Most AI writing tools treat a novel like one long prompt — which is exactly why they unravel after twenty chapters. The hard part of serial fiction isn't generating prose; it's remembering that a character died in chapter 8 and shouldn't be ordering coffee in chapter 40. This plugin reframes the problem as state management rather than text generation, and that single shift is what lets it scale to hundreds of chapters.
What Sets It Apart
- Narrative facts are committed state, not memory. Each chapter writes character status, timeline, and world rules into a versioned
.story-system/store, so continuity becomes a data problem the system can check rather than something the model has to recall. - Writing moves through gates, not a single shot. Eight commands (init, plan, write, review, query, learn, dashboard, doctor) and four internal agents split planning, drafting, review, and fact extraction, with a review gate that blocks contradictions before a chapter is accepted.
- Retrieval is hybrid and local. BM25 keyword search plus semantic embeddings (ModelScope/Jina) run over a local SQLite index, so context recall doesn't depend on cramming the whole story into the prompt window.
Great Fit / Look Elsewhere
Great fit if you write genre serial fiction — it ships presets for 37 Chinese web-novel genres — and value long-range continuity over one-off polish; the versioned audit trail also helps when co-writing across many sessions. Look elsewhere if you want a one-shot prose generator, dislike command-driven workflows, or aren't already in the Claude Code ecosystem, since the whole system is built as a plugin and assumes you're comfortable running its commands and maintaining a local index.
