Professional Long-Form Novel Narrative and Outline Generation

The Problem It Solves

Long-form fiction often scatters context across chat logs, chapter drafts, character notes, and worldbuilding files. Models may treat final-v2-new-really-final.md as canon, confuse temporary character memory with objective history, or load an entire manuscript and dilute the current task. This skill separates system specs from novel projects, forcing the agent to read only the minimum files routed by the project entry and role system.

How It Works

• Project isolation: each novel lives in projects/<novel>/, with no shared canon, character memory, drafts, or review results; start from projects/_template/ by copying and renaming.
• Entry-driven loading: the AI reads SKILL.md, projects/<novel>/00-project/AI-ENTRY.md, the current task context, and the primary role routed by system/README.md, adding at most two directly dependent roles when necessary.
• Artifact layering: world settings, characters, canon, scene plans, drafts, revisions, and published text stay separated; candidate settings must not enter 03-canon/ directly, and prose is not the canon submission path.
• Outline and narrative flow: project.yaml, AI-ENTRY.md, and scene templates structure current goals, role responsibilities, scene blueprints, and writing tasks, useful for outlines, chapter breakdowns, consistency checks, and review drafts.

Boundaries and Caveats

It is better treated as a long-form narrative workspace spec than an automatic whole-book generator. Mixing novel content into system/, creating directly in _template, or stuffing world rules, character memory, and prose into one master document breaks the isolation rules. The skill material lists DeepSeek, ChatGPT, Claude, Gemini, and huyuan as candidate models, but results still depend on how cleanly the entry files, role specs, and canon sources are maintained.