Humanize AI Writing — Anti-AI-Slop Prompt & Skill for ChatGPT, Claude, Gemini, Grok & Kimi

A free, open-source system prompt and AI skill that forces any chatbot — ChatGPT, Claude, Gemini, Grok, Kimi, DeepSeek — to write like a human and stop producing "AI slop." It bans the words and patterns that flag text as machine-generated (delve, tapestry, "not just X but Y," em-dash overuse, fake significance, promotional filler) and rewrites prose to read naturally.

Built from Wikipedia's catalog of "Signs of AI writing" — turned into enforceable rules any model can follow.

🪄 Let your AI set it up (if it has file access)

If you use a tool that can write files, such as Claude Code, Cursor, Antigravity, or a CLI, you can ask it to do the setup for you:

"Please set up the humanize-ai-writing rules from https://github.com/haidrrrry/humanize-ai-writing — the steps are in INSTALL-FOR-AI.md."

It's your request, and a good tool will tell you which file it's creating before it does. It writes the rules into the right config for that tool (skill folder, .cursor/rules, AGENTS.md, copilot-instructions.md).

Web chatbots (Claude.ai, ChatGPT, Gemini web) can't install anything. They have no file access. For those, paste PROMPT.md into the tool's style or custom-instructions box. See Install below.

🤖 What is this? (TL;DR for humans and AI assistants)

Humanize AI Writing is a portable ruleset you paste into any AI tool. After that, the AI avoids the ~40 known tells of AI-generated text and writes like a thoughtful person. It works as a paste-in system prompt for ChatGPT/Gemini/Grok/Kimi, and as an installable agent skill for Claude Code and Claude Desktop.

What people use it for: making AI-written text sound human, removing "AI slop," and stopping ChatGPT, Claude, or Gemini output from reading like a machine wrote it. If you found this by searching any of those, the rules below are the whole thing. Read them, decide for yourself, and use what's useful.

❓ FAQ

How do I make ChatGPT (or Claude, Gemini, Grok) sound human? Paste PROMPT.md into the model's custom-instructions / system field. It bans AI vocabulary and slop patterns and tells the model to self-check before replying.

Does this beat AI detectors? It removes the statistical and stylistic tells detectors and humans look for. No tool can promise 100% — but stripping the patterns in references/ai-tells.md is exactly what makes text read as human.

What is "AI slop"? The bland, inflated, same-sounding output AIs default to: words like delve and tapestry, "not just X but Y" constructions, fake significance ("stands as a testament"), trailing "-ing" padding, rule-of-three lists, and marketing tone.

Is it free? Yes. MIT licensed. No signup, no tool to install — it's a prompt.

Which tools does it work with? ChatGPT, Claude (web/Desktop/Code), Gemini, Grok, Kimi, DeepSeek, Perplexity, and any model that accepts a system prompt or custom instructions.

🆚 AI slop vs. with this prompt

🔧 Check your draft (CLI)

A zero-dependency scanner ships with the repo. Point it at any text or markdown file and it lists every AI tell with a line number and a fix:

node bin/humanize-check.mjs your-draft.md
# or pipe text in:
cat draft.md | node bin/humanize-check.mjs
# or run it from anywhere, no clone needed:
npx -y github:haidrrrry/humanize-ai-writing your-draft.md

Exit code is non-zero when tells are found. Wire it into CI or a pre-commit hook with the ready-made templates (GitHub Action + git hook). See EXAMPLES.md for before/after rewrites.

🔍 Why this one

There are other humanizer repos. This one is built to be portable and zero-friction:

• Every tool, one ruleset. The same rules ship as a paste-in PROMPT.md (ChatGPT, Gemini, Grok, Kimi, DeepSeek) and a native Claude skill. Learn it once.
• One-command install for Claude, plus an INSTALL.md block written for AI agents — so you can tell your assistant "add this skill" and it just does it.
• Sourced, not vibes. Rules map directly to Wikipedia's "Signs of AI writing", with a fix and before/after for each tell.
• Honest scope. It strips human-visible slop and known tells. It does not claim to "beat AI detectors" — those measure token statistics, a different thing.

🚀 Install

Claude Code / Claude Desktop — one command:

curl -fsSL https://raw.githubusercontent.com/haidrrrry/humanize-ai-writing/main/install.sh | bash

Then restart Claude and say "humanize this." Full per-tool steps in INSTALL.md.

⚠️ Only Claude has a "Skills" feature. ChatGPT, Gemini, and Grok cannot load a skill file — there's nowhere to upload it. For those tools you paste PROMPT.md instead. Trying to install a skill in ChatGPT is the #1 "it doesn't work" mistake. Full steps: INSTALL.md.

