Humanizer is a free, open-source skill that names all 33 AI writing patterns and pulls them out of your text in two passes. It installs in one command. Your emails, proposals, and website copy stop carrying the fingerprint that says a machine drafted them.
Your client reads your proposal and something feels off. The content is right. The numbers are right. But they cool on it and you never find out why.
Here is why. The draft still sounds like every other AI draft they have seen this month. Same rhythm, same stock phrases, same tidy little lists. They can feel it even when they cannot name it.
You can name it. All 33 of them, with a tool that runs the check for you before anything leaves your hands.
Why Your AI Copy Reads as AI
Every large model learned to write from the same pile of text. Wikipedia, Reddit, Stack Overflow, a decade of Medium posts. They all absorbed the same habits. So they all produce the same habits back.
The em dash dropped in the middle of a sentence for effect. The "it's not this, it's that" flip. The list of three that always saves the important item for last. The formal register nobody uses out loud.
None of this is wrong. It is uniform. That is the actual problem. Your clients and your audience have now read enough AI output that these patterns register as a signal, and the signal reads as low effort. Trust drops a notch, and you never see the moment it happened.
The fix is not to stop drafting with AI. Drafting with AI is too fast to give up. The fix is to strip the fingerprint after the draft and before the send.
What Counts as an AI Tell?
An AI tell is a writing pattern that shows up far more often in machine-generated text than in anything a person writes on their own. Wikipedia's editors catalogued them while cleaning AI content off the site, and Humanizer is built on that catalogue. So the list is not somebody's hunch about what sounds robotic. It comes from thousands of real, hand-checked examples of AI text getting flagged and removed.
That is the difference between this and a find-and-replace macro. A macro deletes em dashes. Humanizer knows the em dash is one of 33 named habits, catches the other 32, and understands why each one gives you away.
The 33 Patterns, Sorted Into Four Buckets
Every pattern has a name and a fix. They fall into four groups.
| Bucket | How many | What it catches |
|---|---|---|
| Content | 6 | Inflated significance, vague attributions ("experts argue"), promotional language, formulaic "challenges and future" sections, surface-level analysis, name-dropping |
| Language | 7 | AI vocabulary ("delve", "leverage", "robust"), dodging the word "is", negative parallelisms ("not X, but Y"), the rule of three, synonym cycling, false ranges, passive voice |
| Style | 10 or more | Em and en dashes, bold overuse, inline headers, title case, emojis, curly quotes, over-hyphenation, persuasion tropes, signposting, manufactured punchlines |
| Communication and filler | 6 or more | Chatbot leftovers ("I hope this helps"), cutoff disclaimers, servile tone, filler phrases ("it's worth noting"), hedging, generic upbeat conclusions |
Read that table once and you will start seeing the patterns in your own drafts. That recognition is half the value. The tool handles the other half at scale, on every draft, without you having to hold 33 rules in your head.
See One Get Fixed
Here is a line an AI might hand you for a proposal intro:
Our team boasts a proven track record and remains deeply committed to delivering robust, data-driven solutions that unlock lasting value for your organization.
Count the tells. "Boasts" instead of "has". "Deeply committed". "Robust" and "data-driven" stacked. "Unlock lasting value". Every one of those is on the list, and a buyer who has read fifty pitches this quarter feels all of them at once. The sentence says almost nothing and sounds like a brochure.
Cleaned, it reads like a person:
We have done this work before, and we build tools your team can actually run day to day.
Shorter, plainer, and it makes a claim you can defend in the next sentence. That is the shape of every fix. Less ceremony, more of something true.
The Mistake That Makes It Fail
The most common way to waste this tool is to run one pass, see the score drop, and ship. One pass fixes the loud tells and often plants a quiet one in the rewrite. That is exactly why the second pass exists. Let both run.
The second common mistake is turning on autofix with no voice sample and walking away. You get text that is human and generic at the same time, which on a proposal reads as forgettable. Load a sample and read the flags. The tool is a second set of eyes, not a replacement for yours.
Install It in One Command
Three methods pull from the same repo, github.com/blader/humanizer. Pick the one that matches how you already work.
Method 1: as an agent skill. This is the fastest path if you run an agent harness like Claude Code or Codex. One command drops the skill in.
Copy the install command:
Copy this.
npx skills add blader/humanizer
Or add it through the plugin marketplace:
Copy this.
/plugin marketplace add blader/humanizer /plugin install humanizer@humanizer
Then invoke it with /humanizer followed by your text, or just tell the agent to "humanize this."
Method 2: manual copy, any harness. Clone the repo and drop the skill file into your skills directory yourself.
Copy this.
git clone https://github.com/blader/humanizer cp humanizer/SKILL.md ~/.claude/agents/humanizer.md
This works with Claude Code, Cursor, Windsurf, or any tool that reads skill markdown.
Method 3: standalone command line. For running it outside an agent, clone and install the CLI.
Copy this.
git clone https://github.com/blader/humanizer cd humanizer npm install npm install -g .
