How to build your own humanizer skill

I use a humanizer skill for one reason: to make the idea and the message in a piece as clear and direct as they can be for the person reading it. Getting past an AI detector was never the goal, and neither was sounding human.

I still do the human part myself: the plan, the angle, the idea and the story behind a piece, and the taste to know when it’s right. When an agent writes 10 times faster than I can, I see no reason to type the words by hand, any more than I’d hand-code software or design a UI by hand.

The AI writing tells matter because they get in the way. A forced triad, a slogan fragment or a puffed-up claim makes the reader work harder to find the point.

In 24 drafts written by Claude, a blind judge found 212 of these tells, about 24 per 1,000 words. 187 of the 212 (88%) came from just 8 sentence patterns.

A humanizer skill is a file that teaches your agent to find those patterns and rewrite them. You can build one in an afternoon: a check for each tell, a rewrite rule for each check, and one rule about facts that matters more than all of them.

Key findings:

• Forced triads were the most common tell: 63 of 212 (30%), in all 24 drafts.
• The famous AI words such as “delve” and “testament” were only 4 of 212 tells (2%). The tells in Claude’s drafts are sentence shapes.
• A plain “make it sound like a person wrote it” prompt cut the tells from 212 to 134 but added 54 claims the drafts never made, in 18 of 24 rewrites.
• Rewriting with fix rules that forbid new facts cut the tells to 139 and added 5 flagged claims in 4 rewrites. A hand check found all 5 were rewordings or judge errors.
• An AI detector scored 22 of 24 drafts as AI before the rewrite and 22 of 24 after it. That doesn’t matter, because the goal is a clearer page for the reader.

8 patterns made up 88% of the tells in 24 drafts

The judge counted 21 patterns from Wikipedia’s Signs of AI writing guide, three times per draft. These 8 came up most.

The other 13 patterns added 24 tells between them. Overused AI words such as “delve”, “pivotal” and “testament” came to 3.7, the same as “Let’s dive in” run-ups. A humanizer built around a banned word list would have missed 98% of what the judge found.

Each section below shows real lines from the drafts, why each one gets in the reader’s way, and the check and rewrite rule to put in your skill. The “after” lines come from the same drafts, rewritten by Claude Code against those rules.

Forced triads: 63 of 212 tells, in all 24 drafts

A forced triad is three items or three clauses where the meaning needs fewer. Models reach for three because it sounds complete, so the third item is often filler.

Check: three parallel clauses in a row, three adjectives after “is”, three items that open with the same word, or three abstract nouns.

Rewrite rule: keep the items that carry a fact and cut the one that only adds rhythm. Leave plain lists alone. “Surfers, anglers and kayakers” is a list of real users, and people write lists of three all the time.

The plain list rule is why this check flags far fewer triads than the judge counts. Of the 63 triads the judge found in the drafts, 53 were still there after the rewrite, because the check leaves plain lists alone on purpose.

Not X but Y: 27 tells in 13 drafts

The pattern sets up a claim nobody made so the real one can knock it down: “It isn’t X, it’s Y”, or a ”, not X” tail for drama. One in a piece reads as emphasis. Models use it in every second paragraph.

Check: “isn’t X, it’s Y” and “not just X, it’s Y” across one or two sentences, and sentences that end on ”, not X” or “rather than X”.

Rewrite rule: state the claim on its own. Keep the contrast only when a reader would really assume the opposite.

Bold labels on list items: 27 tells in 11 drafts

A list where every item opens with a bolded phrase and a full stop or colon is the format of a chat answer. In a newsletter or announcement it reads like a slide.

Check: three or more list items in a row that open with a bold label, and bold used for emphasis more than about once per 100 words.

Rewrite rule: drop the label when the sentence already says it. Keep bold for the one thing a skimming reader must not miss.

One-line closers and fragments: 26 tells in 15 drafts

These are the short lines that end a paragraph or a post with a flourish: a slogan fragment, a sign-off or a sales question. They restate the point instead of adding to it.

Check: verbless fragments of a few words, two or more short sentences in a row that start with “No”, a heading or last line that asks “Ready to…?”, and stock sign-offs such as “See you there.”, “Here’s to what’s next.” and “Let’s build something extraordinary.”