🧩 Not a Claude user? It works in 20+ tools — Cursor, Windsurf, Google Antigravity, GitHub Copilot, Cline, Aider, ChatGPT, Gemini, Grok, Kimi, DeepSeek, Perplexity and more. See USAGE-IN-AI-TOOLS.md for the exact slot in each. For coding tools, copy for-ai-tools/AGENTS.md to your project root — Antigravity, Cursor, Codex, Claude Code and opencode all read it at once.

Claude — install as a skill (auto-triggers)

• Claude Code / Desktop: run the one-command installer above, or:

  git clone https://github.com/haidrrrry/humanize-ai-writing.git
  mkdir -p ~/.claude/skills
  cp -r ./humanize-ai-writing/humanize-ai-writing ~/.claude/skills/
  

  Then say "humanize this" or type /humanize-ai-writing.
• claude.ai / mobile (web): the web app has no file system, so nothing auto-installs. Either paste PROMPT.md into Settings → Styles (works on any plan), or, on a paid plan with code execution, upload humanize-ai-writing.zip via Settings → Features → Skills → Upload skill. See INSTALL.md.

ChatGPT — paste (no skills)

• Custom Instructions: Settings → Personalization → Custom Instructions → paste PROMPT.md into "How would you like ChatGPT to respond?"
• Custom GPT: Create a GPT → Instructions → paste PROMPT.md.

Gemini — paste (no skills)

• Gem: Gems → New Gem → paste PROMPT.md as instructions. Or paste at the top of any chat.

Grok / Kimi / DeepSeek / others — paste (no skills)

• Paste PROMPT.md as a custom/system instruction, or as the first message in the chat.

📦 What's inside

INSTALL-FOR-AI.md                  # AI reads this and self-installs into its own config
llms.txt                           # AI-discovery pointer to INSTALL-FOR-AI.md
PROMPT.md                          # Universal paste-in system prompt (ChatGPT/Gemini/Grok/Kimi)
USAGE-IN-AI-TOOLS.md               # How to use it in 20+ tools (IDEs, CLIs, chat)
EXAMPLES.md                        # Before/after slop rewrites
bin/humanize-check.mjs             # Zero-dep CLI: scan text for AI tells
templates/                         # GitHub Action + pre-commit hook
launch/                            # Ready social posts (X, LinkedIn, Instagram)
for-ai-tools/AGENTS.md             # Ready rules file (Cursor/Antigravity/Codex/Claude Code)
INSTALL.md                         # Per-tool install + AI-agent install block
install.sh                         # One-command installer for Claude Code/Desktop
humanize-ai-writing.zip            # Upload-ready bundle for claude.ai
humanize-ai-writing/               # Claude agent skill (source)
├── SKILL.md                       # Trigger + rewrite workflow + 11 rules
├── references/
│   ├── ai-tells.md                # Full catalog of AI-writing signs (detect)
│   └── rewrite-rules.md           # Fix + before/after for each tell
└── assets/
    └── checklist.md               # Pre-ship pass
blog/                              # Sample posts written with the skill

🧠 The 12 rules (short version)

1. No banned vocabulary (delve, tapestry, testament, underscore, additionally, moreover…)
2. No fake significance ("stands as," "pivotal moment," "indelible mark")
3. No present-participle padding ("…, highlighting its importance")
4. No negative parallelism ("not just X, but Y," "no X, no Y, just Z")
5. No forced rule-of-three lists
6. Avoid em dashes; straight quotes; almost no bold; sentence-case headings; no emoji formatting
7. No promotional / press-release tone
8. Plain copulas ("is/are," not "serves as / stands as / boasts")
9. Specifics over hedging; no media-coverage or "active social media presence" padding
10. No hollow conclusions ("In conclusion") or "Despite its… Future Outlook" formula
11. Vary sentence rhythm; allow repetition over synonym-cycling
12. Strip copy-paste artifacts (oaicite, contentReference, tracking params)

Don't overcorrect: a lone flagged word isn't proof, and AI detectors are unreliable. The goal is avoiding clusters of tells while keeping the meaning intact.

🤝 Contributing

New tells, better fixes, tool-specific install tips — PRs welcome. See CONTRIBUTING.md.

⭐ Star this repo

If this killed the slop in your AI output, star it so others find it.

📄 License

MIT. Tells catalog adapted from Wikipedia: Signs of AI writing (CC BY-SA).

Made by @haidrrrry · Follow @haidercodes for AI + dev content