work-humanizer skill
A skill that removes AI tone from work writing while keeping every fact, commitment and caveat at the same strength.
- 129 ★ on the repository
- Last commit
- MIT
Minimal access
Why this level
- The skill edits only the text you give it, runs nothing and sends nothing, and the `check.py` script optionally reads your file.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/work-humanizerDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewMinimal accessThe tool runs nothing and makes no network calls.
- The skill edits only the text you give it, runs nothing and sends nothing, and the `check.py` script optionally reads your file.
This is third-party code. Review the repository files before installing.
What it does
work-humanizer edits sales emails, marketing copy, support replies, HR and legal letters, documentation and posts. In rewrite mode it first lists every fact with its certainty level, then writes the text fresh in a plain human voice and ends with a sentence-by-sentence meaning audit. It plants no typos, slang or invented anecdotes. In interview mode the skill asks up to five cheap questions when a draft is AI-composed and thin: a number, an example or what went wrong is missing. The optional `check.py` mechanically sweeps dashes, semicolons, stock vocabulary, filler transitions and sentence-length spread. The README includes a blind test: four AI judges on 12 texts and a count of meaning changes.
Who it is for. Managers, marketers, support staff and lawyers whose wording in a letter equals commitments and for whom an AI draft cannot be edited blindly.
Good fit when
- A client email draft sounds machine-made, and promises and conditions must stay the same.
- You need an email assembled from a few bullets with no facts added.
- A landing page or post was written by AI and you want the author's voice back from their own notes.
- You want to run a text through a script for stock AI tells before sending.
Not a fit when
- You need to evade AI detectors: the author reports that such rewrites still read as AI on Pangram and only your own reasoning helps.
- You need fiction or academic prose: the skill is built for work writing.
- There is no text to edit: the skill does not invent facts.
Example request
Use work-humanizer on this and interview me first: a customer email about a deposit refund.Limitations
The skill does not guarantee a text passes an AI detector: in the author's Pangram test rewrites read as 2 to 4 percent human. Results improve when you give your own notes or a voice-memo transcript: on such material the skill's version scored 100 percent human on 4 of 4 texts. Interview questions were tested with an AI answering for the writer from a fact list, and the mode is untested with real people. The comparison used the Claude Sonnet model and 12 texts, and other models may differ. The `check.py` script needs Python.
How to disable. Delete the `work-humanizer` folder from the agent skills directory: `~/.claude/skills/` or `.claude/skills/` for Claude Code, `~/.codex/skills/` for Codex.
FAQ
How do I install work-humanizer in Claude Code?
Run `npx skills add phuryn/work-humanizer` or copy the `skills/work-humanizer` folder into `~/.claude/skills/` for all projects or `.claude/skills/` for one. For Codex use `~/.codex/skills/`.
How do I call the skill?
Tell the agent "Use work-humanizer on this" and paste the text, or add "and interview me first" for interview mode. You can name a file such as `docs/launch-post.md`.
How is it different from a plain "make it human" prompt?
The skill changes less and adds no facts: in the author's test 18 meaning changes across 12 texts against 42 for a one-line prompt. That prompt reads more human but rewrites freely and mixes in things the writer never said.
Does it help pass an AI detector?
No, by the author's data rewrites still read as AI. Results improve when the text is built on your own reasoning and notes, and worsen when only someone else's draft is edited.
What does check.py check?
The script looks for dashes, semicolons, stock AI vocabulary, filler transitions, pause-and-point lines and uniform sentence lengths. Run it with `python skills/work-humanizer/check.py draft.md`.
Discussion
No questions or reviews yet.
Related
SkillsA skill that strips AI writing tells from text using Wikipedia's list, without adding invented facts
Two skills for English prose: humanize rewrites a model draft in a natural voice, and ai-check flags machine-writing tells
A skill that finds AI writing tells, rewrites the text in a chosen voice, and optionally scores it from 0 to 100
An editor for AI-written Russian text: removes bureaucratese, cliches and 38 AI patterns while keeping facts intact
A skill that edits German AI drafts: it removes stock phrases, keeps facts and the author's voice and checks 72 language patterns.
Humanizer (brandonwise)
humanizer
A skill and Node.js CLI that score English text for AI-writing patterns and statistics and suggest fixes on a 0 to 100 scale
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| work-humanizer (this card) | no | 129 | |
| Humanizer | no | 68,355 | a skill based on the Wikipedia list of AI writing tells; here the focus is work writing with commitments plus an interview mode |
| Humanize | no | 622 | two skills for English text, rewriting and checking; here one skill with rewrite and interview modes and a published blind test |
README in short
The README opens with blind-test results on 12 texts: the share of versions judges picked, the number of meaning changes and the detector result. It continues with install, call examples, a description of rewrite and interview modes, the list of AI tells it catches and the data in `evidence/`.
README badge
Original SKILL.md
--- name: work-humanizer description: | Edit AI-sounding text so it reads like a careful person wrote it, without changing what it says or commits to. Built for work writing where the wording matters: sales and marketing copy, support replies, HR and legal letters, emails, docs, release notes, posts, landing pages. Two modes. Rewrite: edit the text you are given. Interview: when a draft is AI-composed and thin, ask the writer 3-5 sharp questions and build the piece from their answers, in their words. Also writes from the writer's own notes or a voice-memo transcript. Calibrated on measured differences between human and AI writing. --- # work-humanizer Make the text read like a specific person wrote it for a specific reader. Keep every fact. Invent nothing. Three things we measured shape everything below: 1. **Readers notice texture.** Emphasis with no information in it, uniform rhythm, generic claims, and sentences that do the reader's thinking for them. Those are fixable by editing. 2. **Detectors read composition, not words.** A human-written article stayed 99-100% human on a commercial detector after AI formatting, an AI copyedit, and a full AI sentence-by-sentence rephrase. The same piece composed by AI from its outline scored 68% human, and a style-targeted rewrite of that AI version moved it one point. In our October 2026 test, rewrites of AI drafts by this skill, by blader/humanizer, and by a bare prompt all stayed AI on that detector, while versions built from the writer's own sentences scored human roughly in proportion to how many of those sentences survived. An assistant model rewording a draft does not change who composed it. (A paraphrase model trained against detectors did flip the score, and in the same test it cost meaning and quality.) Never promise "undetectable". If the writer needs text that is genuinely theirs, use interview mode. 3. **Specifics carry the human signal.** Density of numbers, names, and dates was the one signal that tracked reader response across the creators we studied. You may only use specifics that are in the source or that the writer gives you. ## Pick the mode **Rewrite mode** (default). The user gives you text; you edit it. **Interview mode.** Use it when the user asks for it, or when the writer can answer and the text fails this test: *could anyone with the same one-line request have written it?* If yes, it is missing what only the writer knows (the real example, the number behind "significantly", what went wrong, what they actually think), and no edit can supply that. This is the usual case for posts, newsletters, essays, bios, talk abstracts, pitch emails, landing pages, job posts, and launch stories. Rewrite without asking when the text is operational and its facts are all present: a Slack update, a support answer, a release note, a policy, a doc. ## Interview mode Ask one round of at most 5 questions, then write. - **Each question fills a gap a reader would feel.** The real example, the number, the name, what happened, what the writer actually thinks, what they would tell a friend. Quote the line the answer will replace.