The growing international container logistics industry has even intensified the need of efficient and easy to implement on a local scale Decision Support Systems (DSS). Although LLMs have a significant potential application, their implementation through conventional methods like Retrieval-Augmented Generation (RAG) or fine-tuning impose significant costs, data, and expertise limitations especially for small and medium logistics companies. This research focuses on exploring the ability of Few-shot In-Context Learning (ICL) to tune a performant base LLM-Gemma-3-27b-it to intricate logistics reasoning problems without the adequacy of these dependencies. Carefully designed dataset of more that 300 cases scenarios, based on real-life of the seven global ports, was synthesized. LLM-as-a-Judge (Gemini-2.5-pro-exp-03–25) was introduced to test Gemma-3 under a Baseline and a Few-shot ICL condition over forty runs where solutions were scored using a 10-point scale regarding logical soundness, explanatory quality, and structure clarity. The results indicate that Few-shot ICL significantly surpassed the baseline, especially, on the Explanatory Quality criterion, where the average increased from 7.20 up to 8.43 (+17.1%). That has resulted in the overall quality score being higher (8.81 vs. 8.27) and the success rate being higher (88.64% vs. 79.55%). The paper illustrates that sophisticated prompt engineering is a powerful and lightweight method of improving the reasoning and explanatory capacities of effective LLMs. It offers a feasible solution to reaching trustworthy and locally implementable AI assistance systems capable of providing reliable and explainable decision support in the logistics domain.