Task-oriented chatbots are increasingly used to access all sorts of services – like booking a flight, or setting a medical appointment – through natural language conversation. There are many technologies for implementing task-oriented chatbots, including Dialogflow, Watson, and Rasa. They rely on an explicit definition of the user intents, conversation flows, and chatbot outputs, which is costly to specify, and sometimes results in suboptimal user experiences and artificial conversations with limited diversity of chatbot responses. Recently, the advances in generative artificial intelligence fostered by Large Language Models (LLMs) have enabled a new range of open-domain chatbots, like ChatGPT, able to converse fluently on any topic. However, they are general-purpose, and therefore not directly usable to solve specialised tasks reliably. In this paper, we study the power of LLMs to build task-oriented chatbots, resulting in lighter specifications – no intent definition required – and more natural conversations than in intent-based approaches. To this end, we propose a lightweight domain-specific language based on YAML to specify chatbots using modules of different types (e.g., menus, question-answering, data gathering). These specifications are compiled into structured LLM prompts that use the ReAct framework to inform our runtime how to interpret the user input and coordinate the tasks that the chatbot must perform. The paper presents the design and realisation of our framework, and an assessment that encodes a set of existing intent-based chatbots using our approach, showing its benefits in terms of specification size, conversation flexibility and output diversity.