What problem does it solve? LLMs write good prose but hallucinate structure: invented place names, inconsistent numbers, and malformed system panels. This Skill separates structure from prose by having the orchestrator write a pre-validated chapter skeleton (scenes, numbers, panels, hooks) and letting Claude Code only fill in the prose, reducing per-chapter hallucinations from 3-5 to 0-1. ## Core Features & Use Cases - Skeleton Generation: Writes chapter skeletons with scene breakdowns, whitelisted fictional place names, verified numbers, and fixed-format system panels read from the book's configuration file. - Capacity Pre-check: Enforces minimums before drafting—at least 4 effective plot turns, 3 shifts in stakes, and 1 proactive protagonist decision—blocking under-capacity skeletons from entering the DRAFT stage. - Pre-review & Dispatch: Runs review_scan.py on the skeleton before dispatching it to Claude Code CLI with three short ironclad rules (whitelist place names, untouched panels, 2000-2800 word target). - Use Case: Writing chapter 42 of a system-flow web novel: generate the skeleton with the exact settlement panel values, pre-scan it, then dispatch to Claude Code which outputs prose that cannot alter numbers or invent real Chinese place names. ## Quick Start Generate the skeleton for chapter N of my current book, pre-validate it with the review scan, then dispatch it to Claude Code to write the prose draft.
