We devised annotation guidelines for the de-identification of German clinical documents and assembled a corpus of 1,106 discharge summaries and transfer letters with 44K annotated protected health information (PHI) items. After three iteration rounds, our annotation team finally reached an inter-annotator agreement of 0.96 on the instance level and 0.97 on the token level of annotation (averaged pair-wise F1 score). To establish a baseline for automatic de-identification on our corpus, we trained a recurrent neural network (RNN) and achieved F1 scores greater than 0.9 on most major PHI categories.