Best practice under scrutiny: Prompt structure and the underestimation of LLM performance

Large language models (LLMs) are commonly evaluated and deployed using tightly structured prompting, based on the assumption that such constraints improve output reliability and quality. This study tested that assumption across multiple models and 25 historical topics, comparing structured prompts (A-conditions) with less structured, relational prompts (B-conditions), and a compressed summary condition (C). Across all models, structured prompting did not improve performance. Under less constrained prompting, factual output increased substantially by approximately 3.5 to 11.5 times, without loss of coherence. Instead, responses showed a qualitative shift: facts were embedded within integrated explanations rather than presented as isolated statements. Hallucinations were only detected under A-conditions. This contradicts the assumption that longer outputs increase the risk of factual errors. Under B-conditions, one model produced approximately 82,000 words without observed hallucinations. C-condition results further showed that structure alone did not account for these differences. Although C-conditions imposed the same format constraints as A-conditions, they were based on summaries of outputs generated under B-conditions. Despite this limitation, outputs matched A-condition outputs quantitatively, while retaining qualitative improvements in structure and coherence. The findings indicate that LLM outputs are strongly condition-dependent. This has direct implications for both evaluation and use. Current practices based on constrained prompting may systematically underestimate model capabilities, particularly for tasks requiring explanation, integration, and contextual understanding. In applied settings, differences of this magnitude may substantially affect how effectively LLM systems perform.