A deep-scan tool for identifying and replacing AI-generated prose patterns in fiction. Goes
far beyond general prose quality review by targeting the specific, cataloged fingerprints
that large language models leave in creative writing.

## Key Points

- User wants to "de-AI" their manuscript or make it sound more human
- User says "find AI tells", "detect AI writing", "scrub AI patterns", "make this sound human"
- User is preparing for publication and wants to eliminate robotic prose
- As a complement to the Novel Audit's Module 6, for deeper pattern analysis
1. "a symphony of [noun]"
2. "a tapestry of [noun]"
3. "a dance of [noun]"
4. "a mosaic of [noun]"
5. "a kaleidoscope of [noun]"
6. "a cascade of [noun]"
7. "the weight of [noun] settled over/on"
8. "the fabric of [noun]"

AI Tell Detector Skill

A deep-scan tool for identifying and replacing AI-generated prose patterns in fiction. Goes far beyond general prose quality review by targeting the specific, cataloged fingerprints that large language models leave in creative writing.

When to Use This Skill

• User wants to "de-AI" their manuscript or make it sound more human
• User says "find AI tells", "detect AI writing", "scrub AI patterns", "make this sound human"
• User is preparing for publication and wants to eliminate robotic prose
• As a complement to the Novel Audit's Module 6, for deeper pattern analysis

The AI Tell Taxonomy

Category 1 — Ornamental Metaphor Syndrome (OMS)

AI models default to flowery, abstract metaphors that sound literary but carry no specific meaning. These are the most recognizable AI fingerprints:

The Blacklist — Ornamental Phrases:

1. "a symphony of [noun]"
2. "a tapestry of [noun]"
3. "a dance of [noun]"
4. "a mosaic of [noun]"
5. "a kaleidoscope of [noun]"
6. "a cascade of [noun]"
7. "the weight of [noun] settled over/on"
8. "the fabric of [noun]"
9. "in the grand scheme of things"
10. "the ebb and flow of"
11. "a whirlwind of emotions"
12. "painted across her/his face"
13. "hung heavy in the air"
14. "cut through the silence"
15. "pierced the veil of"

Category 2 — Emotional Stage Directions (ESD)

AI tells the reader what to feel instead of creating the feeling through action and detail:

16. "couldn't help but [verb]"
17. "a smile that didn't quite reach [their] eyes"
18. "let out a breath [they] didn't know [they'd] been holding"
19. "something shifted in [their] eyes"
20. "a flicker of [emotion] crossed [their] face"
21. "[their] heart hammered/raced/pounded in [their] chest"
22. "a knot formed in [their] stomach"
23. "tears pricked at the corners of [their] eyes"
24. "a chill ran down [their] spine"
25. "[they] swallowed hard/thickly"

Category 3 — Filler Transitions and Padding (FTP)

AI uses these to bridge scenes when it doesn't know what happens next:

26. "as the days turned into weeks"
27. "little did [they] know"
28. "it was then that [they] realized"
29. "the silence stretched between them"
30. "time seemed to stand still"
31. "the world seemed to fall away"
32. "and just like that, everything changed"
33. "the question hung in the air"

Category 4 — Pseudo-Profound Closers (PPC)

AI loves ending paragraphs and chapters with lines that sound deep but say nothing:

34. "and perhaps, that was enough"
35. "some things were better left unsaid"
36. "but that was a story for another day"
37. "and in that moment, [they] understood"
38. "the journey was only beginning"
39. "nothing would ever be the same"
40. "[they] knew, deep down, that..."

Category 5 — Structural Tells

These are patterns in how AI organizes prose, not specific phrases:

• The triple beat: describing everything in groups of three adjectives or three actions
• Mirror paragraphs: opening and closing a scene with nearly identical imagery
• Epiphany dumps: characters suddenly understanding complex truths in a single moment
• Dialogue sandwich: action beat — dialogue — internal thought, repeated identically
• The enumeration impulse: listing items when narrative would be stronger
• Synonym cycling: using three different words for the same thing in consecutive sentences to appear varied ("the car / the vehicle / the sedan")

Scanning Process

Per-Chapter Analysis

For each chapter:

1. Phrase scan: Count occurrences of every blacklisted phrase and close variants.
2. Pattern scan: Identify structural tells (triple beats, mirror paragraphs, etc.).
3. Density calculation: (total AI tells found) / (total word count) * 1000 = AI Tell Density Score (ATDS) per thousand words.
4. Heat mapping: Mark the densest paragraphs for priority revision.

Scoring Interpretation

ATDS Range	Rating	Interpretation
0-2	Clean	Reads as human-written
2-5	Light	Occasional AI flavor; minor polish needed
5-10	Moderate	Noticeable AI patterns; systematic revision recommended
10-20	Heavy	Reads as AI-generated to attentive readers
20+	Saturated	Extensive rewriting required

Replacement Strategy

For every flagged instance, provide a context-appropriate replacement. Do not simply swap one cliche for another. The replacement must be specific to the scene, use concrete sensory detail, match the character's voice, and be shorter than the original when possible.

Examples: "a symphony of emotions played across her face" becomes "Her jaw tightened. She looked at the letter again, then folded it in half." / "he couldn't help but smile" becomes "He grinned before he'd even decided to." / "the silence stretched between them" becomes "Neither of them reached for the check."

Output Format

# AI Tell Detection Report
**Title**: [Novel title]
**Date**: [Today]
**Chapters scanned**: [N]
**Total AI tells found**: [N]
**Manuscript ATDS (overall)**: [score]

## Per-Chapter Density

| Chapter | Word Count | AI Tells | ATDS | Rating |
|---------|-----------|----------|------|--------|
| 1 | ... | ... | ... | ... |
| ... | ... | ... | ... | ... |

## Flagged Passages with Replacements

### Chapter [N]
**[location]**: "[flagged text]" — Tell: [category] — Replace: "[suggestion]"

## Summary Recommendations
[Overall assessment and revision strategy]

Anti-Patterns

Flagging intentional literary language. Some authors genuinely write in an ornate style. If the manuscript consistently uses elevated prose with specificity and purpose, that is style, not an AI tell. AI tells are generic and interchangeable — real literary prose is precise and earns its complexity.

Providing equally generic replacements. Replacing "a symphony of emotions" with "a storm of feelings" solves nothing. Every replacement must be grounded in the specific scene.

Treating the blacklist as exhaustive. New AI patterns emerge constantly. If you spot a recurring phrase that feels machine-generated but isn't on the list, flag it anyway and note it as an emerging pattern.

Ignoring context frequency. A single "heart pounded" in a 90,000-word novel is fine. The same phrase appearing twelve times is the problem. Always report frequency, not just presence.

Over-correcting into bland prose. The goal is not to eliminate all figurative language. The goal is to replace generic AI metaphors with specific, earned imagery. Flat, purely functional prose is not the target.