In this paper, a new syntax-based approach to belief revision is presented. It is developed within a nonmonotonic framework that allows a two-steps handling of inconsistency to be adopted. First, a disciplined use of non-monotonic ingredients is made available to the knowledge engineer to prevent many inconsistencies that would occur if a standard logical interpretation and representation of beliefs were conducted. Remaining inconsistencies are considered unexpected and revised by weakening the formulas occurring in any minimally inconsistent subbase, as if they were representing exceptional cases that do not actually occur. While the computation of revised knowledge bases remains intractable in the worst case, our approach benefits from an efficient local search-based heuristic technique that empirically proves often viable, even in the context of very large prepositional applications.