The Garden of Forking Prompts: How Users Explore Narrative Space in Story Generation

Large language models (LLMs) have changed the way people engage with stories. Users of modern chatbots iteratively edit their prompts to explore narrative possibilities, adjusting characters, redirecting plots, and swapping fictional universes to produce rich traces of creative preference at scale. Yet story generation evaluation benchmarks rely on static, one-shot prompts that cannot capture this exploratory behavior. In this work, we study how users revise consecutive story prompts in the wild. We extract paired prompts from real chatbot conversations and develop a framework of edit types through open coding of observed edit pairs, identifying the dimensions along which users navigate narrative space: four edit directions (adding, removing, changing, and extending content) crossed with thirteen targets (from plot and character description to genre and fandom). We then use this framework to build a prompt permutation pipeline that generates structured story variations grounded in real editing patterns rather than artificial or idealized prompt differences, and we analyze large sets of edit chains, identifying consistent shifts in lexical patterns and narrative shapes.