Large language model (LLM) behavior is shaped at use time through prompting rather than through model training. Yet, in Design Science Research (DSR), prompts are often treated as incidental implementation details: prompt design rationales, iterations, and evaluation conditions remain under-specified, which weakens attribution and limits cumulative knowledge building. While prompt engineering provides techniques for crafting effective prompts, it mainly focuses on prompt structure and model performance, while neglecting domain-informed development within the broader socio-technical context. This paper conceptualizes prompts as DSR artifacts and translates established DSR knowledge into guidance for prompts across problem formulation, design, and evaluation in non-deterministic LLM environments. We propose a two-level conceptualization that comprises task support (user expectations, task boundaries, context, and constraints) and prompt specification (the reusable instruction structure that encodes these elements). We conclude by outlining a research agenda for prompts as design science artifacts.