Intuitive prompting lifts LLM fidelity on Serbian social media reactions from 3x to 7x human variance compression

Eight Serbian participants completed questionnaires, interviews and self-presentations. Their reactions to 68 posts were recorded. Four models received the same profiles under five prompt regimes that varied attitudinal depth and reasoning style. Attitudinal content alone outperformed demographic backstories. Intuitive instructions produced the single highest match rate and narrowed the gap between agent and human variance from sevenfold compression to threefold. The result held on posts covering unprofiled topics, where intuitive agents exceeded the crowd baseline by the largest margin recorded. Consistency between agent and stated profile proved independent of fidelity once profile data was supplied. This pattern indicates that reduced analytical scaffolding preserves individual signal rather than averaging it toward training priors. Related work on reasoning traces shows similar degradation when chain-of-thought is applied to preference or affect tasks. The Serbian study supplies the first direct comparison of intuition versus analysis inside a within-subject social simulation design. Operational consequence is that platforms can maintain smaller, topic-agnostic agent panels instead of retraining specialists for each policy domain. Next steps include replication on non-Serbian cohorts and measurement of downstream effects on A/B tests of moderation rules. Threshold for deployment relevance is agent-human agreement exceeding 0.75 Spearman on held-out posts within a single platform.