A structured LLM prompt engineering framework for generating quality,performant simulation code for classical and quantum computing

Abstract

Reliable, performant code is essential in modeling and simulation, and integration of AI-generated code into simulation software development must not undermine the stability or integrity of models. Generative AI produces variable outputs, resulting in LLM-generated code that does not reliably satisfy standards of correctness or performance. This paper presents an LLM prompt engineering framework, Goal, Performance, Exclusion Architecture (GPE-A), that structures intent, constraints and performance criteria to guide the generation of reliable, performant simulation code. Rather than attempting reproducibility of an implementation, the framework steers generative outputs toward convergence on required behavioral metrics, consistent with requirements-driven development. The framework is evaluated using LLM-generated random number generation code as a representative simulation component across classical and quantum computing. Random number generation is a foundational simulation primitive and domain where quantum methods allow a comparable proof-of-concept implementation against classical implementations. Metric-based comparisons are made to production baselines, assessing correctness, statistical stability, and computational performance. The GPE-A is evaluated across multiple LLMs, through prompt ablation, and against modern prompting methods. Results indicate that structured LLM prompt engineering can increase predictability and quality in AI-generated simulation code, exceeding production-standard baselines for random number generation, and indicate potential extensibility toward emerging computational paradigms.