GitHub repo humanizer lists 33 AI writing tells

A GitHub project called humanizer packages 33 specific writing tells into an agent skill file. It picked up 1,504 stars in the week ending July 27, per GitHub, reaching 31,581 total.

Anyone shipping AI-drafted copy under their own byline has a checklist for the patterns that give it away, from inflated claims to stray curly quotes.

Per the skill file, the list splits into four categories. Content problems: significance inflation, vague attribution, promotional language dressed as analysis. Language habits: copula avoidance, like 'serves as' instead of 'is', and hedges stacked on hedges. Style tells: em-dash overuse, title-case headers, emoji, curly quotes. Communication artifacts: sycophantic tone and filler that pads without saying anything.

The skill runs on plain Markdown, so it drops into Claude Code, Cursor or Codex without extra tooling. It also adds voice calibration and a no-fabrication rule, so rewrites can't introduce facts, numbers or claims absent from the original.