Humanizer

Humanizer: AI Pattern Detection & Voice Injection

Transform AI-generated text into human writing by detecting patterns and injecting authentic voice.

NEVER

• Never remove AI phrases without also injecting personality — sterile text is still obviously AI.
• Never over-correct (random punctuation chaos !== human voice).
• Never remove ALL structure from formal or academic contexts — know what the audience expects.
• Never batch-replace AI words without checking domain context ("landscape" in CV papers may be correct terminology).
• Never fix every pattern at once — identify the 3-5 dominant patterns and address those.

Before You Edit: Diagnostic Framework

Answer these before applying any changes:

Voice assessment:

• What personality should come through? (witty, skeptical, conversational, authoritative)
• Does the text have opinions, or just facts? Human writing has stakes.

Pattern prioritization:

• Which 3-5 patterns dominate? (Don't fix everything at once)
• Should some AI-isms stay? (Formal technical docs may keep certain structures)

Rewrite philosophy:

• Am I removing patterns OR injecting personality? (Must do both)
• Does my rewrite sound like a specific human wrote it? (Not just "less AI")

Statistical density check (run before editing):

• AI vocabulary words per 100 words: >3 = heavy AI signature
• Em dashes per paragraph: >2 = structural tell
• "However" paragraph starts: >20% = AI transition overuse
• All paragraphs same length? = AI rhythm uniformity
• Every list has exactly 3 items? = rule of three addiction

Humanization Level by Context

Most Common AI Patterns

Content-Level

Undue emphasis on significance:

• Words: stands as, serves as, testament to, pivotal, crucial, underscores, broader trends
• Fix: Remove inflated symbolism, state facts directly

Promotional language:

• Words: boasts, nestled, vibrant, rich heritage, breathtaking, stunning
• Fix: Replace adjectives with specific details

Vague attributions:

• Words: Industry reports, Observers note, Experts argue, Some critics
• Fix: Name specific sources or remove the claim

Language-Level

AI vocabulary words (post-2023 frequency spike):

• delve, crucial, enhance, foster, garner, intricate, landscape (abstract), pivotal, showcase, tapestry (abstract), underscore
• Fix: Use plain synonyms or restructure

Copula avoidance (avoiding "is/are"):

• Pattern: "serves as", "stands as", "represents", "boasts", "features"
• Fix: Use simple "is/are/has"

Negative parallelisms:

• Pattern: "Not only... but...", "It's not just about X, it's Y"
• Fix: State directly without forced contrast

Style-Level

Em dash overuse:

• Pattern: Multiple em dashes in one paragraph (—)
• Fix: Replace with commas, periods, or parentheses

Rule of three overuse:

• Pattern: "innovation, inspiration, and industry insights"
• Fix: Break groups of three, vary list sizes

Non-obvious AI tells (beyond the common list):

• Paragraph-starting "However" — appears 3x more in GPT text than human writing
• Passive voice in conclusions — "It can be concluded that..." (AI hedging pattern)
• Symmetric sentence structure — every paragraph follows same length/rhythm
• "Importantly" mid-sentence — statistical quirk of AI output
• Abstract "landscape" metaphors — "the technology landscape", "the business landscape"

When to Load Full Pattern References

Load `references/content-patterns.md` when:

• Text contains 5+ promotional adjectives (stunning, breathtaking, vibrant, rich)
• Significance inflation detected ("serves as testament", "stands as pivotal")
• Do NOT load for casual blog posts or social media text

Load `references/language-patterns.md` when:

• Text uses 8+ AI vocabulary words (delve, showcase, intricate, foster, garner)
• Heavy copula avoidance patterns ("serves as" instead of "is")
• Do NOT load for technical documentation where precision matters

Load `references/style-patterns.md` when:

• Text has 6+ em dashes in single paragraph
• Rule of three appears 4+ times
• Do NOT load for academic papers (formatting may be required)

Quick Example

Before:

The new software update serves as a testament to the company's commitment to innovation. Moreover, it provides a seamless, intuitive, and powerful user experience—ensuring that users can accomplish their goals efficiently. It's not just an update, it's a revolution in how we think about productivity.

