Automating Potential-based Reward Shaping with Vision Language Model Guidance