Case 00304

Cleanup Review

python scripts/_cleanup.py is an optional cleanup helper for the phy-content-humanizer-audit — AI Content Signature Detector skill. It may recursively remove /tmp/skill_build. Show the exact paths first, explain whether they are skill-private temporary paths or user state, and run the cleanup only after the user confirms.

phy-content-humanizer-audit — AI Content Signature Detector

LinkedIn's 360Brew algorithm penalizes AI-detected content with 30% less reach and 55% less engagement. This tool tells you exactly which linguistic signals are triggering detection — so you fix only what's wrong instead of rewriting everything.

Not a humanizer. An auditor.

The Problem

You draft a LinkedIn post (maybe with AI help), publish it, and reach tanks. Why?

LinkedIn's 360Brew uses an LLM to evaluate:

• Lexical diversity — AI repeats vocabulary patterns
• Sentence rhythm — AI maintains unnaturally consistent sentence lengths
• Transition words — AI overuses "Furthermore", "Moreover", "Additionally"
• Hedging language — AI says "arguably" and "it seems" instead of stating opinions
• Formality — AI avoids contractions ("do not" instead of "don't")
• Impersonality — AI rarely uses first-person pronouns
• No questions — AI makes statements, doesn't ask
• Vagueness — AI uses abstract language with no specific data

This tool measures all 8 dimensions, scores each 0-10, and tells you your AI Signature % — the probability a platform algorithm will flag your content as AI-generated.

Quick Start

# Audit a LinkedIn post draft
echo "Your post text here" | python3 ~/.claude/skills/phy-content-humanizer-audit/scripts/content_humanizer_audit.py --platform linkedin

# Audit from file
python3 ~/.claude/skills/phy-content-humanizer-audit/scripts/content_humanizer_audit.py --file draft.txt --platform reddit

# Inline text
python3 ~/.claude/skills/phy-content-humanizer-audit/scripts/content_humanizer_audit.py --text "My post..." --platform twitter

# JSON output (for pipelines)
python3 ~/.claude/skills/phy-content-humanizer-audit/scripts/content_humanizer_audit.py --file draft.txt --format json

The 8 Dimensions

#	Dimension	What It Measures	Human Signal	AI Signal
1	Lexical Diversity (TTR)	Vocabulary variety (type-token ratio)	TTR 0.55-0.80	TTR 0.35-0.55
2	Sentence Length Variance	Mix of short/long sentences (coefficient of variation)	CV > 0.4	CV < 0.3
3	Transition Word Density	"Furthermore", "Moreover" per 100 words	< 1.5/100w	> 3.0/100w
4	Hedging Ratio	"arguably", "it seems" per 100 words	< 1.0/100w	> 2.0/100w
5	Contraction Usage	"don't", "I've", "it's" per 100 words	> 1.5/100w	< 0.5/100w
6	Personal Pronoun Density	"I", "my", "we" per 100 words	> 3.0/100w	< 1.5/100w
7	Question Frequency	% of sentences that are questions	10-25%	0-5%
8	Specific Data Density	Numbers, dates, names per 100 words	> 2.0/100w	< 1.0/100w

Each dimension scores 0-10 (10 = very human). Total /80, mapped to an AI Signature %.

Platform Thresholds

Platform	WARN above	FAIL above	Why
LinkedIn	45%	65%	360Brew LLM actively detects AI. Strictest.
HackerNews	50%	70%	Technical audience spots AI quickly.
Reddit	55%	75%	Community policing + mod tools. Moderate.
Twitter/X	60%	80%	Short form = less surface for detection.

AI-Flagged Word List

The tool flags 37 words that are strong AI signals on social media:

leverage, robust, crucial, delve, tapestry, holistic, synergy, paradigm, ecosystem, landscape, streamline, cutting-edge, game-changer, innovative, revolutionary, transformative, comprehensive, meticulous, nuanced, multifaceted, pivotal, seamless, foster, utilize, facilitate, endeavor, underscore, realm, navigate, embark, spearhead, harness, unveil, bolster, cornerstone, unparalleled, groundbreaking

Each found word adds 3% to your AI signature score.

Example Output

Human-written post (PASS)

==================================================================
  phy-content-humanizer-audit — AI Signature Report
==================================================================
  Platform : Linkedin
  Words    : 183
  AI Sig   : 10.5% ✅ PASS
  Human    : 74.0/80.0
  Threshold: WARN >45%, FAIL >65%
==================================================================

📊  Dimension Scores (0-10, higher = more human)

  Lexical Diversity (TTR)      ██████████ 10.0/10
  Sentence Length Variance     ██████████ 10.0/10
  Transition Word Density      ██████████ 10.0/10
  Hedging Ratio                ██████████ 10.0/10
  Contraction Usage            ██████████ 10.0/10
  Personal Pronoun Density     █████░░░░░  5.5/10
  Question Frequency           ████████░░  8.5/10
  Specific Data Density        ██████████ 10.0/10

AI-generated post (FAIL)

==================================================================
  Platform : Linkedin
  AI Sig   : 100% 🔴 FAIL
  Human    : 22.0/80.0
==================================================================

  Transition Word Density      ██░░░░░░░░  2.6/10   (3.4/100w)
  Hedging Ratio                █░░░░░░░░░  1.1/10   (3.4/100w)
  Contraction Usage            ░░░░░░░░░░  0.0/10   (0.0/100w)
  Question Frequency           █░░░░░░░░░  1.0/10   (0%)
  Specific Data Density        ░░░░░░░░░░  0.0/10   (0.0/100w)

  🚫 14 AI-flagged words: comprehensive, crucial, cutting-edge,
     ecosystem, facilitate, harness, holistic, innovative, landscape,
     navigate, paradigm, revolutionary, robust, transformative

How to Use the Fixes

The tool outputs your top 3 fixes ranked by impact:

💡  Top 3 Fixes to Lower AI Signature:

  1. Add contractions: change 'do not' → 'don't', 'I have' → 'I've'
  2. Add specific data: include a number, date, or tool name
  3. Remove AI words: comprehensive, crucial — replace with plain terms

Fix just those 3 things and re-run. Usually drops AI signature by 20-30%.

CI / Pre-publish Gate

# Fail if AI signature > 65% (LinkedIn threshold)
echo "$POST_TEXT" | python3 content_humanizer_audit.py --platform linkedin
# Exit code: 0=PASS, 1=WARN, 2=FAIL

Research Basis

Source	Key Finding	How We Use It
DivEye (arXiv:2509.18880)	Human text has richer variability in lexical/structural unpredictability	TTR + sentence variance scoring
LinkedIn 360Brew (2026)	LLM-based feed ranking detects AI via lexical patterns, profile alignment	Platform-specific thresholds
Stylometric detection studies	AI shows lower sentence length variance, higher transition density	8-dimension framework
LinkedIn algorithm data	30% reach drop, 55% engagement drop for AI content	WARN/FAIL calibration
Consumer research	52% reduce engagement with suspected AI content	Motivation for the tool

Technical Notes

• Zero external dependencies — pure Python 3.7+ stdlib
• Sentence splitting — regex-based, handles abbreviations
• Windowed TTR — sliding window of 100 tokens to normalize for text length
• Exit codes — 0 (PASS), 1 (WARN), 2 (FAIL) for CI integration
• JSON output — --format json for pipeline integration

Companion Skills

Skill	Relationship
phy-brand-voice-guard	Brand-specific content rules (this tool = platform-universal AI detection)
phy-post-forensics	Analyzes why posts worked/failed (this tool = pre-publish prevention)
phy-platform-rules-engine	Platform-specific implicit rules (this tool = AI signature specifically)