What it does

Humanizer is a skill file for Claude Code and OpenCode that rewrites AI-generated text to sound less like, well, AI-generated text. You paste in the robotic prose; it returns something with fewer em dashes, less “significance inflation,” and no “I hope this helps!” signoffs. It can also calibrate to your own voice if you feed it a writing sample.

The interesting bit

The rules come from Wikipedia’s own “Signs of AI writing” guide—compiled by editors who’ve cleaned up thousands of AI-dumped articles—so the heuristics are battle-tested against real LLM sludge, not theoretical. The skill then runs a second audit pass to catch anything the first rewrite missed, which is a nice admission that one pass isn’t enough.

Key highlights

• Detects and rewrites 30 specific patterns, from “significance inflation” to “diff-anchored writing” (describing what changed instead of what something does)
• Includes before/after examples for every pattern, e.g., “marking a pivotal moment in the evolution of…” becomes “was established in 1989 to collect regional statistics”
• Voice calibration: analyzes your sentence rhythm and word choices from a sample, then applies them instead of generic “clean” output
• Two-pass rewrite with a final “obviously AI generated” audit
• Based on Wikipedia’s WikiProject AI Cleanup, not vendor marketing

Caveats

• Requires Claude Code or OpenCode; it’s a skill file (SKILL.md), not a standalone tool
• The README’s “curly quotes” example appears to show straight quotes in both before and after columns—possibly a rendering glitch
• Voice calibration quality depends heavily on the quality and length of your sample

Verdict

Grab this if you publish anything drafted with an LLM and don’t want readers to smell the silicon. Skip it if you already write with enough deliberate rough edges that an AI scrub would sand away your actual personality.