How do practitioners incorporate requirements and design decisions in LLM-assisted code generation?

This paper originally appeared in the IEEE 33rd International Requirements Engineering Conference (RE) UKV25. In an interview study with 18 participants from 14 companies, we examined how practitioners incorporate information from requirements and software design artifacts during LLM-assisted development. Based on our findings, we propose a theory that explains the processes developers employ, the artifacts they rely on, and the collaboration patterns that emerge. Our results indicate that requirements, as typically documented, are too abstract for direct input into LLMs. Instead, they must be manually decomposed into programming tasks and enriched with design decisions and explicit architectural constraints. Otherwise, LLMs tend to produce solutions requiring substantial modifications. We also identify three modes of LLM interaction in practice: technical exploration (to investigate solution approaches), code generation (LLMs produce most of the code, with developers correcting parts), and manual coding (developers program while using LLMs for auto-completion). Our theory is important for contextualizing scientific approaches to automating requirements-and design-centric SE tasks, and for guiding companies on how to integrate LLMs in their software engineering processes.