This research introduces CORSE (Contextual Orchestration and Reward Shaping Engine), a novel decision-support architecture that integrates structured prompt routing with quantitative reward-based evaluation to enhance the reliability of large language model (LLM) recommendations. Anchored in a multi-cycle Action Design Research (ADR) methodology, the study first diagnoses fragilities in generic LLM prompting, then develops and formalizes CORSE’s modular architecture—including context-aware orchestration (CO), a reward shaping evaluation (RSE) framework, and iterative refinement mechanisms. A complete prototype is implemented and experimentally validated in 50 synthetic financial advisory scenarios using LLaMA 3.3 and Claude 4.5. CORSE significantly outperforms baseline and partial variants in reward scores, pass rates, and failure correction, offering measurable and auditable improvements in decision quality. The study derives finance-grounded design principles for context-sensitive LLM decision support; applicability beyond financial advising remains a proposition for future domain-specific adaptation and validation. This work contributes a transparent, metrics-driven design artifact that demonstrates how contextual orchestration and reward shaping can be integrated into a closed-loop architecture to improve the reliability and auditability of LLM-based decision support.