Fragment assembly is attracting intense research interest as a means of predicting protein structure from sequences. Success requires a good library of local structures (i.e. fragments) to serve as building blocks for whole protein reconstruction. Here, we describe a novel two-stage clustering scheme for identifying recurrent local structures and deriving fragment libraries. Parameters that represent cluster compactness, distinctiveness, and coverage were used for evaluating library quality. According to test results, our system was capable of approximating protein structures with smaller library sets at a level of accuracy comparable to or better than methods previously reported by others.