Anfonwch hwn fel neges destun: Rethinking visual prompting for multimodal large language models with external knowledge