The challenge of sparsity in user-item interactions has consistently posed an obstacle for recommendation systems, known as the cold-start problem. Existing methods address this by introducing auxiliary data, such as heterogeneous information networks (HIN), or employing rapidly adaptable models to novel tasks, such as meta-learning. Recently, the emerging prompt-tuning paradigm sheds light on efficiently guiding tasks with few-shot data. With just a few labeled examples, prompt-tuning learns soft prompts for different tasks and performs task inference under the task-specific soft prompt while freezing the pre-trained model. Thus, we are motivated to design a HIN-based prompt-tuning framework, called HINPrompt, to address the cold-start recommendation. In the training stage, we distill complex high-order semantic relations into semantic tokens, serving as the fundamental units of prompts for the different user-based recommendation tasks. Subsequently, a mixture layer is employed to compose different semantic tokens into user-specific prompts, guiding the cold users to find the most relevant prior from the pre-trained recommendation model. We conduct comprehensive experiments across various cold-start scenarios, demonstrating the effectiveness of HINPrompt for the cold-start problem.