Humanizer skill for Claude Code, Codex and ChatGPT
A free Agent Skill that makes your AI agent edit like a person: it removes the signs from Wikipedia's "Signs of AI writing" guide, plus the tells that showed up with GPT-6 in 2026. One file, one-line install.
Download SKILL.md Check text for free
Install
Claude Code
mkdir -p ~/.claude/skills/humanizer curl -o ~/.claude/skills/humanizer/SKILL.md https://gounslop.com/humanizer-skill/SKILL.md
Codex
mkdir -p ~/.codex/skills/humanizer curl -o ~/.codex/skills/humanizer/SKILL.md https://gounslop.com/humanizer-skill/SKILL.md
Cursor
mkdir -p .cursor/rules curl -o .cursor/rules/humanizer.mdc https://gounslop.com/humanizer-skill/SKILL.md
ChatGPT, Claude.ai, Gemini
Paste everything below the --- header into Custom Instructions, a Project, or a Gem. Then: "Humanize this: …"
The skill
---
name: humanizer
description: Rewrite or write text so it reads like a careful human wrote it. Removes the signs of AI writing listed in Wikipedia's "Signs of AI writing" guide and gounslop's 2026 measurements. Use when asked to humanize, de-AI, unslop or edit prose, or when writing docs, emails, posts or copy.
---
# Humanizer
Edit for a reader, not a detector. Keep every fact, name, number, quote, link and code block. Never add claims, anecdotes, typos or slang to "sound human".
## 1. Content
- Cut inflated significance: "pivotal moment", "enduring legacy", "stands as a testament", "reflects broader trends".
- Cut notability padding (lists of coverage, "online presence") and vague attributions ("experts say", "observers note"). Name the source or drop it.
- Cut "-ing" commentary tails: ", highlighting its importance", ", underscoring the need for".
- No "Challenges and future prospects" wrap-ups. End when the content ends.
## 2. Words
- Replace AI vocabulary with plain words or delete it:
2023-25 set: delve, tapestry, testament, intricate, pivotal, underscore, vibrant, crucial, align with, enhance, fostering, showcase, emphasizing, highlighting, additionally.
2026 (GPT-6) set: challenge, clearer, dependable, echoed, foster, leverage, matters, multifaceted, practical, prioritize, quietly, steady, universally.
- Use is / are / has instead of "serves as", "stands as", "boasts", "features", "marks a".
## 3. Sentence patterns
- No negative parallelisms: "it's not X, it's Y", "X isn't A. It's B.", "not just X, but Y", "not X, but Y", "Y rather than X". Say the point directly.
- No reflexive lists of three. Use the true number of items.
- No corrective disclaimers ("should be treated as X rather than Y", "does not establish", "should not be confused with") unless the reader truly needs the warning.
- Vary sentence length. GPT-6-style text sits at 8-20 words per sentence; let some run long and some stay short.
## 4. Chat residue and formatting
- Delete offers and sign-offs: "If you'd like, I can...", "Let me know if...", "I hope this helps", "Great question".
- No **bold run-in labels**, "###" headings or bullet lists in prose unless the content is a real list or the format requires it.
- Em dashes: at most one per paragraph; prefer commas, periods, colons or parentheses. (ChatGPT's app default rarely uses them now, but GPT-5.5 to GPT-6.1 via the API and Claude still do.)
## 5. Check before returning
- Re-read for any pattern above. If a sentence only survives because it sounds nice, cut it.
- Confirm every number and name from the input is still there.
Online checker (free, highlights all of the above): https://gounslop.com/ai-slop-detector
Rewrite tool with before/after slop score: https://gounslop.com/
Humanizer skills compared
| Skill | Based on | 2026 updates | Checker |
|---|---|---|---|
| blader/humanizer | Wikipedia's Signs of AI writing | Depends on repo updates | No |
| hardikpandya/stop-slop (guide) | Banned phrases and structures | Depends on repo updates | No |
| gounslop unslop skill | Our 100+ measured slop phrases | Yes | Yes |
| gounslop anti-slop | Count, fix, re-count with numeric caps | Yes | Yes |
| gounslop humanizer (this page) | Wikipedia guide + our Nov 2026 model measurements | GPT-6 vocabulary, "rather than" disclaimers, per-model em dash data | Yes |
They aren't exclusive. Many people run a skill while drafting and then check the result with the AI slop detector.
Why the 2026 updates matter
Older lists lead with "delve" and em dashes. In our November 2026 test, ChatGPT's default model used 0.75 em dashes per 1,000 words, but the GPT-5.5 to GPT-6.1 API models used 3.8 to 9.2 (human answers to the same questions: 0.21). Wikipedia moved em dashes to its historical section and added the GPT-6 vocabulary and "rather than" disclaimers. A skill that only bans dashes misses half of what current models do, and one that ignores dashes misses API output.
Skill vs. the online tool
The skill keeps an agent's own writing clean. For finished text from somewhere else, the unslop tool rewrites it in your tone and shows a before/after slop score and a meaning check.
The skill and checklists are free. For finished text from ChatGPT or Claude, the unslop tool rewrites with a slop score. After the 3 free daily rewrites: $5 for 50 or Pro $12/month.
FAQ
What is a humanizer skill?
A SKILL.md file (an Agent Skill) that teaches an AI agent like Claude Code, Codex or Cursor to edit or write text without the signs of AI writing. The agent loads it when you ask it to humanize, edit or write prose.
Which humanizer skill is best?
The best known are blader/humanizer on GitHub (built on Wikipedia's Signs of AI writing) and hardikpandya/stop-slop. This one is free too and adds what changed in 2026: the GPT-6 vocabulary list, "rather than" corrective disclaimers, and em dashes treated by model rather than as a universal sign. Try two on the same draft and keep the one whose output you'd sign.
How do I install it in Claude Code?
mkdir -p ~/.claude/skills/humanizer, then save SKILL.md from gounslop.com/humanizer-skill/SKILL.md into that folder. Restart Claude Code and ask it to "humanize README.md".
Does it work with Codex or ChatGPT?
Codex reads Agent Skills from ~/.codex/skills/<name>/SKILL.md; save it there or paste it into AGENTS.md. In ChatGPT or Claude.ai, paste the body into Custom Instructions or a Project's instructions.
Will a humanizer skill get text past AI detectors?
We don't claim that and you shouldn't rely on it. The goal is text that's pleasant to read. Using it to pass off AI work as your own where that's prohibited breaks our acceptable use policy.
Is this the same as stop slop or anti slop skills?
Same idea, different rule sets. Stop-slop focuses on banned phrases and structures; humanizer skills follow Wikipedia's guide. This file combines both with our 2026 measurements.