Humanizer Claude Skill — AI skill for Claude Code

Humanizer Skill

A Claude skill that takes LLM-generated text — a ChatGPT draft, a Claude response, anything with that flat, over-polished AI cadence — and rewrites it to read like a person wrote it. It also scores any draft 0–100 on how human it reads, so you can check text before you publish it or verify a rewrite actually worked.

It treats "sounding like AI" as three separate problems:

• Register — surface tells: em dashes, delve, uncontracted formal prose, uniform mid-length sentences, three-item lists. These are patternable, so humanizer/scripts/humanize.py catches and fixes them mechanically.
• Genericness — writing that commits to nothing, names nothing, and could be about any company in any industry. No regex catches this one. It needs judgement, which is what the skill's workflow in SKILL.md walks Claude through.
• Structure — what the piece decides, not how it phrases things: whether the narrator explains what it all meant, who resolves the plot, whether emotion is ever named plainly. This is the durable layer. Strip every cliché and purple sentence out of an AI story and a classifier built only on narrative decisions still catches it, because editing out purple prose doesn't change who resolves the plot.

Most "humanize this" requests are really about genericness, even when the person only consciously notices the register. Structure is the layer nobody notices and nothing else fixes.

Research foundation

The structural layer isn't guesswork — it's built from a real study of what separates AI-written fiction from human-written fiction at the level of narrative decisions, not word choice.

**The paper.** Jenna Russell, Rishanth Rajendhran, Chau Minh Pham, Mohit Iyyer (University of Maryland) and John Wieting (Google DeepMind), *StoryScope: Investigating idiosyncrasies in AI fiction*, COLM 2026 (arXiv:2604.03136v6). The team built a parallel corpus — 10,272 human stories from Books3, each mirrored by five LLM