Given the vast amount of information available today, the spread of fake news significantly undermines societal trust and decision-making processes. Moreover, this news appears across multiple domains, such as health, politics, or social media, with different intrinsic characteristics, making fake news detection (FND) an even harder problem. This paper investigates the efficiency of transfer learning (TL) and prompt tuning in improving FND across multiple domains. We utilized and fine-tuned the Llama 2 7B model, incorporating prompt tuning to enhance detection accuracy. First, we classify the domain of news articles using a DistillBERT model, which has been demonstrated to be a highly accurate domain classifier. Then, we employ domain-specific prompts tailored to guide the fine-tuned Llama 2 7B model in assessing news credibility. The second layer of our approach involves fine-tuning Llama 2 for FND and subsequently applying prompt tuning. This combination allows us to compare three scenarios: blind testing with the original Llama 2, the performance of the fine-tuned Llama 2 model, and the enhanced accuracy achieved through prompt tuning in the second stage. Our results demonstrate substantial improvements in precision, recall, accuracy, and F1 scores across all tested domains. This study highlights the potential of integrating domain-specific prompt tuning with fine-tuning-based TL to advance the accuracy and reliability of automated FND systems.