A method for processing wafers includes learning a first pattern at a de-skew site on a first wafer layer, saving the first patterns in a recipe for de-skewing wafers, learning a second pattern at the de-skew site a second wafer layer, and saving the second pattern in the same recipe for de-skewing wafers. Learning the first pattern may include determining a score of uniqueness for the first pattern. The method further includes finding the de-skew site on the second wafer layer using the first pattern before learning the second pattern. Finding the de-skew site includes determining a score of similarity between the first pattern and the second pattern. Learning the second pattern is performed when the score of similarity is less than a threshold value. A recipe for de-skewing wafers includes multiple patterns of a de-skew site of a wafer, wherein the patterns include a first pattern at the de-skew site on a first wafer layer and a second pattern at the de-skew site on a second wafer layer.