Understanding and supporting how developers prompt for LLM-powered code editing in practice

Despite the rapid adoption of LLM-based coding assistants, real-world developer prompting behavior remains underexplored. This paper investigates how developers interact with an LLM-powered code editing feature within a large-scale industrial IDE. Through a multi-phase study of telemetry logs, we find that frequent re-prompting signals significant developer struggle in communicating intent. We perform a qualitative analysis of unsatisfactory requests to derive a taxonomy of five common under-specification categories. Leveraging these insights, we develop AutoPrompter, a pipeline that automatically enhances sparse developer prompts by inferring missing context. Our evaluation on a real-world dataset shows that AutoPrompter improves code edit correctness by 27\% without additional user input.