This paper addresses the optimal execution problem with limit orders, focusing on the challenge of setting limit prices in a hybrid action space that is both continuous (percentage changes) and discrete (tick sizes). The authors propose HALOP (Hybrid Action-space Limit Order Placement), a two-stage reinforcement learning method. In Stage 1, a discretized Gaussian policy selects a target percentage change in price, scoping a subset of discrete actions for generalization across stocks. In Stage 2, a Gaussian-and-Softmax policy chooses a specific tick-based action from the subset, enabling specialization for individual stock characteristics. The method uses PPO for policy optimization and a neural network with 1D-CNN and self-attention for state representation. Experiments on CSI 300 data show that HALOP significantly outperforms market orders, vanilla PPO with continuous or discrete actions, and previous learning-based methods like ODP. It achieves an average excess return of 4.41 bps over TWAP with high statistical significance and improved stability. A grouping study confirms that HALOP balances generalization and specialization, particularly benefiting low-priced stocks. The work provides a practical framework for applying RL to real-world trading where action spaces have both continuous and discrete aspects.