A comprehensive quality-assurance workflow for AI-generated novels. Covers narrative consistency,
character integrity, factual plausibility, prose quality, and structural compliance.

## Key Points

- Full manuscript audits before publication
- Mid-draft consistency checks
- Outline compliance verification
- Character/name consistency sweeps
- Duplicate content detection
- Any single-concern check (always run everything unless told not to)
- **The manuscript** — full text, or chapter files (.txt, .md, .docx, or plain paste)
- **The outline** (optional but strongly recommended) — chapter-by-chapter or scene-by-scene plan
- **Character bible / cast list** (optional) — names, descriptions, relationships
- **World-building notes** (optional) — settings, rules, timeline, lore
1. Extract every proper noun that appears to be a character name (use capitalization + context clues).
2. Build a name registry: canonical_name → [all variants found]

Novel Audit Skill

A comprehensive quality-assurance workflow for AI-generated novels. Covers narrative consistency, character integrity, factual plausibility, prose quality, and structural compliance.

When to Use This Skill

Use whenever the user wants to catch AI-generation errors in long-form fiction, including:

• Full manuscript audits before publication
• Mid-draft consistency checks
• Outline compliance verification
• Character/name consistency sweeps
• Duplicate content detection
• Any single-concern check (always run everything unless told not to)

Input Requirements

Before starting, collect:

• The manuscript — full text, or chapter files (.txt, .md, .docx, or plain paste)
• The outline (optional but strongly recommended) — chapter-by-chapter or scene-by-scene plan
• Character bible / cast list (optional) — names, descriptions, relationships
• World-building notes (optional) — settings, rules, timeline, lore

If the user has none of these supporting documents, proceed with manuscript-only auditing and note which checks were skipped or degraded.

Audit Modules

Run all modules unless the user says otherwise. Each module produces a findings section in the final report.

Module 1 — Character & Name Consistency

Goal: Ensure every character is named consistently throughout the entire manuscript.

Steps:

1. Extract every proper noun that appears to be a character name (use capitalization + context clues).
2. Build a name registry: canonical_name → [all variants found]
3. Flag:
   • Spelling drift (e.g., Kaelin vs Kaelyn vs Kaylin)
   • Mid-story name changes with no in-text explanation
   • Characters referred to by title in some scenes and name in others, inconsistently
   • Minor characters whose names change silently (easy AI mistake)
   • Pronouns that shift for a character (he/she/they) without narrative reason
4. If a character bible was provided, cross-reference every character against it.

Output: table of Character | Canonical Name | Variants Found | Chapters Affected | Severity

Module 2 — Duplicate & Repeated Content Detection

Goal: Find passages that are copy-pasted, near-duplicated, or thematically rehashed.

Steps:

1. Exact duplicates: Search for any paragraph or sentence that appears verbatim more than once.
2. Near-duplicates: Flag passages that are >70% similar in wording (same scene described twice with minor word swaps — a classic AI failure mode).
3. Idea recycling: Flag scenes where the same emotional beat, revelation, or plot event is effectively repeated (e.g., the hero "decides to fight back" in chapter 3 and again in chapter 7 with no meaningful distinction).
4. Repeated phrases / verbal tics: Identify stock phrases the AI overuses (e.g., "a smile that didn't reach her eyes", "his heart hammered", "she let out a breath she didn't know she was holding"). Flag if any phrase appears more than 3x in the manuscript.

Output: list of duplicate/repeated items with chapter locations and exact or paraphrased excerpts.

Module 3 — Narrative & Outline Compliance

Goal: Verify the story actually follows the provided outline (if any) and is internally coherent.

Steps:

1. Outline mapping (if outline provided): For each outline beat, confirm it appears in the manuscript. Flag missing beats, out-of-order beats, and beats that appear but are underdeveloped (< 1 paragraph where a full scene was planned).
2. Plot hole detection: Look for events that rely on information the characters couldn't have, actions with no established motivation, or consequences that contradict earlier events.
3. Cause-and-effect chains: Verify that major decisions/actions have visible consequences in later chapters.
4. Subplot tracking: List all subplots introduced; flag any that are opened but never resolved ("Chekhov's gun" failures).
5. Timeline coherence: Flag temporal contradictions (character is in two places at once, events happening in impossible order, stated durations that don't add up).

Output: outline compliance table + list of plot holes/inconsistencies with chapter citations.

Module 4 — World-Building & Factual Plausibility

Goal: Catch internal contradictions in the fictional world and AI hallucinations about the real world.

Steps:

1. World-rule violations: If world-building notes exist, flag any scene where the rules are broken (magic system, technology limits, geography, etc.).
2. Setting consistency: Does the physical environment stay consistent? (Room layout, city geography, travel times, weather.)
3. Real-world facts (if novel is set in the real world or historical period): Flag any claim that sounds like an AI hallucination — wrong dates, invented places, incorrect historical events, implausible science. Note: you cannot verify everything; flag suspicious claims for human review.
4. Anachronisms: Technology, language, or cultural references that don't fit the setting's era.
5. Proper nouns for real places/organizations: Flag if the AI invented a real-sounding institution that doesn't exist, in a context where it might mislead readers.

