Conversational breakdowns often force users to go through frustrating loops of trial and error when trying to get answers from chatbots. Although research has emphasized the potential of conversational repair strategies in helping users resolve breakdowns, design knowledge for implementing such strategies is scarce. To address this challenge, we are conducting a design science research (DSR) project to design effective repair strategies that help users recover from conversational breakdowns with chatbots. This paper presents the first design cycle, proposing, instantiating, and evaluating our first design principle on identifying the cause of conversational breakdowns. Using 21,736 real-world user messages from a large insurance company, we conducted a cluster analysis of 5,668 messages leading to breakdowns, identified four distinct breakdown types, and built a classifier that can be used to automatically identify breakdown causes in real time. Our research contributes with prescriptive knowledge for designing repair strategies in conversational breakdown situations.