Enviar aquest missatge de text: Rethinking visual prompting for multimodal large language models with external knowledge