This study examined the written work of students in the introductory calculus-based electricity and magnetism course at the University of Arkansas. The students' solutions to hourly exams were divided into a small set of countable features organized into three major categories, mathematics, language, and graphics. Each category was further divided into subfeatures. The total number of features alone explained more than 30% of the variance in exam scores and from 9% to 15% of the variance in conceptual posttest scores. If all features and subfeatures are used, between 44% and 49% of the variance in exam scores is explained and between 22% and 28% of the variance in conceptual posttest scores. The use of language is consistently positively correlated with both exam performance and conceptual understanding. (Contains 5 tables.)