work-humanizer

A humanizer for work writing. It takes the AI tone out of your text without changing what it says or commits to.

Sales emails, marketing copy, support replies, HR notes, legal letters: at work, the exact words are commitments. Ask an AI to make them sound human and it changes them on the way. It apologizes where you only acknowledged, attaches a condition to an offer, blames the customer. work-humanizer rewrites in a plain human voice and is built to keep every fact, commitment and caveat at the same strength. When a draft is missing what only you know, it asks you a few questions first.

An Agent Skill: works in any agent that loads SKILL.md, including Claude Code, Codex and Cursor.

What the test found

12 new AI-written texts, rewritten by each tool on the same model, judged blind by four AI models from different companies, with every meaning change checked against the original.

• Judges would send work-humanizer's version most often. Shown what the writer meant and asked which rewrite they'd send, they picked work-humanizer 48% of the time and blader/humanizer, the most popular humanizer skill, 20%. A one-line "rewrite this to sound human" prompt got 1%.
• It changes the least. 18 meaning changes across the 12 texts, against 33 for blader/humanizer, 42 for the one-line prompt and 120 for Emulate-1. work-humanizer invented no facts in any of the 56 texts it rewrote across five test sets.
• The one-line prompt reads more human, and that's the catch. It beat the AI draft on 90% of judgments to work-humanizer's 75% because it rewrites freely: it changes more and adds things the writer never said. Judges who could see the original sent it once in 96 picks.
• No rewrite gets past an AI detector. On Pangram, work-humanizer, blader/humanizer and the one-line prompt all leave the text reading as AI (2 to 4% human). Emulate-1, sold as undetectable, reaches 87% by changing the most.
• Your own reasoning does. Given the writer's own notes, work-humanizer's version scored 100% human on Pangram on 4 of 4 texts. It builds the piece on the writer's reasoning (their claims, in their order, with their own because and so) and rewrites sentences where they need it. Word choice matters least: in an earlier test, an AI rephrase of a human article still read 99–100% human.
• Only work-humanizer asks you for what's missing. When a draft lacks facts only the writer knows, the other tools rewrite around the gap, and some make things up. work-humanizer asks a few questions and uses your answers: on 16 texts it asked on 9, used 39 of the writer's facts and invented none, and judges would send its version 69% of the time (blader/humanizer 22%). In the test, an AI answered for each writer from a list of facts they knew; with real people it is untested.

What each tool fixed, and what it changed

When you give it bullets instead of a draft

Most of the time nobody hands AI a draft to fix. They type a few bullets and ask for the email. Here every tool got the same brief and wrote the email itself (3 runs each; run 1 shown, the notes count all 3):

One brief, checked by hand against it, so read it as an illustration rather than a measurement. Data: evidence/brief-to-email/.

Install

Copy the skills/work-humanizer folder into your agent's skills folder: