Accurate clinical diagnosis is of vital importance in the medical field. Large Language Models (LLMs) usually need to be fine-tuned and improved before they can be applied in the medical vertical field, and this process means additional data annotation and huge computing power overhead. Small Language Models (SLMs) can serving as an efficient alternative to fine-tuning LLMs through model distillation. However, during the distillation process, the teacher models often fail to stimulate their own vertical domain knowledge due to the lack of guidance, thereby causing the performance bottleneck problem of the student model. Aiming at the problems of incomplete internal knowledge stimulation of the teacher model and low quality of knowledge samples in model distillation, this paper proposes a zero-shot knowledge distillation framework PeKDiL (Prompt-enhanced Knowledge Distillation with Loss-based filtering) for Chinese clinical diagnosis, aiming to enhance the performance of SLMs through prompt engineering and loss-based data filtering, meanwhile reducing training costs and avoid manual annotation. This framework first guides LLMs to generate higher-quality knowledge through prompt engineering, and then uses a detection model to screen the confidence level of the generated knowledge data, retaining high-quality samples for the final model distillation. The experimental results show that SLMs distilled by PeKDiL perform better than general LLMs and other methods, verifying the effectiveness of this framework in improving the diagnostic accuracy and generalization ability of the model.