Teaching AI to Forget It's AI: Inside the Humanizer Skill for Claude

Hook

Wikipedia editors have reviewed thousands of AI-generated articles and documented exactly how machines write differently than humans. Someone turned that knowledge into a Claude skill that can detect—and remove—all 29 patterns.

Context

Large language models have an uncanny valley problem. They're good enough to draft documentation, blog posts, and technical content at scale, but not quite good enough to sound genuinely human. The tells are subtle: an overreliance on em dashes for emphasis, phrases like "it's worth noting that" appearing three times in one article, or the peculiar habit of cycling through synonyms ("additionally," "moreover," "furthermore") in consecutive paragraphs.

As AI-generated content flooded Wikipedia, editors at WikiProject AI Cleanup began documenting these patterns systematically. They weren't guessing—they were analyzing thousands of confirmed AI-written articles and cataloging the linguistic fingerprints. The blader/humanizer project takes this real-world curation and turns it into executable instructions for Claude Code and OpenCode. Instead of building a separate detection tool or training a new model, it's a SKILL.md file that teaches Claude to recognize and rewrite its own species' writing patterns. The irony is deliberate: using AI to remove signs of AI.

Technical Insight

The humanizer works as a skill file for Claude-based coding assistants. When loaded, it adds a new directive to Claude's system prompt: analyze text against 29 specific patterns and rewrite accordingly. The architecture is deliberately simple—no API calls, no external dependencies, just a structured markdown file with instructions.

The pattern taxonomy divides AI tells into four categories. Content patterns include significance inflation ("revolutionary," "game-changing"), vague attributions ("studies show" without citations), and manufactured uncertainty ("appears to," "seems to"). Language patterns cover copula avoidance (replacing "is" with "serves as"), synonym cycling, and buzzword density. Style patterns flag em dash overuse, listicle structures, and inappropriate emoji use. Communication patterns catch chatbot artifacts like "As an AI language model" or XML-style tags that leaked from training data.

Here's what the skill execution looks like in practice:

# Input text
It's worth noting that machine learning has revolutionized 
the field of data science. Additionally, it appears to have 
transformed numerous industries—fundamentally changing how 
organizations approach problem-solving.

# Humanizer analysis
Patterns detected:
- Significance inflation: "revolutionized," "fundamentally changing"
- Filler phrase: "It's worth noting that"
- Manufactured uncertainty: "appears to have"
- Em dash emphasis (1 occurrence)
- Sequential transition: "Additionally" following filler

# Rewritten output
Machine learning changed data science by making pattern 
recognition practical at scale. Companies now use it for 
predictions that were previously impossible.

The two-pass approach is crucial. First pass: apply pattern-specific rewrites. Second pass: run an "obviously AI generated" audit with fresh context, checking if the rewrite still feels robotic. This catches higher-order patterns—the prose might avoid individual tells but still have that characteristic AI rhythm where every sentence is perfectly balanced and grammatically optimal.

The voice calibration feature adds personalization. You feed it samples of your actual writing, and it extracts stylistic preferences: sentence length distribution, punctuation habits, vocabulary range, paragraph structure. Instead of producing generic "clean" prose, it targets your specific voice. The skill includes a template for this:

## Voice Calibration
Analyze these writing samples:
[paste your articles/emails/documentation]

Extract:
- Average sentence length and variation
- Paragraph structure preferences  
- Common transition phrases
- Punctuation patterns
- Technical vocabulary density

Apply these characteristics to rewrites.

This transforms it from a pattern-removal tool into a style-matching system. The underlying model isn't just deleting AI markers—it's emulating a target voice based on examples.

The skill file itself is remarkably lightweight, under 500 lines of markdown. It's a testament to prompt engineering: you don't need fine-tuning or specialized models when you can articulate the task precisely. The 29 patterns are enumerated with examples, counter-examples, and edge cases. The skill even includes meta-instructions about what NOT to do—don't make text overly casual, don't remove all formal language, don't add errors for authenticity.

Because it runs inside Claude Code or OpenCode, it has access to file context. You can point it at an entire documentation directory, and it'll process files while maintaining cross-references and technical accuracy. It's not doing blind paraphrasing; it understands the semantic content and preserves meaning while shifting style.

Gotcha

The fundamental limitation is the paradox at its core: you're asking an AI to remove signs of AI writing. Claude is being instructed to avoid patterns that are, in some cases, emergent properties of transformer architectures. It's like asking someone to avoid thinking about purple elephants—the instruction itself creates awareness that can influence output.

There's no validation mechanism. You get a rewritten text, but no confidence score, no detection metrics, no guarantee it'll pass human review or AI detection tools. The 29-pattern checklist is based on Wikipedia's 2023-2024 observations; new AI models may develop different tells. GPT-4's writing patterns aren't identical to Claude's, which differ from Gemini's. The skill is frozen in time, targeting yesterday's AI fingerprints.

The voice calibration feature assumes your writing samples are genuinely human-written. If you feed it AI-generated examples, it'll learn to match AI patterns while thinking it's matching human style. There's also the risk of overfitting—if your samples are too consistent, the rewrite might sound like a parody rather than natural variation. For high-stakes content where AI detection has real consequences, this is a prompt template, not a guarantee.

Verdict

Use if: you're a technical writer who drafts with AI but publishes under your name and wants to remove robotic tells before human review. Use if you're contributing to Wikipedia or platforms where AI-generated content requires disclosure. Use if you have a body of human-written work and want AI assistance that matches your existing voice rather than generic clean prose. Use if you're comfortable with manual quality checks and understand this is a revision tool, not a magic eraser. Skip if: you need guaranteed undetectability for high-stakes contexts like academic integrity situations—the lack of validation makes this unsuitable for adversarial scenarios. Skip if you're writing for audiences that don't care about AI markers or contexts where AI generation is openly accepted. Skip if you want set-it-and-forget-it automation; this requires thoughtful input samples and review of outputs. Skip if you're working with AI models other than Claude in coding assistant contexts—the skill is specifically designed for Claude Code and OpenCode integration.