Output: list of world-rule violations and suspected hallucinations, with chapter + quote.

Module 5 — Character Arc & Behavior Consistency

Goal: Ensure characters act like themselves throughout the book.

Steps:

1. Personality drift: Does a character's voice, attitude, or decision-making style shift without justification? (AI often forgets character traits between chapters.)
2. Knowledge consistency: Does a character "know" something in chapter 8 that they had no way of learning before chapter 10? Conversely, do they forget something they definitively learned?
3. Relationship continuity: Do character relationships evolve logically, or do they jump inexplicably?
4. Character voice: Does dialogue remain distinct per character, or do all characters start to sound like the same narrator?
5. Physical descriptions: Eye color, hair, scars, and other physical details — flag contradictions.

Output: per-character consistency summary + list of behavior/voice breaks.

Module 6 — Prose Quality & AI Tell Detection

Goal: Surface passages that read as low-quality AI output.

Common AI tells to flag:

• Purple prose avalanche: Dense, overwrought description with no narrative function
• Filler transitions: "Meanwhile...", "As the days passed..." used repeatedly to paper over plot gaps
• Emotion-telling over showing: "She felt sad." with no physical/behavioral grounding
• Sudden POV shifts mid-scene (head-hopping) not consistent with the book's style
• Unearned epiphanies: A character suddenly "realizes" something with no build-up
• Passive voice clusters: More than 3 passive constructions in a paragraph
• Chapter endings that summarize instead of hook: AI often ends chapters with a paragraph summarizing what just happened rather than creating tension

Output: list of flagged passages with module label + chapter location.

Module 7 — Structural & Formatting Issues

Goal: Catch mechanical/structural problems.

Steps:

1. Chapter length variance: Flag chapters significantly shorter or longer than the median.
2. Scene breaks: Check that scene-break markers are used consistently.
3. POV labeling (if multi-POV): Confirm each chapter is clearly attributed to a POV character.
4. Dialogue formatting: Consistent use of quotation marks, em dashes, and dialogue tags.
5. Chapter/section numbering: No skipped or duplicate numbers.

Output: structural issues list with locations.

Report Format

# Novel Audit Report
**Title**: [Novel title if known]
**Date**: [Today]
**Chapters audited**: [N]
**Word count (approx.)**: [N]
**Supporting documents provided**: [Outline: yes/no | Character bible: yes/no | World notes: yes/no]

---

## Executive Summary
[3-5 sentence overview of overall quality, biggest risks, and recommended priority fixes.]

## Severity Legend
🔴 CRITICAL — Breaks the story; must fix before publication
🟠 MAJOR — Noticeable error that will bother most readers
🟡 MINOR — Small inconsistency; fix if time allows
🔵 STYLE — Prose/craft suggestion; optional

---

## Module 1: Character & Name Consistency
[findings]

## Module 2: Duplicate & Repeated Content
[findings]

## Module 3: Narrative & Outline Compliance
[findings]

## Module 4: World-Building & Factual Plausibility
[findings]

## Module 5: Character Arc & Behavior Consistency
[findings]

## Module 6: Prose Quality & AI Tell Detection
[findings]

## Module 7: Structural & Formatting Issues
[findings]

---

## Master Issues List
[All findings consolidated, sorted by severity, with chapter citations]

## Recommended Fix Order
[Prioritized action list: fix X before Y because...]

Workflow for Large Novels

If the manuscript is > 50,000 words or > 20 chapters, process in passes:

Pass 1 — Structural read (fast): Skim chapter-by-chapter. Build character name registry, chapter summaries (1-2 sentences each), timeline of major events, list of subplots.

Pass 2 — Deep modules: Run Modules 1-7 using the registry and summaries from Pass 1 as anchors.

Pass 3 — Cross-check: Verify findings from Pass 2 against actual manuscript text before reporting.

Handling Missing Inputs

Missing input	Degraded behavior
No outline	Skip outline compliance; note in report
No character bible	Audit names from manuscript only; flag uncertain cases
No world-building notes	Flag only obvious internal contradictions
No manuscript (only chapters)	Process chapter by chapter; produce per-chapter + aggregate report

Anti-Patterns

Over-flagging style choices as errors. Module 6 is advisory. The author's intentional stylistic choices should be distinguished from AI slop. When uncertain, flag as a question for the author rather than a definitive error.

Skipping modules for a "quick check." Always run all modules unless explicitly told otherwise. A name-only check often misses the real problems (plot holes, duplicates).

Reporting without citations. Every finding must include chapter number and a brief quote or paraphrase so the author can locate it instantly. Vague findings are useless.

Losing context in long manuscripts. Use the three-pass workflow for books over 50K words. Maintain a running state document between sessions if processing chapter by chapter.

Flagging every passive voice instance. Some passive voice is fine. Only flag clusters (3+ in a paragraph) that indicate AI padding, not individual occurrences that serve the narrative.