Human Eyes

human-eyes

A [Claude Code skill](https://docs.anthropic.com/en/docs/agents-and-tools/claude-code/skills) for auditing and rewriting prose that carries common AI-writing tells.

human-eyes is a pattern detector, not an AI detector. It flags specific patterns in text that statistically give readers the "AI ick" when used without intent.

Inspired by [blader/humanizer](https://github.com/blader/humanizer), built around a deterministic grader, an iteration harness, and a corpus of matched human and AI essays for testing.

Install

Clone (preferred)

Get updates with `git pull`.

git clone https://github.com/amaezey/human-eyes.git ~/.claude/skills/human-eyes

Claude Code CLI

npx skills@latest add amaezey/human-eyes

Codex and Claude Code desktop

Download the latest release zip from .

Why this exists

Much of the text we consume and create today is generated by AI. While there are some obvious signs of AI writing, there are subtler tells that are harder to diagnose and repair. The effect is cumulative: AI fingerprints on the voice, rhythm, structure, and phrasing all compound.

The patterns aren't the problem. Good writers use triads, em dashes, and contrast framing on purpose. AI uses them without intent. human-eyes views AI writing as inferior to human writing but accepts the reality that it is here to stay. Rather than rewriting blind (like most AI writing skills), it checks your text against a catalogue of research-backed and house-style patterns and gives you options to act on each flag.

Surface patterns are detected by a deterministic Python script. Deeper problems, like structure and tone, are judged against a fixed list of questions, so the model cannot improvise, and the grader rejects answers that are missing or no longer match the text.

What it does for you

When you invoke the skill and provide text, human-eyes runs a complete **Audit** first: a summary