blader/humanizer

Humanizer – What It Does

Humanizer is a skill/plug‑in for AI‑powered coding assistants (Claude Code, Codex, Gemini CLI, GitHub Copilot, etc.). Its purpose is to take text that was generated by a large language model and rewrite it so the prose sounds like a real person wrote it, while preserving the original meaning. It does this by detecting 26 “AI‑writing tells” (e.g., “Not X but Y”, forced triads, over‑used buzzwords, chat‑bot residue) and then producing a cleaned‑up version that follows a more natural human style.

How to Install

The repository provides a Skills‑CLI package that can be added to any supported agent.

Agent	Install command
Claude Code	`/plugin marketplace add blader/humanizer` `/plugin install humanizer@humanizer`
Codex	`npx skills add blader/humanizer --global --agent codex`
Claude.ai / Claude Desktop	Download the repo as a ZIP and upload it via the Settings → Skills UI
Other agents (Gemini CLI, GitHub Copilot, Windsurf, etc.)	`npx skills add blader/humanizer --global --agent '*'
`

Add --global to make the skill available to all projects, or omit it to install locally.

Using the Skill

Once installed, invoke it with the /humanizer command:

/humanizer

[paste AI‑generated text]

You can also ask in plain language:

Please humanize this text: <your text>

To rewrite a whole file, give the file path:

Humanize the prose in docs/launch‑post.md

Voice‑matching

If you provide a short sample of your own writing (2‑3 paragraphs), Humanizer will mimic your rhythm, word choice, punctuation, and even personal quirks such as dash usage.

Core Mechanics (From the README)

Detect tells – The engine scans the input for any of the 26 patterns, ranking the strongest first.
Draft a rewrite – It creates an initial version that removes or rephrases the detected tells while keeping the factual content unchanged.
Critique & refine – The draft is checked again against the pattern list; any remaining artificial‑sounding fragments are flagged, and a final version is produced.
Safety – Humanizer never invents new facts. If a sentence needs a name, date, or quote that isn’t present, it asks the user instead of hallucinating.
Selective editing – When given a file path, only prose sections are altered; code, front‑matter, data tables, and link targets stay untouched.

The 26 Patterns (Brief Overview)

The README groups them into six categories:

A. Staging instead of stating – e.g., “It’s not just X, it’s Y”.
B. Rhythm by rule – forced triads, repeated sentence starts, over‑use of dashes, stacked qualifiers, excessive hyphenation, passive voice.
C. Inflation & borrowed authority – buzzwords, inflated significance, vague associations, sales‑y language, citing unnamed experts.
D. Formatting by rule – decorative bold, emojis, heading case, curly quotes.
E. Leftovers from chat/draft – chatbot sign‑offs, knowledge‑limit disclaimers, meta‑talk about the document.
F. Writing for the wrong reader – re‑explaining what the audience already knows.

Each pattern includes a “before” (AI‑style) and an “after” (humanized) example.

Example Transformation

Before (AI‑heavy):

"I’m thrilled to announce that shared drafts are finally here! 🚀 … Let that sink in."

After (Humanizer):

"Shared drafts are out today. For months our own team passed around files named final_v7.docx, so we built a way for two people to edit the same doc at once, with each person's changes showing up live for the other. … It's free on every plan."

In a blind test, judges preferred Humanizer’s rewrite 16/16.

License & Community

License: MIT – free for personal or commercial use.
Stars: Badges show current GitHub stars.
Support: Issues such as #229 discuss evaluation results; the repo links to the Wikipedia page that defines the pattern list.

TL;DR

Humanizer is a plug‑in that cleans up AI‑generated prose, making it read like a human author by stripping out common “AI tells”. It works with many LLM‑powered agents via the Skills CLI, can be voice‑matched with a user sample, and guarantees factual fidelity by never hallucinating new information.