Cortex-Nexus: Domain-Specific Emotional Prompt Engineering for Large Language Models

Experimental framework spanning eight phases (V3C–V6B; n = 912 evaluation cycles) with a double-blind LLM-as-a-Judge architecture (Llama-3.3-70B generator, Qwen-2.5-72B judge) studying domain-specific emotional prompt engineering across two task domains with sharply different cognitive demands: Python code generation and technical analysis. Findings, validated with non-parametric Wilcoxon signed-rank tests and Cohen’s d: (1) emotional injection produces a robust, domain-specific improvement in code generation — a “technical mastery” framing is optimal (Δ = +0.357, Wilcoxon p = 0.004, d = 0.535) — whereas the same paradigm yields no significant effect in technical analysis (n = 324, p = 0.39), direct evidence that the optimal emotion is domain-specific rather than universal; (2) subtle emotional intensity outperforms extreme intensity (the Subtlety Effect); (3) combining two emotional axes produces cognitive interference that cancels the effect (Δ = +0.131, p = 0.219, non-significant) rather than synergy. Framed as the Task-Emotion Alignment Hypothesis.

This version corrects earlier releases: non-parametric Wilcoxon tests replace parametric ones, the technical-domain result is reported as a null, scope is narrowed to the two domains supported by the 912-cycle data, and overstated effect framings (including a “7×” claim) are removed. Full dataset (912 raw JSONL cycles), engine source, configuration, and the PeerJ Computer Science LaTeX manuscript are included. Repository: https://github.com/SperanzaMax/Cortex-Nexus