A method of automatic optimization of an examination recipe includes obtaining inspection data of a given layer of a semiconductor specimen acquired by an inspection tool during runtime examination, the inspection data including inspection images representative of defect candidates from a defect map of the given layer, extracting inspection features characterizing the inspection images, and using a classifier to classify the defect candidates based on the inspection features, giving rise to a list of defect candidates having a higher probability of being defects of interest (DOIs). The semiconductor specimen includes multiple layers, and the classifier is a general-purpose classifier (GPC) usable for runtime classification of inspection data from any layer of the multiple layers of the semiconductor specimen, the GPC being previously trained using training data including inspection features characterizing training inspection images of various types of DOIs and nuisances collected from the multiple layers and label data associated therewith.