Complex task planning and execution in dynamic environments remain challenging due to sparse rewards, temporal dependencies, and complex preconditions. Traditional reinforcement learning methods struggle to converge efficiently under these constraints. To address these issues, we propose a Hierarchical Task Planning and Multi-Skill Scheduling (HTPMSS) framework powered by large language models (LLMs). Our method decomposes high-level tasks into logically structured sub-tasks through natural language-based semantic reasoning, forming executable hierarchical sequences. At the lower level, we dynamically reassess and reorder skills based on real-time feedback provided by the LLM, ensuring adaptability and efficiency during execution. Extensive experiments conducted in robotics manipulation tasks and simulated industrial workflows demonstrate that HTPMSS achieves superior performance compared to baseline methods, significantly improving success rates and convergence speeds. Ablation studies further validate the indispensable roles of hierarchical decomposition and dynamic scheduling, highlighting their complementary effects. Our work illustrates the effectiveness of combining LLM-driven semantic planning with adaptive multi-skill scheduling, offering a promising solution for scalable and flexible autonomous systems.