This paper presents our approach to Subtask 1 of SemEval-2026 Task 9 on multilingual polarization detection in social media texts in English and Spanish. We model the task as a prompt-based binary classification problem and systematically compare zero-shot, one-shot, and few-shot strategies across multiple large language models accessed via commercial APIs, without task-specific fine-tuning. Our controlled experimental setup enforces strict data separation and consistent decoding conditions to analyze the impact of in-context supervision across architectures and languages. Results indicate that well-structured prompting enables competitive performance, though implicit and culturally nuanced polarization remains challenging.