A challenge in computational protein folding is to assemble secondary structure elements—helices and strands—into well-packed tertiary structures. Particularly difficult is the formation of β-sheets from strands, because they involve large conformational searches at the same time as precise packing and hydrogen bonding. Here we describe a method, called Geocore-2, that (1) grows chains one monomer or secondary structure at a time, then (2) disconnects the loops and performs a fast rigid-body docking step to achieve canonical packings, then (3) in the case of intrasheet strand packing, adjusts the side-chain rotamers; and finally (4) reattaches loops. Computational efficiency is enhanced by using a branch-and-bound search in which pruning rules aim to achieve a hydrophobic core and satisfactory hydrogen bonding patterns. We show that the pruning rules reduce computational time by 103- to 105-fold, and that this strategy is computationally practical at least for molecules up to about 100 amino acids long.