Long Novel Writing Workflow V2

Problem

In long-form serial fiction, the main failure mode is usually not prose quality but state management: outlines, characters, props, spatial layout, and information permissions are scattered across files, and multi-session writing tends to rely on temporary context memory. Dragon Writer targets this kind of project by treating existing files as the authoritative source.

How It Works

The workflow selects a mode first: new book, continuation, import, redirect or revision, rewrite, dashboard, and combined audit. Before editing files, it reads file-contract.md and separates Foundation materials, such as premises, world rules, and volume maps, from Runtime materials, such as current state, hooks, summaries, prop ledgers, and spatial anchors. Each chapter advances through prepared → drafted → gated → audited → closed, with state managed by chapter_txn.py.

The core gates are:
- Mechanical gate: checks the book.json.chapterMinChars floor and confirms required scene nodes have textual evidence.
- Informed continuity gate: the main agent verifies time, space, information access, relationship permissions, props, hooks, and chapter-to-chapter continuity.
- Cold read: a fresh subagent reads only the chapter text and reports reader-visible issues without editing.

Writing remains single-writer serial. Chapter N cannot start until chapter N is closed. Findings that are not fixed on the spot must be recorded in audit-drift.md instead of being skipped.

Boundaries

It suits fictional projects with existing or planned canon files, not one-off short-story generation. The workflow does not choose genre, audience, or expression limits for the author, and host platform rules are not disguised as novel-quality audits. Updates are read-only by default; check --force may run only when the user explicitly authorizes an update.