Small-probe contact dot surface analysis, with all explicit hydrogen atoms added and their van der Waals contacts included, was used to choose between the two possible orientations for each of 1554 asparagine (Asn) and glutamine (Gln) side-chain amide groups in a dataset of 100 unrelated, high-quality protein crystal structures at 0.9 to 1.7 A resolution. For the movable-H groups, each connected, closed set of local H-bonds was optimized for both H-bonds and van der Waals overlaps. In addition to the Asn/Gln "flips", this process included rotation of OH, SH, NH3+, and methionine methyl H atoms, flip and protonation state of histidine rings, interaction with bound ligands, and a simple model of water interactions. However, except for switching N and O identity for amide flips (or N and C identity for His flips), no non-H atoms were shifted. Even in these very high-quality structures, about 20 % of the Asn/Gln side-chains required a 180 degrees flip to optimize H-bonding and/or to avoid NH2 clashes with neighboring atoms (incorporating a conservative score penalty which, for marginal cases, favors the assignment in the original coordinate file). The programs Reduce, Probe, and Mage provide not only a suggested amide orientation, but also a numerical score comparison, a categorization of the marginal cases, and a direct visualization of all relevant interactions in both orientations. Visual examination allowed confirmation of the raw score assignment for about 40 % of those Asn/Gln flips placed within the "marginal" penalty range by the automated algorithm, while uncovering only a small number of cases whose automated assignment was incorrect because of special circumstances not yet handled by the algorithm. It seems that the H-bond and the atomic-clash criteria independently look at the same structural realities: when both criteria gave a clear answer they agreed every time. But consideration of van der Waals clashes settled many additional cases for which H-bonding was either absent or approximately equivalent for the two main alternatives. With this extra information, 86 % of all side-chain amide groups could be oriented quite unambiguously. In the absence of further experimental data, it would probably be inappropriate to assign many more than this. Some of the remaining 14 % are ambiguous because of coordinate error or inadequacy of the theoretical model, but the great majority of ambiguous cases probably occur as a dynamic mix of both flip states in the actual protein molecule. The software and the 100 coordinate files with all H atoms added and optimized and with amide flips corrected are publicly available.