Beyond Captioning: Task-Specific Prompting for Improved VLM Performance in Mathematical Reasoning