Humanizer — Remove AI Writing Patterns

Your AI writes like an AI. Everyone can tell.

Not because the grammar is wrong. Because the patterns are wrong. Em dashes everywhere. Everything comes in threes. Every announcement is "thrilling." Every product is "innovative, scalable, and reliable." These patterns are documented, cataloged, and detectable. This skill teaches your agent to find and kill them.

Version: 2.2.0

How It Works

This is not a "sound more human" prompt. That is vague and produces vague results. This is a systematic checklist of 24 specific patterns across 5 categories that make writing sound machine-generated. Each pattern has detection rules, words to watch, and before/after rewrite examples.

Your agent loads this skill once and applies it to everything it writes. Blog posts, social media, documentation, emails, App Store descriptions, landing page copy. It becomes a permanent writing editor that catches AI tells before they go live.

The 24 Patterns (5 Categories)

Content Patterns:

• Inflated symbolism ("This represents a paradigm shift")
• Promotional language ("We are thrilled to announce")
• Superficial -ing analyses ("The study, highlighting key trends, showcasing...")
• Vague attributions ("Many experts believe")
• Significance inflation ("This changes everything")

Language Patterns:

• Em dash overuse, the single biggest AI tell
• Rule of three lists ("innovative, scalable, and reliable")
• AI vocabulary words (leverage, utilize, streamline, foster, delve, tapestry, landscape)
• Negative parallelisms ("It is not just X, it is Y")
• Filler conjunctive phrases ("Moreover," "Furthermore," "It is worth noting")

Style Patterns:

• Uniform sentence length and structure
• Sycophantic openers ("Great question!")
• Hedging everything into oblivion
• Lists that always have exactly 3-5 items
• Bolding the first word of every bullet point

Communication Patterns:

• Starting with a summary of what you just said
• Ending with "Let me know if you need anything else"
• Offering unsolicited alternatives after answering
• Thanking people for asking basic questions

Filler Patterns:

• "In today's [X] landscape"
• "At its core"
• "It is important to note that"
• Closing with a call to action nobody asked for

Each pattern includes specific words to watch for, an explanation of why it reads as AI, and concrete rewrite examples.

The Two-Pass Anti-AI Audit

After editing, your agent runs two final checks:

1. "What makes this obviously AI generated?" identifies remaining tells
2. "Now make it not obviously AI generated." final revision pass

This catches patterns that slip through individual checks. The combination of systematic detection plus holistic review produces writing that reads like a human wrote it.

Adding Soul (Not Just Removing Patterns)

Removing AI patterns is half the job. Sterile, voiceless writing is just as obvious as slop. This skill also teaches your agent to:

• Have opinions instead of neutral reporting
• Vary sentence rhythm (short punchy + longer exploratory)
• Acknowledge complexity and mixed feelings
• Use first person when appropriate
• Let some imperfection in, because perfect structure feels algorithmic

Built in Production

Battle-tested at Custom Signal. Our marketer agent runs every piece of public content through this skill. App Store descriptions, X posts, Reddit copy, landing pages, inter-agent messages. If our public writing reads like a human wrote it, this skill is why.

Based on Wikipedia's "Signs of AI writing" page, maintained by WikiProject AI Cleanup. The most comprehensive public catalog of AI writing patterns available.

What This Does NOT Do

• Will not give your agent a personality (that is the Agent Starter Kit, free)
• Will not fix bad ideas or weak arguments. Just bad writing patterns.
• Will not make short-form content longer. It rewrites, not pads.
• Works best with Claude and GPT-4 class models. Smaller models may miss subtle patterns.
• Not a plagiarism checker or factual accuracy tool.

Who This Is For

• Content creators using AI for drafts
• Marketers writing copy with AI assistance
• Developers writing documentation
• Founders posting on X, LinkedIn, Reddit
• Anyone whose audience can tell when AI wrote something

Who This Is NOT For

• You want AI detection (this makes writing pass detectors, not run them)
• You need grammar or spelling checking (use Grammarly)
• Your writing is already good and you just want minor polish

Common Pitfalls

• Over-correcting: Removing ALL em dashes makes writing feel artificially restricted. The goal is natural frequency, not zero.
• Losing meaning: Some AI-flagged phrases are the clearest way to say something. Preserve clarity over anti-AI purity.
• Inconsistent application: Run it on everything or nothing. Partially humanized text is more obvious than fully AI text.

Guardrails

• Preserves the original meaning and intent of the text
• Maintains the target tone (formal, casual, technical) and does not impose a single voice
• Never fabricates information or changes factual claims
• Flags when a rewrite significantly alters the message

File Manifest

• SKILL.md Complete 24-pattern detection guide with examples (~15KB)
• README.md Installation and usage guide (~2KB)

2 files. Drop SKILL.md in your skills directory. Point your agent at it. Watch your writing stop sounding like a robot wrote it.

---

This skill runs on the same framework as the AI Agent Starter Kit (free). For the complete multi-agent writing and publishing pipeline, check out The Full Factory ($199).