Humans have the amazing capacity to perform new tasks with little or no instruction. To explain this remarkable ability, I propose that people, when faced with a new task, compose the necessary knowledge for that task using cognitive skills as building blocks. In our cognitive modeling research, we have shown how a small set of skills can instantiated into a variety of task models, and provide explanations for phenomena such as attentional blink and task switching costs without having to rely on assumptions about limitations of the brain. If cognitive skills are the building blocks of cognition, is important to be able to identify them, and study how they are learned. To identify cognitive skills, we use a hybrid approach, in which we use bottom-up machine learning methods to use individual differences in student performance to construct a knowledge graph, in which each node represents a combination of skills, and a possible knowledge state of the student. A first pilot study has shown the benefits of this approach in teaching arithmetic to students in the Dutch Vocational education.