Novel Weaver

The Problem

Long-form LLM writing often fails not because a single paragraph is weak, but because the finished story drifts from the outline, character timelines break, sub-scene causality is missing, and the ending does not close. novel-weaver targets multi-chapter novels and structured stories by turning writing into an auditable state-machine workflow.

How It Works

The skill centers on novel_state.json, phase gates, and serial context:
- Confirm before writing: scene setup, first-level outline, and chapter substructures require user confirmation and plan-chapter checks; missing writing_prompt or unregistered characters trigger HOOK-BLOCK.
- Pipeline gates: entering writing requires outline_causality and sub_causality; full completion requires fidelity and ending_verify.
- Serial writing: context_loader loads the next substructure only after the previous one is completed, preventing parallel drift.
- Chapter checks: finalize-chapter automatically runs continuity, cross-chapter promise, style, logic, semantic, and reasoning reviews; HARD issues block and write to _fixes.json.
- Optional models: semantic and reasoning steps use locally cached models only, skip when unavailable, and run on CPU.

Boundaries

It fits long stories with multiple chapters, characters, causal chains, and ending validation. It is not intended for simple rewriting, polishing, translation, or short-answer tasks. The workflow is constraint-heavy and requires user confirmation, serial progression, and hard blocking on failures.