TombWriter reinvents AI co-writing with beat-level story archeology

TombWriter tackles a core problem in AI co-writing: disposable prompts that vanish after use, leading to imprecise results and diminished author agency. The paper proposes LLM-based story archeology, drawing on fossil theory where stories exist as latent structures writers excavate through craft. Instead of prompting for full prose, writers define and refine hierarchical prompts at the beat level—character actions in scenes—letting the LLM simulate emergent possibilities or allowing direct nudges. This separates structure from style, with prose generated later based on genre and style settings.

The result is TombWriter, a web-based tool with a five-stage narrative pipeline (inciting incident, rising action, etc.) that visualizes stories as navigable cards for characters, scenes, and beats. In a qualitative study with five experienced writers over three days, researchers found that writers framed AI as a generation engine rather than a collaborator, claimed ownership over the story but reported some voice loss, and valued the system primarily for structural discovery rather than prose production. The findings highlight a tension between control and creative voice that future co-writing tools will need to address.