Traditional Chinese medicine (TCM) prescription recommendation methods are favored in the era of Artificial Intelligence (AI). Recent methods ignore the problem of imbalanced herb labelling data and lack of attention to less frequently occurring principal medicines, which is not in line with the medication patterns of clinical TCM practitioners. More importantly, it has been observed that existing methods relying on graph learning may introduce unsuitable herb embeddings as they focus on capturing symptom-herb relationships. Therefore, in this paper, we propose a knowledge graph-guided diffusion model (DM) for TCM prescription recommendation (TCM-KGDR). We first utilize the TCM knowledge graph-guided diffusion module (TCM_Diff) for integrated modelling of TCM representations. Then, we leverage the Reinforcement Learning (RL)-conditioned TCM community detection module (TCM_CRL) to analogize herbs into communities from the node level based on the integrated similarity score to consider the similarity and compatibility among herbs. Finally, we conduct the diffusion-guided TCM graph denoising module (TCM_DEN) for symptom-herbal medicine relationships to learn their patterns of collaboration from the edge level. Compared with several baseline models on two real datasets, extensive experiments demonstrate the effectiveness of our method while it can provide decision support for clinical herb prescription recommendation (HPR) methods.