Deepfake detection aims to generalize across unseen deepfake image domains without requiring training samples from the target distribution. While vision-language models like CLIP exhibit strong generalization capabilities, their emphasis on semantic alignment hampers their ability to capture subtle low-level artifacts indicative of generation. In this work, we propose PromptFake, a simple yet effective method that enhances CLIP’s deepfake detection performance through two key innovations. First, we introduce learnable textual prompts that guide the model to discriminate between real and fake content across diverse generative models. Second, we propose a lightweight Feature Separation Module that operates on intermediate transformer layers to decouple generate-related cues from semantic information, improving feature sensitivity to deepfake artifacts. Extensive evaluations on a range of datasets—including those generated by GANs and diffusion models—indicate that PromptFake delivers encouraging performance in deepfake detection, with notable gains in both accuracy and mean average precision.