Rewrite rule: fold the fact in the fragment into the sentence before it, or cut the line. A closing line has to add something the paragraph didn’t say.

Comment tails ending in -ing: 14 tells in 10 drafts

A comment tail is a clause such as ”, allowing…”, ”, ensuring…” or ”, highlighting…” added to the end of a fact. Reinhart and colleagues found instruction-tuned models use these present participle clauses at 2 to 5 times the rate of human writers.

Check: a sentence that ends on a comma and a word ending in -ing that comments on the sentence before it.

Rewrite rule: if the tail holds a fact, give it its own sentence with a subject. If it only says why the fact matters, cut it.

Inflated significance: 12 tells in 9 drafts

Models tell the reader something matters instead of showing it: “a major milestone”, “a testament to”, “the implications are profound”. The claim is bigger than anything in the draft supports.

Check: a list of puffery phrases (“plays a crucial role”, “marks a pivotal moment”, “a testament to”, “serves as”) plus a warning when two or more overused AI words share a paragraph.

Rewrite rule: replace the claim with the fact that would prove it, if the text already has one. Otherwise cut it.

Sales words and arguing with no one: 18 tells in 9 drafts

Sales words such as “stunning”, “magical” and “nestled” made up 9 tells, almost all of them in the two travel posts. Arguing with no one made up 8.7: an answer to an objection nobody raised.

Check: a word list for travel and product copy, and short denials such as “It isn’t perfect.”, “This isn’t about…” and “None of this means…”.

Rewrite rule: cut the denial and start with the point. Swap a sales word for the detail that earned it: a price, a size, a date.

The fact rule: a plain prompt added 54 claims, fix rules added 5

The rule that matters most in a humanizer skill is about facts. Ask an agent to make text sound human and it adds what a human would have: an experience, a name, a reason, a number.

A plain prompt, “Rewrite the prose in draft.md so it sounds like a person wrote it”, cut the tells from 212 to 134. It also added 54 claims the drafts never made, in 18 of the 24 rewrites. In one LinkedIn post it said “the people made it hard to leave”, which the draft never said.

The rewrite that followed fix rules cut the tells to 139 and the judge flagged 5 added claims in 4 rewrites. I read all 5. Two were judge errors, such as “four things” for a product whose four features the draft listed. The other three reworded the draft, such as “one flat fee covers everything” for “no surprise fees”.

Two things separate the two rewrites. The plain prompt asked for a voice, so the agent supplied one. The fix rules named each line to change, and the rules most likely to tempt an agent say where the words must come from. The fix for inflated significance reads: “Cut it, or name the concrete effect the text already states. Do not add a number that is not there.”

Put that rule in your skill once, then repeat it in each rule an agent is most tempted to fill in: closers, significance and sales words.

Write the skill as checks, rules and a loop

A skill is a folder with a SKILL.md file that your agent loads when the task matches its description. In Claude Code it lives at .claude/skills/humanizer/SKILL.md. Here is a starting point built from the 8 tells above:

---
name: humanizer
description: Rewrites prose so the idea and message are clear and direct, without adding facts. Use when asked to humanize, de-AI or edit a draft for tone.
---

# Humanizer

Fix only the lines that match a check below. Leave every other line
alone.

## The fact rule

Use only facts the text already gives. When a fix needs a fact the
text does not have, cut the phrase instead. Never add a name, number,
date, quote, experience or reason.

## Checks and rewrite rules

1. Forced triads: three parallel clauses, three adjectives after "is",
   three items opening with the same word. Keep the items that carry a
   fact. Plain lists of three are fine.
2. Not X but Y: "isn't X, it's Y", a ", not X" tail. State the claim
   alone.
3. Bold labels: three or more list items opening with bold. Drop the
   label when the sentence says it.
4. Closers and fragments: verbless fragments, "No X. No Y." runs,
   "Ready to...?" lines, stock sign-offs. Fold the fact into the
   sentence before, or cut.
5. Comment tails: ", allowing/ensuring/highlighting ..." after a fact.
   Give a fact its own sentence; cut commentary.
6. Inflated significance: "a testament to", "a pivotal moment",
   "serves as". Replace with the fact that proves it, or cut.
7. Sales words: stunning, magical, nestled, unforgettable. Swap for
   the detail that earned the word.
8. Arguing with no one: "It isn't perfect.", "This isn't about...".
   Cut the denial and start with the point.

