How can we explain that people are capable of performing new tasks with no or little instruction? Earlier work has proposed that new tasks can be acquired by a rapid composition of cognitive skills, and implemented this in the ACT-R and PRIMs cognitive architectures. Here, we discuss a possible application of rapid composition in building tutoring systems. The goal is to identify underlying skills through unsupervised machine learning from a dataset of arithmetic learning for students in a Dutch vocational program. The resulting skill graph is used as a basis for a tutoring system. The results show evidence for predictive power of the system and tentative evidence of a learning benefit compared to control groups.