Predictive Maintenance (PdM) has become a cornerstone of Industry 4.0, aiming to reduce operational costs and downtime through accurate prognostics and optimized scheduling. While deep learning methods have advanced Remaining Useful Life (RUL) estimation, translating prognostic insights into actionable maintenance policies remains challenging due to reward mis-specification, safety-critical constraints, and domain-specific cost models. Reinforcement Learning (RL) provides a principled framework for sequential decision-making, but conventional designs are highly sensitive to reward formulation and often fail to generalize across diverse industrial contexts. In this study, a framework for Prompt-Driven Deep Reinforcement Learning in Modern Predictive Maintenance is introduced. Natural language prompts, interpreted through large language models (LLMs), are employed to dynamically reconfigure cost models, reward shaping, and action constraints during training. The framework integrates an LSTM-based RUL predictor with DQN/DRQN agents on the C-MAPSS benchmark. Experimental evaluation indicates that, although the prompt-driven agents do not yet achieve cost superiority over static baselines, they deliver improved stability in policy behavior and demonstrate adaptability through operator-guided prompt specifications (e.g., risk-averse or cost-saving modes). This work is positioned as a proof-of-concept and, to the best of current knowledge, represents the first systematic attempt to couple LLM-guided prompts with RL in PdM. The findings highlight feasibility, interpretability, and adaptability as key contributions, while cost optimization is identified as a critical direction for future research.