LLMs Choose the Right Stack: From Patterns to Tools

Choosing suitable architectural patterns and the technologies that implement them is a complex design task. We evaluate how well current LLMs can support such decisions by empirically evaluating six LLMs (five open-source, one closedsource) on three scenarios: (i) naïve versus prompt-engineered pattern recommendation, (ii) decision-tree-guided selection via the CAPI method, and (iii) mapping patterns to concrete tools from a supplied list. We assess reasonableness, consistency, pattern specificity, and output structure. We show that even minimal prompts yield reasonable suggestions, while prompt engineering improves focus on architectural (rather than lowlevel design) patterns and consistency. CAPI guidance expands coverage and approaches human-expert performance, though models exhibit a strong bias toward micro-services and tend to over-suggest patterns. All models propose plausible tools when a curated list is provided. Overall, LLMs-especially when combined with structured prompts and decision-tree guidancecan meaningfully augment architectural decision-making, while highlighting the need for tighter output control and broader, less biased pattern coverage.