Cranial implant design is aimed to repair skull defects caused by brain related diseases like brain tumor and high intracranial pressure. Researches have found that deep neural networks could potentially help accelerate the design procedure and get better results. However, most algorithms fail to handle the generalization problem: deep learning models are expected to generalize well to varied defect patterns on high-resolution skull images, while they tend to overfit to some specific defect patterns (shape, location, etc.) in the training set. We employ principle components analysis (PCA) to model the shape of healthy human skulls. We assume that defective skulls have similar shape distributions to healthy skulls in a common principle component (PC) space, as a defect, which usually only occupies a fraction of the whole skull, would not substantially deviate a human skull from its original shape distribution in a compact PC space. Applying inverse PCA to the principal components of defective skulls would therefore yield their healthy counterparts. A subtraction operation between the reconstructed healthy skulls and the defect skulls is followed to obtain the final implants. Our method is evaluated on the datasets of Task 2 and Task 3 of the AutoImplant 2021 challenge (https://autoimplant2021.grand-challenge.org/ ). Using only 25 healthy skulls to create the PCA model, the method nonetheless shows satisfactory results on both datasets. Results also show the good generalization performance of the proposed PCA-based method for skull shape modelling. Codes can be found at https://github.com/1eiyu/ShapePrior.