Easing Cloud‑Scale Software Development with LLM‑Powered Architecture Guidance

In “LLMs Choose the Right Stack: From Patterns to Tools,” we investigate whether today’s large language models can help engineers navigate the complex task of picking appropriate architectural patterns and the corresponding technologies for distributed / cloud‑scale systems. By testing six models (five open‑source and one closed‑source) across three realistic scenarios—plain versus prompt‑engineered pattern recommendation, decision‑tree‑guided selection using our CAPI method, and mapping patterns to a curated list of tools—we find that even minimal prompts yield reasonable suggestions, while prompt engineering and CAPI guidance markedly improve focus on high‑level architectural decisions and bring model performance close to that of human experts. The study also uncovers systematic biases, such as an over‑emphasis on micro‑service architectures, and highlights the need for tighter output control and broader pattern coverage. These results demonstrate that LLMs, when combined with lightweight prompting techniques, can serve as valuable assistants in architectural decision‑making, opening new research directions for AI‑augmented software design.