novel-editing-patterns

Diagnose the chapter against five payoff checks (money landing, shock reactions, face-slapping, income comparison, system mystery). Then inject low-disruption fixes like a system data comparison, an internal resource-counting monologue, or an external observer's viewpoint, limited to two injection points per chapter.

Use the dual-gate standard: 2200-2400 characters plus content density of at least one plot change per 500 characters. Never pad with scenery or repetition; instead add valid plot beats like conflict escalation, character decisions, or financial specifics, verified with review_scan.py.

After two consecutive Claude Code failures, the main agent takes over drafting directly. For over-compressed output below 2200 characters, apply targeted patches with valid plot content rather than re-dispatching full rewrites, following the word count oscillation stop-loss ladder.

When one error is found, search all chapters for the same keyword pattern and related entities before fixing. Verify geographic feasibility against each character's current position, and compute timeline windows explicitly from anchor dates before allowing result-type events.

Yes, it includes a dialect audit workflow: scan chapters for dialect characters, classify each line as keep-or-replace based on character identity, batch-patch with backups, then update the book configuration so future chapters follow the quantified dialect limits.