detecting-ai-slop

Detecting AI Slop

Guidelines: See CLAUDE.md for writing standards. Last Updated: 2026-06-16 See also: reference.md (pattern dictionary), examples.md (usage examples)

Wrong skill? Rewriting to remove slop → eliminating-ai-slop. Profanity/inappropriate language → professional-language-audit.

Detection Approach

This skill analyzes text and produces a slop score (0-100) with detailed breakdown by detection dimension. Use it to quantify AI slop before deciding whether to rewrite.

Core principle: Detection is read-only. This skill flags patterns but does not rewrite. Use eliminating-ai-slop for active rewriting.

When to Use

• Score a CV or resume for AI-generated content
• Analyze cover letters for generic patterns
• Audit marketing copy for slop density
• Review your own AI-assisted drafts before editing
• Compare before/after versions of edited text
• Triage documents: which need the most cleanup?

Content Type Detection

The skill auto-detects content type from context:

Content Type	Detection Signals
Document	Default fallback
Email	"email", "to:", "subject:"
LinkedIn	"linkedin", "post", "connections"
CV/Resume	"resume", "cv", "experience"
Cover Letter	"cover letter", "dear hiring"
README	Filename is "README"
PRD	"requirements", "PRD", "product"

Override: "Analyze this as a [type]: [text]"

Output Format

Slop Score: 73/100

Breakdown:
├── Lexical:      28/40  (14 patterns in 500 words)
├── Structural:   18/25  (formulaic intro, template sections)
├── Semantic:     12/20  (3 hollow examples, 1 absolute claim)
└── Stylometric:  15/15  (low sentence variance, flat TTR)

Top Offenders (showing 10 of 23):
 1. Line 12: "incredibly powerful" [Generic booster]
 2. Line 34: "leverage synergies" [Buzzword cluster]
 3. Line 56: "it's important to note" [Filler phrase]
 ...

Stylometric Measurements:
├── Sentence length σ: 7.3 words (target: >15.0) ⚠️
├── Type-token ratio: 0.48 (target: 0.50-0.70) ⚠️
└── Hapax rate: 31% (target: >40%) ⚠️

Verdict: Heavy slop. Substantial rewrite needed.

Scoring Algorithm

Dimension	Max Points	Calculation
Lexical	40	min(40, pattern_count * 2)
Structural	25	5 * content_structure_patterns + sum(style_tell_weights) — content patterns (Cat. 9, formulaic intro, etc.) score 5 pts each; style-level tells (random bolding, one-sentence paragraphs, etc.) use variable weights from reference.md
Semantic	20	5 * semantic_patterns_found
Stylometric	15	5 * stylometric_flags

Total: Sum of dimensions, capped at 100.

Score Interpretation

Score	Interpretation
0-20	Clean: minimal AI patterns detected
21-40	Light: some patterns, minor editing needed
41-60	Moderate: noticeable AI fingerprint, edit recommended
61-80	Heavy: significant slop, substantial rewrite needed
81-100	Severe: text reads as unedited AI output

Stylometric Thresholds

Based on StyloAI (Opara, 2024) and Desaire et al. (2023) research.

Metric	Flag If	Target
Sentence length σ	σ < 15.0	σ > 15.0
Paragraph length SD	SD < 25	SD > 25
Type-Token Ratio	TTR < 0.50 or TTR > 0.70	0.50 ≤ TTR ≤ 0.70
Hapax legomena rate	Below user baseline	At or above baseline

Structural & Semantic Patterns (45 points max)

Pattern	Description	Points
Formulaic Introduction	Rephrasing topic → importance → overview	+5
Template Sections	Overview → Key Points → Best Practices → Conclusion	+5
Over-Signposting	"In this section...", "As mentioned earlier..."	+5
Staccato Paragraphs	>50% are 1-2 sentences	+5
Symmetric Coverage	Equal weight to all options without prioritization	+5
Hollow Specificity	"Many companies have seen improvements" (which?)	+5
Absent Constraints	Absolute claims without limitations	+5
Balanced to a Fault	Every pro has matching con of equal weight	+5
Circular Reasoning	Rephrases thesis without new evidence	+5
Structural Contrast	"It's not about X. It's about Y." slogan forms (see Cat. 9 in reference.md)	+5

Lexical Pattern Categories

For the complete pattern dictionary, see reference.md.

Category	Examples	Action
Generic Boosters	incredibly, extremely, very	Delete or replace with metrics
Buzzwords	robust, seamless, leverage, elevate, harness, pivotal, impactful	Replace with plain language
Filler Phrases	"It's important to note that", "In today's ever-evolving world"	Delete entirely
Hedge Patterns	of course, arguably, seems to	Commit or remove
Sycophancy	"Great question!", "Happy to help!"	Delete
Transitional Filler	Furthermore, Moreover, Additionally, However, Indeed	Use sparingly or cut
Vague Abstraction	the frame, the lens, the narrative, the space	Replace with the specific noun
Structural Contrasts	"It's not about X. It's about Y."	Rewrite as direct claim
Style Tells	one-sentence paragraphs, random bolding, abstract noun stacking	Restructure
Typographic Tells	em-dash (—), smart quotes	Replace with standard punctuation

Dictionary Integration

This skill reads from .slop-dictionary.json if present in workspace root.

• Custom patterns are included in detection
• Exceptions are skipped during detection
• Weight affects scoring: score = base_score * weight

Note: This skill reads from the dictionary but does not write. Use eliminating-ai-slop to add patterns or exceptions.

Semantic Quick Tests

Use these when reviewing AI text qualitatively (merged from reviewing-ai-text):

Test	Slop Signal	Real Signal
Specificity	"Use appropriate caching strategies"	"Use Redis with 5-minute TTL for session data"
Asymmetry	"Both options have merits"	"Use Postgres unless >10M writes/day"
Constraint	"Implement microservices for scalability"	"Microservices add 3x ops overhead. Stay monolith unless dedicated platform team."
First-Person	Generic enough to apply anywhere	Grounded in specific context

Companion Skills

• eliminating-ai-slop: Active rewriting to remove detected patterns
• professional-language-audit: Profanity and inappropriate language detection
• readme-authoring: README generation
• incorporating-research: Score research quality before incorporating

Example

# Score text for AI patterns (read-only analysis)
# GVR = Generality + Verbosity + Repetition
echo "Check for: hedging ('It is worth noting'), filler ('In order to'),
  superlatives ('incredibly powerful'), and vague claims ('comprehensive')"

Failure Modes

• False positives on domain jargon: Flagging legitimate technical terms (e.g., "robust" in a load-testing context) as slop
• Score inflation: Giving a passing score to text with subtle but pervasive AI patterns
• Detection without action: Scoring text as sloppy but not invoking eliminating-ai-slop to fix it