AI-generated text has tells. Specific words, rhythms, and sentence structures that trained readers and detection tools catch. The Humanizer skill strips those patterns out, using research from Wikipedia’s WikiProject AI Cleanup to identify and rewrite 24 documented AI writing habits.

Here’s how to install it, use it, and get the most out of it.

What it actually does

The skill scans for 24 specific patterns from the WikiProject AI Cleanup research. These aren’t vague heuristics. They’re documented and named.

A few examples: significance inflation (“groundbreaking,” “transformative,” “paradigm shift” — words AI reaches for when summarizing anything), vague attribution (“experts say,” “many believe” with no source), vocabulary tells (“delve,” “utilize,” “facilitate” appearing at rates no human matches), and hedge stacking where qualifiers pile up instead of stating a position.

The skill flags each instance, explains which pattern it matched, and rewrites the section with a natural alternative. You keep the meaning. You lose the robot fingerprint.

Installation

Clone it into your skills directory:

git clone https://github.com/blader/humanizer.git ~/.claude/skills/humanizer

That’s it. No config files, no API keys. Start a new Claude Code session and ask the agent to list skills — you should see humanizer in the output.

Basic usage

Paste or reference your text and tell the agent to humanize it:

/humanize

[your text here]

The skill runs an eight-pass audit. Each pass targets a different category of AI pattern. You get the rewritten text plus a summary of what changed and why.

You can also point it at a file:

/humanize src/content/blog/my-draft.md

This reads the file, runs the audit, and rewrites in place. Check the diff before committing — the skill optimizes for naturalness, and occasionally a precise technical term gets swapped for a more casual synonym.

The pattern report

Before the rewrite, the Humanizer produces a pattern report. Honestly, this is the most useful part — even if you plan to rewrite manually, it shows you exactly where the AI fingerprint lives.

A typical report:

Patterns found: 7
- Significance inflation (3): "groundbreaking," "transformative," "game-changing"
- AI vocabulary (2): "delve," "comprehensive"
- Hedge stacking (1): "it's worth noting that this could potentially"
- Filler transition (1): "moreover"

Each flagged phrase gets a line number and a suggested replacement. You can accept all rewrites or cherry-pick.

Pairing with AI detection

The workflow I’ve seen work best:

  1. Generate your draft with an AI agent
  2. Run /humanize to strip patterns
  3. Run an AI detection check to score the result
  4. If the score is still high, run /humanize again on the flagged sections

Two passes usually does it. If the score stays above 2/10 after that, the problem is structural — paragraph rhythm, section shape — not vocabulary.

The Humanizer also pairs well with the Brand Voice Guardian skill. Humanizer removes the AI. Brand Voice Guardian adds your team’s tone back in. Different jobs, good sequence.

When to use it

Published content — blog posts, marketing copy, client-facing docs, landing pages, email sequences. Anything that will be read by people who notice (or run detectors).

Skip it for internal docs, READMEs, code comments, technical specs, and first drafts you plan to rewrite by hand anyway. The eight-pass audit takes 15-30 seconds on a 1,000-word post. Worth it for published work. Overkill for a Slack message.

Tuning the output

The default mode rewrites aggressively — every flagged pattern gets replaced. Sometimes you want a lighter touch.

You can run it selectively: after the pattern report, tell the agent which categories to fix and which to leave alone. “Fix the significance inflation and AI vocabulary, but leave the transitions.” It respects that.

Or let it run the full pass, review the diff, and revert specific changes where the original phrasing was intentional. This works well in Claude Code where you see the edit diff before accepting.

Combining with other skills

The Long-Form Writer handles structure and argument flow. The Humanizer handles voice. Generate a structured post, then strip the AI patterns.

For headlines, the Headline Writer generates variants and the Humanizer cleans the top picks. AI-generated headlines lean on significance inflation hard — “groundbreaking” and “revolutionary” show up constantly.

Social copy from the Social Copy Writer usually needs the lightest pass, since social writing already trends informal.

Common issues

If the rewrite changes your meaning, it’s usually in technical content where a precise term got swapped for a more natural synonym. Review the diff, revert those specific changes. The skill optimizes for naturalness, not domain accuracy.

If the score is still high after humanizing, run a second pass. Some patterns nest — a hedge stack wrapped in significance inflation takes two passes to unwind.

And sometimes the Humanizer flags a word you actually wanted. The 24 patterns are statistical flags, not rules. Humans do use “crucial” sometimes. If the word fits your context, put it back. The skill catches the pattern; you make the call.

FAQ

Does this work with agents other than Claude Code? Yes — Claude Code, Cursor, Windsurf, and any universal-compatible agent. Installation paths differ slightly, but the core skill is the same.

How is this different from just asking an AI to “rewrite in a human voice”? A generic rewrite prompt doesn’t target specific patterns. The Humanizer uses a documented research base — 24 named patterns from WikiProject AI Cleanup — so it catches things a general prompt misses. It also produces a pattern report showing exactly what it found, which you don’t get from a vague instruction.

Can I use this in a CI pipeline? Yep. The pattern report output is structured enough to parse programmatically. Score the output, set a threshold, block merges that fail. The exact implementation depends on your CI setup, but the output is consistent enough to automate around.