Recent studies have shown that large language models perform excellently on downstream tasks. However, applying large models to named entity recognition (NER) through fine-tuning faces significant cost barriers. Therefore, we introduced a Three-Stage Prompt-based Low-Resource Named Entity Recognition (TPL-NER) model, aimed at improving the performance of zero-shot and few-shot NER tasks through contextual learning. TPL-NER addresses zero-shot and few-shot NER problems through a three-tiered step-by-step reasoning strategy. First, it identifies the possible entity types in a sentence, then recognizes which entities belong to each category within the sentence, and finally confirms the entity type for predicted entities that are easily confused across multiple categories. Experimental results on datasets from multiple domains and different languages show that TPL-NER’s superior performance in zero-shot and few-shot NER tasks.