In this article, we discuss how AI21 Labs' Jurassic-1 language model can be customized to improve its performance on specific tasks such as news article topic classification. We demonstrate that by carefully crafting a prompt and providing the model with examples of correctly labeled articles, we can achieve significantly better accuracy than using the model in a zero-shot setting. Additionally, we show how using a custom model offers safety advantages over general-purpose models when it comes to mitigating prompt injection attacks. Finally, we discuss how even without any labeled data, one can still train a custom model by automatically labeling unlabeled examples with high confidence and using them for training.