Your system prompt isn't instructions. It's data.

In a detailed exploration of LLM behavior, developer Natuworkguy argues that system prompts are processed by models as data rather than rigid instructions. Through iterative testing of a 31B parameter model, the author discovered that models often treat examples and phrasing as patterns to replicate rather than guidelines to follow. Key findings include the observation that negative constraints often fail, while concrete examples and the physical positioning of rules within a prompt significantly impact output quality. The author emphasizes the importance of using pinned seeds for reproducibility and implementing automated testing suites to verify model performance. By treating system prompts as data samples, developers can better understand how to structure instructions to avoid common pitfalls like style bleeding or incorrect code generation. The article concludes by encouraging developers to move problematic rules to the top of their prompt sections and prioritize empirical testing over subjective evaluation.