Structured Prompt and Agent AI Framework for Reliable LLM-Based RC Beam Design

This study evaluates a Structured Prompt (S-Prompt) approach for LLM-based RC beam design under Korea Design Standard (KDS) provisions. Across 1,200 trials (two models, two prompt styles, and 300 runs per condition), S-Prompt increased end-to-end correctness, defined as the simultaneous correctness of $${M}_{n}$$ , $${V}_{n}$$ , and governing failure mode, from approximately 37% with a one-shot prompt to approximately 88–90%. For the benchmark case, S-Prompt solutions matched manual references within rounding, while hallucinations and omissions were substantially reduced. In an additional moment–curvature example, the S-Prompt-guided workflow showed close agreement with a structural analysis program. These findings indicate that structured prompting can improve the reliability and auditability of LLM-assisted RC beam calculations within a controlled benchmark setting.