Due to the complexity of 12-lead signals, the diversity of cardiac conditions, and the semantic gap between waveforms and clinical language, existing large language model (LLM)-based electrocardiogram (ECG) report generation remains challenging. In this paper, we propose a novel approach to ECG report generation through diagnostic knowledge-enhanced prompt learning. Specifically, a knowledge-aware module is constructed to extract waveform features and diagnostic cues via multilabel classification. These clinical semantic features are then fused with textual descriptions to form input prompts, thereby enhancing the semantic richness of the prompts. Moreover, by incorporating signal augmentation to capture fine-grained waveform semantics, we perform both intra-modal and cross-modal alignment between high-dimensional ECG signals and the generated text during model training, thereby improving the accuracy and relevance of report generation. A constraint-aware loss is further introduced to ensure the inclusion of essential diagnostic elements. Experiments on benchmark datasets demonstrate that our proposed method achieves superior performance on natural language generation metrics (e.g. BLEU, CIDEr-D), validating its effectiveness in both generation accuracy and clinical relevance. The code is available at: https://github.com/pangpanqiqi/ECG-Report-Prompt .