Paper #29: Bonsai: Cultivating Author Intent in LLM-Based Interactive Digital Narratives

Authors of LLM-based interactive digital narratives (IDNs) struggle to preserve creative intent as player choices and real-time generation pull storylines in unpredictable directions. We propose cultivation as a design metaphor for LLM-based IDN authoring. Rather than specifying intent upfront, authors edit AI-improvised content, and inline edits can yield natural-language preference rules that steer future generations. We instantiate cultivation in Bonsai, an IDN authoring environment for studying preference learning from creative edit traces. Across three projects with simulated author preferences, intent rules parsed from those preferences produce a project-specific signal and transfer to small held-out test scenes. We also find that priming extraction on IDN authoring categories raises world-constraint recovery from 8\% to 83\% over naive extraction, though chance-corrected analyses qualify the size of this gain. By treating AI-generated content as persistent, editable material, cultivation reimagines the static IDN as a living artifact where alignment becomes a byproduct of use.