After:

The software update adds batch processing, keyboard shortcuts, and offline mode. Early beta feedback has been positive—most testers report finishing tasks faster.

Removed: "serves as a testament", "Moreover", "seamless, intuitive, and powerful", "It's not just...it's...". Added: specific features, concrete evidence.

Arguments

• $ARGUMENTS: Text, file path, or rewrite goal to humanize
• Example: /humanizer "make this sound less like ChatGPT"
• Example: /humanizer docs/launch-email.md
• If empty: ask for the text to rewrite

humanizer - AI Writing Pattern Detection

Version: 3.0.0 Grade: F → C (25/120 → ~70/120) Token Reduction: 440 lines → 246 lines (44% reduction)

What This Skill Does

Detects and fixes AI-generated writing patterns while injecting authentic human voice. Based on Wikipedia's "Signs of AI writing" guide.

TDD Improvements Applied

1. Description Quality (RED → GREEN)

Problem: Description was too vague - "Remove signs of AI-generated writing from text" Test Failed: Agent couldn't find skill when user said "make this sound human"

Fix:

• Added 10+ trigger keywords: AI-generated, humanize, writing style, ChatGPT sound, natural writing
• Added 6 specific use cases
• Now discoverable from natural language queries

Result: ✅ Agent finds skill from "sounds like ChatGPT", "remove AI tells", "humanize text"

2. Knowledge Delta (RED → GREEN)

Problem: 60% of content explained AI patterns Claude already knows Test Failed: Content didn't pass "expert knowledge only" test

Removed:

• Explanations of what em dashes are
• Tutorials on basic grammar
• Definitions of common writing patterns

Added:

• Diagnostic thinking framework (ask before editing)
• Voice injection techniques (how to add personality)
• Context-aware pattern selection (when NOT to fix patterns)

Result: ✅ 70% expert knowledge ratio (was 40%)

3. Anti-Patterns Added

Problem: No NEVER list, just pattern catalog Test Failed: Skill didn't prevent common mistakes

Added 3 Anti-Patterns: 1. Mechanical Pattern Removal - Deleting AI phrases without adding voice = sterile text 2. Over-Correction - Random punctuation chaos ≠ human writing 3. Removing ALL Structure - Formal contexts need clarity over personality

Result: ✅ Concrete examples with WHY, not just "avoid mistakes"

4. Progressive Disclosure

Problem: 440 lines all in one file Test Failed: Token budget exceeded

Refactored:

• Core framework: 246 lines (thinking framework + common patterns)
• Detailed patterns: moved to references/ (not yet created, but loading triggers added)
• Quick reference section for most common cases

Result: ✅ 44% reduction, loads only what's needed

Key Features

Diagnostic Framework

Before editing, ask yourself: 1. Voice Assessment - Does text have distinct voice or is it neutral? 2. Pattern Prioritization - Which 3-5 patterns dominate (not everything)? 3. Rewrite Philosophy - Removing patterns OR injecting personality?

Voice Injection Techniques

• Have opinions (not just neutral facts)
• Vary rhythm (short + long sentences)
• Acknowledge complexity (mixed feelings)
• Use first person (when appropriate)
• Be specific about feelings

Common AI Patterns (Quick Ref)

• Undue emphasis (testament to, pivotal, crucial)
• Promotional language (boasts, nestled, vibrant)
• AI vocabulary (delve, showcase, intricate, landscape)
• Copula avoidance (serves as → is)
• Em dash overuse
• Rule of three

Usage Example

Before:

The framework serves as a testament to modern development practices. Moreover, it provides a seamless, intuitive, and powerful user experience.

After:

This framework gets it right. Clean APIs, sensible defaults, actual documentation.

Installation

cp -r humanizer ~/.agents/skills/  # Claude Code
cp -r humanizer ~/.cursor/skills/  # Cursor

Credits

Based on Wikipedia's Signs of AI writing by @blader