Cultural heritage sites are increasingly embracing playful approaches such as treasure hunt games to enhance interactive educational experiences, proving effective in engaging diverse audiences, encouraging exploration, and promoting knowledge retention. However, designing such experiences remains a complex task that requires careful integration of narrative, interactivity, and factual accuracy. In this article, we propose an approach that leverages large language models (LLMs) to automate the initial drafting of treasure hunt games for cultural heritage sites. Our method allows content curators to specify key parameters—such as the target artefacts to be included in the hunt, intended audience, and narrative styles—after which the system generates a structured sketch of the game. We compare two generation strategies: a basic sequential method and a revised approach that incorporates a pre-planning phase. Our evaluation assesses the resulting drafts in terms of their correctness, consistency, and stylistic coherence. Results suggest that pre-planning improves the quality of the generated content, producing generally more structured and contextually appropriate outputs. Moreover, we describe some of the remaining challenges, such as the need for interactive and validated co-design mechanisms or the introduction of factual accuracy guarantees in the adventures.