Recent studies have used prompt learning to enhance few-shot NER, focusing on template and label word design. However, exhaustive entity enumeration in templates and unrepresentative label words can limit time efficiency and performance. This paper introduces TLprompt, a prompt-based NER model that addresses these challenges with EPCT (Entity Position de-Construction Templates) and DaPreMD (Dataset and Pre-trained Model Mixture Distribution). EPCT locates entities with a pre-trained model and creates templates based on these positions. DaPreMD combines dataset and model-predicted distributions for more representative label words. Experiments on two NER benchmarks show that TLprompt achieves 1.49 to 1927.74 times faster inference with superior performance, balancing speed and accuracy effectively.