Analyzing the Flow of Developer Prompts in ChatGPT Conversations

This study explores how developers engage with ChatGPT across multiple turns, analyzing 645 prompts from real coding conversations. It categorizes developer messages into seven roles—ranging from introducing tasks to refining queries and giving feedback—and uses a Markov model to map their conversational flow. Results show that most dialogues begin with task definition, followed by iterative follow-ups or refinements that improve ChatGPT’s responses. These findings reveal how developers co-evolve solutions with AI through layered, feedback-driven exchanges rather than single-shot queries.