In this paper a new pruning method for a rule-based parser is described that relies on separating the underlying grammar rules into several mutually competing levels. This method has been developed and exploited for Czech in the syntactic parser Synt to reduce the number of possible output derivation trees. The algorithm behind operates on a so called packed forest of trees, a compressing data structure used for internal representation of parallel analyses, and thus performs very effectively. An evaluation of its contribution has been performed on the Brno Phrasal Treebank showing that the algorithm significantly prunes the resulting tree space while preserving perspective parses.