WiFi-based action recognition has become increasingly powerful due to rapid advancements in deep learning, especially for applications in smart homes, health monitoring, and security. However, existing studies frequently lack detailed guidelines for platform setup and dataset construction, and they seldom discuss the impact of window length and latency on recognition accuracy. To address these gaps, this study explores the implementation of platforms and the construction of datasets, incorporating movement sequences of varying levels of action completeness to simulate different recognition latencies. We propose a high-accuracy algorithm, named CNN-ResNet-BiLSTM (C-R-BL), which significantly improves recognition accuracy while ensuring timely responses. To validate the effectiveness and superiority of the proposed algorithm, we conduct ablation and comparative experiments. Results indicate that the algorithm achieves accuracy rates ranging from 93.8% to 96.0% under low-latency conditions, demonstrating significant performance enhancements. This study provides a practical guide for new researchers and emphasizes the importance of these factors in WiFi-based action recognition.