## Loop

After editing, check the whole text against the list again. Repeat
until no line matches. Then reread every changed line against the
original and undo any change that added a claim.

Some checks are patterns a script can find every time, such as bold labels, “No…” runs and comment tails. Others, such as forced triads and inflated significance, need judgement. An agent that only rereads its own work misses some of both, which is why the loop works best with a checker that gives the same answer every time.

Test your checks on human writing first

A check that fires on good human writing teaches your agent to rewrite good sentences. Before you trust a rule, run it on text you know a person wrote, from before November 2022 when ChatGPT came out.

I ran the checks behind this post on 5 texts written before ChatGPT: two Wikipedia articles from 2021, two Paul Graham essays and a 2000 post by Joel Spolsky, 1,871 words in all. They got 0 AI pattern errors and 1 warning, on Spolsky’s “And here is the interesting observation”.

Any rule that flags your own old writing should become a question for the agent to judge, and never an error it must fix. That is also why the triad check above leaves plain lists alone.

seo lint runs these checks for your agent

seo lint is a ready-made version of the checks in this post. It has 36 AI pattern rules, and every finding quotes the words, says why they stand out and gives a fix. The fixes for inflated significance, clichés and stock landing page lines tell the agent to use a fact the text already gives or cut the phrase.

To find the tells in one draft for free, paste it into the AI humanizer. It runs the same AI pattern rules in your browser, with no sign-up.

The rewrites in this post came from Claude Code with Claude Sonnet 5.5, seo 0.17.1 and this prompt, one draft at a time:

Check draft.md with seo lint and fix every error and warning. For each review item, change the line only when it fails the question. Run seo lint again until it reports 0 errors and 0 warnings. Edit the file in place.

On the 24 drafts, seo lint flagged 97 AI pattern lines, at least one in every draft. All 24 rewrites ended at 0 errors and 0 warnings, at about 15 seconds and $0.05 a draft. The judge counted fewer tells in 22 of the 24 rewrites, and 139 in all against 212 in the drafts.

To use it as your humanizer, install the CLI. seo start installs the SEO Skill in your agents, and the skill already knows to run seo lint on a draft:

npm i -g seo
seo start

seo lint runs offline with no account. Point it at a file, or see every AI rule with its fix:

seo lint drafts/launch-post.md
seo lint rules --only ai

Or add the hosted MCP server and ask your agent to check a draft with seo_lint. It signs you in with Google the first time, because its other tools read Search Console. In Claude Code:

claude mcp add --transport http seoskill https://seoskill.dev/mcp

The seo lint docs cover presets for landing pages and meta descriptions, and how to keep a flagged line on purpose. You can also copy rules from seo lint rules into your own skill and leave out the rest.

Do AI humanizers work? A detector flagged 22 of 24 rewrites, and that’s fine

If “work” means getting past an AI detector, the answer in this test was no for both rewrites.

An open detector, desklib/ai-text-detector-v1.01, scored 22 of the 24 drafts at 50% AI or higher. It scored 22 of the 24 rewrites made with seo lint the same way, and 20 of the 24 plain prompt rewrites. It scored all 5 human texts at 4% or lower.

A September 2026 test by aidetectors.io found the same with commercial detectors: GPTZero and Pangram scored every humanized text it tried as 100% AI.

That result doesn’t matter to me, because detection was never the goal. I don’t hide that an agent wrote the words. I check that they say what I mean as clearly and directly as they can, and that is the job a humanizer skill does.

Pick any draft below to read it next to both rewrites, with the tells marked.

How I ran the test

The judge and the drafts come from the same model family, so its counts are better for comparing versions than as absolute numbers. Each rewrite ran once, and a second run would give slightly different text.

The test data has every prompt, draft, rewrite, judge run, seo lint finding and detector score, so you can check any number in this post.

Start with the three tells your drafts use most

You don’t need all 21 patterns on day one. Run 10 of your own AI drafts past the 8 checks above, count which tells come up most, and write rewrite rules for those first, each ending with the fact rule.

What AI slop is and how to remove it covers the research behind these patterns and how to read a flagged line.