Humanize Writing

A prompt-based writing editor that rewrites AI-generated text to sound like a real human wrote it.

It catches the patterns that make AI writing obvious — the vocabulary, the sentence structures, the too-perfect tone — and fixes them through a systematic 3-pass editing process.

Works with any LLM: Claude, ChatGPT, Gemini, Cursor, Windsurf, or anything that accepts a system prompt. Also available as a one-command install for Claude Code users.

How It Works

The system runs your text through three passes:

Pass 1: Kill the AI Vocabulary — Replaces words like “delve,” “leverage,” “tapestry,” and 30+ other statistically overused AI words with simpler human alternatives.

Pass 2: Break the AI Structures — Eliminates structural patterns like parallel negation (“Not X, but Y”), tricolons (groups of three), em dash overuse, rhetorical Q+A, and mirror structures. These are stronger tells than vocabulary.

Pass 3: Add Human Texture — Varies sentence length, adds contractions, lets some thoughts stay unresolved, and makes the author’s actual opinion visible.

Also includes special rules for LinkedIn posts and a 14-point quality checklist that runs before returning any rewrite.

Quick Start (Any LLM)

The entire system is two markdown files. You can use them right now:

1. SKILL.md — The main instructions (the 3-pass process, LinkedIn rules, quality checklist)
2. ai-patterns-dictionary.md — The reference dictionary (36+ banned words, 10 structural patterns, tone tells)

Copy the contents of both files into your LLM’s system prompt or custom instructions, then paste any text you want humanized.

Setup by Platform

Claude Code (Plugin Install)

One command, and it works automatically whenever you ask to humanize text:

claude plugin marketplace add https://github.com/lguz/humanize-writing-skill

After that, just say things like “humanize this” or “make this sound less AI” and Claude Code will use the skill automatically.

Claude Code (Manual Install)

Copy the skill files into your Claude Code skills directory:

mkdir -p ~/.claude/skills/humanize-writing/references
curl -o ~/.claude/skills/humanize-writing/SKILL.md https://raw.githubusercontent.com/lguz/humanize-writing-skill/main/skills/humanize-writing/SKILL.md
curl -o ~/.claude/skills/humanize-writing/references/ai-patterns-dictionary.md https://raw.githubusercontent.com/lguz/humanize-writing-skill/main/skills/humanize-writing/references/ai-patterns-dictionary.md

ChatGPT / Custom GPT

1. Go to ChatGPT → create a new GPT (or use “Custom Instructions” in settings)
2. Paste the contents of SKILL.md into the system prompt / instructions field
3. Paste the contents of ai-patterns-dictionary.md below it (or upload it as a knowledge file)
4. Start a conversation and paste any text you want humanized

Cursor / Windsurf

1. Create a rules file in your project (e.g., .cursor/rules/humanize-writing.md or .windsurfrules)
2. Paste the contents of both files into it
3. When editing text in the IDE, reference the rule or ask the AI to humanize your writing

Any Other LLM

Copy the contents of both markdown files into whatever “system prompt,” “custom instructions,” or “context” field your tool provides. The instructions are model-agnostic — they work with any LLM that can follow structured prompts.

What’s Inside

skills/humanize-writing/
├── SKILL.md                              # 3-pass rewriting process + quality checklist
└── references/
    └── ai-patterns-dictionary.md         # Banned words, structural patterns, tone tells

The AI Patterns Dictionary Covers

• Tier 1 Banned Words (18 words) — Strongest AI signals like “delve,” “tapestry,” “pivotal,” “testament”
• Tier 2 Banned Words (18 words) — Moderate signals like “crucial,” “leverage,” “seamless,” “robust”
• Tier 3 Transitions — Words that are fine alone but AI clusters unnaturally (“Furthermore,” “Moreover,” “Additionally”)
• 10 Structural Patterns — Parallel negation, tricolons, em dash overuse, rhetorical Q+A, mirror structures, dramatic reveals, and more
• Tone & Voice Tells — Uniform sentence length, hedging, over-positivity, absence of personal voice
• Content Red Flags — Generic conclusions, missing concrete details, regression to generic statements

Writing Philosophy

Based on the principles behind Paul Graham’s writing style: clarity over cleverness, directness over decoration. Good writing sounds like a smart person thinking out loud. The goal is invisible editing — the reader should never think about how something was written.

Sources

The AI patterns dictionary draws from:

• Wikipedia: Signs of AI Writing
• GPTZero AI Vocabulary Research
• Chris Herbert’s ChatGPT Overused Words
• Sabrina Ramonov’s Humanizer Prompt
• God of Prompt: 500 ChatGPT Overused Words

License

MIT