Humanizer

A Claude skill that rewrites AI-generated (or AI-sounding) text so it reads like it was written by an experienced human — without changing its meaning, facts, names, dates, numbers, or quotations.

Originally based on blader/humanizer; v3.0.0 is a substantial extension built on top of that foundation, adding explicit fact-preservation guardrails, automatic style/register detection, a structured output format, and a deterministic pre-scan script.

Features

• 24+ AI-writing patterns detected, from Wikipedia's "Signs of AI writing" taxonomy — significance inflation, promotional language, vague attributions, overused AI vocabulary, em dash overuse, formulaic conclusions, and more
• Fact-preservation guardrails — names, dates, numbers, and quotations pass through untouched; no invented statistics or examples, ever
• Automatic style detection across nine registers (academic, technical, documentation, business, email, casual, storytelling, marketing, professional reports), each with its own guidance on what to flex and what to keep fixed
• Self-audit pass — after the first rewrite, the skill asks "what still makes this obviously AI-generated?" and revises once more against its own answer
• Structured output — every rewrite comes with a categorized Summary of Improvements (Flow, Tone, Readability, Repetition removed, Grammar, Sentence diversity), so you can see exactly what changed and why
• Deterministic pre-scan script (scripts/scan_ai_patterns.py) for a quick word/phrase-frequency report on long documents before the creative rewrite pass

Installation

Claude Code

mkdir -p ~/.claude/skills
cp -r humanizer ~/.claude/skills/humanizer

Claude.ai / Cowork

Upload the packaged humanizer.skill file — the Save skill button installs it to your profile.

Usage

Ask Claude to humanize text directly:

Please humanize this text: [your text]

Or just describe the goal — "make this sound less like AI," "remove the AI patterns from this draft," "this report reads too robotic" — the skill triggers on any of these without needing the literal word "humanize."

For a long document, you can also run the pre-scan yourself first to see where the tells cluster:

python scripts/scan_ai_patterns.py my_draft.txt

Contents

humanizer/
├── SKILL.md                    — skill definition: constraints, workflow, output format
├── README.md                   — this file
├── CHANGELOG.md                — version history
├── CLAUDE.md                   — architecture notes for future editors
├── references/
│   ├── patterns.md             — full 24+ pattern catalog with before/after examples
│   └── style-guide.md          — per-register guidance for all nine styles
└── scripts/
    └── scan_ai_patterns.py     — deterministic pre-pass hit-count scanner

Overview

Based on Wikipedia's "Signs of AI writing" guide, maintained by WikiProject AI Cleanup — a taxonomy built from observing thousands of instances of AI-generated text.

"LLMs use statistical algorithms to guess what should come next. The result tends toward the most statistically likely result that applies to the widest variety of cases." — Wikipedia:Signs of AI writing

Removing those statistically-likely-but-flat patterns is only half the job. The skill also actively writes toward the things that make text sound like a specific person wrote it: varied sentence rhythm, an actual opinion where one belongs, and register-appropriate voice — see SKILL.md Step 3 and references/patterns.md's "soulless vs. has a pulse" example.

References

• Wikipedia: Signs of AI writing — primary source for the pattern catalog
• WikiProject AI Cleanup — maintaining organization
• blader/humanizer — original upstream skill this was built on

See CHANGELOG.md for full version history.

License

MIT