<p>Object detection for autonomous driving continues to encounter substantial difficulties when operating in complex degradation conditions, including rain, fog, snow and low light. Existing detectors rely on the integrity of input images and tend to suffer performance collapse under unknown degradations, while traditional image enhancement methods depend on visual metrics and incur high computational cost, making them unsuitable for real-time applications. To tackle these challenges, we introduce PDA-Net (Prompt-guided Degradation-Aware Network), which performs dynamically adaptive enhancement via explicit degradation modeling and conditionally generated prompts. We design a Degradation-Aware Dynamic Recalibration (DADR) module, which employs a lightweight degradation encoder to explicitly extract degradation representations, while weak supervision ensures that the extracted features accurately capture degradation characteristics. The DADR output simultaneously enables FiLM-based feature modulation and multi-step prompt generation, ensuring that the prompts remain strongly aligned with both the type and intensity of degradation. By integrating a dual-path dynamic modulation strategy, PDA-Net enables comprehensive adaptation over multi-scale features, spanning from low-level details to high-level semantic representations. Experiments conducted on real-world harsh-weather datasets show that PDA-Net outperforms existing state-of-the-art approaches in terms of mAP, generalization, and inference efficiency, providing a novel pathway toward safer autonomous driving deployment in challenging conditions.</p>