A writing skill for AI agents that catches and removes the tells of AI-generated text: puffed-up significance claims, "delve"-style vocabulary, em dash overuse, inline-header bullet lists, and about twenty other patterns.
It's built directly from Wikipedia's Signs of AI writing page, maintained by WikiProject AI Cleanup. That page is the closest thing that exists to a field guide for this, put together from thousands of real flagged edits across Wikipedia. This skill turns that field guide into something an AI agent can actually apply to its own output.
This README was written by an AI agent, then edited using this exact skill. If it doesn't read like a press release, that's the point.
- Puffed-up significance ("marks a pivotal moment," "stands as a testament")
- Notability name-dropping ("has been featured in Wired, Forbes, and other outlets")
- Superficial "-ing" analysis tacked onto sentences for fake depth
- Promotional language ("nestled in the heart of," "boasts a vibrant")
- Vague attributions ("experts believe," "researchers say")
- The formulaic "Despite its challenges, X continues to thrive" section
- Era-specific AI vocabulary (yes, the overuse of specific words has shifted by model generation, and the skill tracks that)
- Copula avoidance ("serves as" instead of "is")
- Negative parallelisms ("it's not just X, it's Y")
- Rule of three, everywhere, always
- Title case headings, boldface overuse, inline-header bullet lists
- Em dashes doing the work commas should be doing
- Emoji used as section decoration
- Curly quotes (with the caveat that this one's weak on its own)
- Chatbot leftovers like "I hope this helps!" or "Let me know if you'd like me to expand"
- Knowledge-cutoff hedging and unfilled placeholder text like
[Your Name]
It also documents what's not a reliable tell, perfect grammar, formal vocabulary, and unsourced claims are all things good human writers produce too, so the skill is built to avoid false accusations, not just to strip vocabulary.
Drop SKILL.md into your skills directory:
~/.claude/skills/humanizer/SKILL.md
That's it. Claude Code picks up skills from that folder automatically, and it'll show up as humanizer whenever you invoke it or whenever the description matches what you're asking for.
If you only want it in one project rather than globally, put it at .claude/skills/humanizer/SKILL.md inside that project instead.
Paste the contents of SKILL.md into your custom instructions or rules file (.cursorrules, .windsurfrules, or your tool's equivalent). These tools don't have a separate "skill" concept the way Claude Code does, so the instructions just become part of the standing system prompt.
SKILL.md is plain markdown with a YAML frontmatter block up top. Strip the frontmatter if your tool doesn't use it, and feed the rest in as a system prompt or a preamble before your actual request. It works the same way regardless of which model is on the other end, it's instructions, not a fine-tune.
Once installed, just ask for it by name: "humanize this" or "run this through humanizer." In Claude Code specifically, you can also invoke it directly as /humanizer.
If you want the output to match your own writing style rather than a generic "human" default, give it a sample of your own writing first. It'll read your sentence length, punctuation habits, and word choice, then rewrite toward that instead of a one-size-fits-all voice.
Most "AI detector" tools work by guessing statistically whether text was AI-generated, and they're not that reliable, a 2025 study found humans distinguishing AI text from human text perform barely better than a coin flip. This skill takes the opposite approach: instead of detecting after the fact, it edits before the fact, using the same patterns researchers and Wikipedia editors have already catalogued as tells.
The result isn't just "text that fools a detector." It's text that reads like it came from someone who actually thought about what they were saying, varied their sentence rhythm, and had an opinion instead of a neutral summary. Removing the tells is half the job. Adding an actual voice is the other half, and the skill covers both.
MIT. Use it, fork it, put it in your own agent's skill folder, whatever's useful.
Every pattern in SKILL.md traces back to Wikipedia:Signs of AI writing, which itself cites primary research (Reinhart et al. 2025, Juzek & Ward 2025, Kobak et al. 2025, and others) for the empirical claims about word frequency shifts. This repo just repackages that work into something an AI agent can act on directly.