Fill-in-the-blank questions are one of the main assessment devices in REAP.PT tutoring system. The problem of automatically generating the stems, i.e. the sentences that serve as basis to this type of question, has been studied mostly for English, and it remains a challenge for a language as morphologically rich as European Portuguese (EP), for which additional data scarcity problems arise. To address this problem, a supervised classification technique is used to model a classifier that decides whether a given sentence is suitable to be used as a stem in a cloze question. The major focus is put in the feature engineering task, describing both the development of new criteria, and the adaptation to EP of features already explored in the literature. The resulting classifier filters out inadequate stems, allowing experts to build and personalize their instruction focusing on a set of potentially good sentences.