This paper examines how design can support the collection of AI training data during mental health consultations. Drawing on Interactive Machine Teaching, the researchers worked with clinicians to develop and refine an interface for labeling clinical judgments as they happen. Qualitative analysis explored how the interface helped clinicians make their reasoning explicit and organize data creation. The study also discusses practical limits, including potential disruption to conversations and the importance of non-verbal information. It proposes design directions for recording expert decisions in human-centered AI development.