Enhancing CLIP with GPT-4: Harnessing Visual Descriptions as Prompts