TAD density, WP combinations for zeolite structure prediction can be reduced by 1-4 orders of magnitude. Taking advantage of this discovery, the efficiency of zeolite structure prediction based on the enumeration of WP combinations can be significantly improved.