<p>Zero-shot anomaly detection task is to identify anomaly regions without using target domain data. The detection methods based on large vision language model typically rely on text encoder to extract text features, which significantly increases the computational complexity of the model. This paper proposes a zero-shot anomaly detection method based solely on learnable text query. Text features are directly represented by designed learnable text query rather than encoded by the text encoder. Partial tokens from the input space of the image encoder are replaced by designed learnable vision prompt. Image features are extracted by the improved image encoder. Anomaly regions are calculated based on cosine similarity between text query and multiple adjusted image features. Four publicly available datasets are used to evaluate the zero-shot anomaly detection performance. Experimental results show that the proposed method achieves better results than the methods using the text encoder, and it has strong zero-shot generalization ability in different detection scenarios.</p>