Which one should you use? Here is the short version.
| You are | Use | Why |
|---|---|---|
| Already in an agent all day | Method 1 | One line, no maintenance, invoke inline while you draft |
| On a harness the plugin skips | Method 2 | Same skill file, dropped in by hand |
| Cleaning files in bulk, no agent | Method 3 | Scores and fixes whole folders from the terminal |
All three run the same 33 patterns and the same two-pass logic. There is no "lite" version. Done. The skill is installed.
Run the Two Passes
The first pass strips the obvious. The second pass reads its own output and catches what is still hiding.
That second pass is the whole point. Most cleanup tools run once and stop. The trouble is that fixing one pattern can plant another. You cut a bloated sentence and the rewrite lands on a rule-of-three. Humanizer treats its own first draft as fresh input and runs the check again, so the repair does not sneak a new tell back in.
If you are on the agent skill, paste your text and run it:
Copy this.
/humanizer [YOUR DRAFT PROPOSAL, EMAIL, OR PAGE COPY]
If you are on the command line, you have a few moves:
Copy this.
# Quick score: are you in the danger zone? humanizer score < draft.txt # Full read: see every pattern it flagged humanizer analyze -f draft.md # Fix with suggestions applied humanizer humanize --autofix -f draft.txt # Sweep a whole folder and fail anything too AI-heavy humanizer scan docs --fail-above 45
If the output still reads stiff after both passes, feed it back with a plain instruction: "read every line as if you were saying it to a client on the phone. Cut anything you would never actually say out loud. Return only what is left." That one sentence recovers most of what a mechanical pass misses.
Calibrate It to Your Own Voice
Out of the box, Humanizer makes text sound like a generic human. That is a step up, but generic is not you. Your proposals should sound like your proposals.
Give it a sample first. Two or three paragraphs of your own writing, the real stuff you have already sent, then the draft to clean.
Copy the voice-calibration prompt:
Copy this.
/humanizer Here is a sample of my writing: [2 OR 3 PARAGRAPHS YOU ACTUALLY WROTE AND SENT] Here is the text to clean: [THE AI DRAFT]
Now it matches your sentence lengths, your word choices, your openings, instead of flattening everything toward the middle. For anything high-stakes, a proposal, a landing page, a sales script, load the sample every time. The lift in quality is worth the extra ten seconds.
Read the Score Before You Publish
Humanizer gives your text a 0 to 100 AI-likelihood score. Pattern detection carries 70 percent of that weight. The other 30 percent comes from statistical uniformity, meaning how even and machine-like your sentence rhythm is.
Here is how to read the number.
| Score | Reading | What to do |
|---|---|---|
| Under 25 | Reads human | Ship it |
| 25 to 45 | Some patterns showing | Skim the flags, fix the easy ones |
| Above 45 | AI patterns clearly present | Run the two passes before you send |
| 76 to 100 | Heavily machine-shaped | Rewrite, do not just patch |
Target under 25 for anything a client will read. Human writers land there naturally. If your draft is sitting above 45, it is carrying tells you would rather your prospect not clock.
How I Actually Use It
Every piece of AI-assisted writing goes through Humanizer before it leaves my hands.
I draft most emails, proposals, and site copy with AI because the first draft is fast and structured. It is also, every single time, full of tells. I noticed it in how clients responded. Something read as off even when the content was solid.
So the loop is short:
- Draft with the AI. Fast, on point, structured.
- Run it through Humanizer with my voice samples loaded.
- Read the flagged patterns. Sometimes I fix them myself, sometimes I let the autofix run.
- Read it out loud once before it goes, on anything that matters.
That read-aloud step is the backstop. If I cannot say a line naturally, it reads as AI on the page. Humanizer catches most of it mechanically. My own mouth catches the rest. The two together mean nothing slips out with my name on it.
Honest: Where Humanizer Falls Short
A cleanup pass is not a content check. Humanizer makes writing sound human. It does not make a weak argument strong or a wrong number right. If the draft says the wrong thing clearly, you now have the wrong thing in a very human voice. Judgment stays your job.
It can also over-correct. A clean human writer sometimes hits a few of these patterns on purpose. If you let the autofix run unread on your best writing, it can sand off a phrasing you actually wanted. So read the flags. Do not blind-accept every fix, especially on copy you are proud of.
And scoring is a guide, not a verdict. A 30 is not a rejection and a 20 is not a guarantee. The score points you at the sentences worth a second look. You still decide.
None of this is a reason to skip it. It is the reason to keep yourself in the loop instead of handing the whole thing to a tool. The tool flags. You choose.
If You Do One Thing This Week
Install it with Method 1 and run one real proposal through it. Not a test paragraph. The actual next thing you were about to send.
Watch what it flags. You will recognize half the patterns the moment you see them named, and you will not un-see them in your writing again. That alone changes how you draft.
If it does nothing for you, you spent five minutes and one command. If it does what it did for me, every email and proposal you send from now on stops sounding like it came out of a machine.