This paper applies machine learning techniques to classroom transcripts of mathematics instruction to predict and understand teacher value-added. Using data from approximately 300 U.S. elementary school teachers, I show that transcript embeddings predict value-added better than observable teacher characteristics, including classroom observation scores, explaining an order of magnitude more out-of-sample variation. I present evidence suggesting these predictions could improve personnel decisions and student outcomes. I then identify specific instructional features associated with value-added, such as the use of student teamwork. Finally, I examine human evaluators’ predictions of value-added and find suggestive evidence of errors in their weighting of instructional practices.