A text can be correct and read like a machine wrote it. The difference is not a single word; it is a density of patterns, and five of them carry most of the impression. Finding ai text markers is the first step of any honest edit.
Code: humanizer-ua, MIT licensed.
The skill is Ukrainian: 30 hard bans, 12 normative rules, 80 markers in eight groups, and a scanner with no dependencies. The reasoning applies to any language.
Sign 1: importance instead of a fact
“У сучасному світі” and “варто зазначити” promise that something important follows, and then nothing arrives. The sentence inflates a claim that has no detail.
The fix is deletion. The paragraph usually survives without the wrapper, and it reads faster.
Sign 2: an empty contrast
“This is not just an update, it is a new stage.” The negative half names something nobody claimed, and the positive half sounds bigger for it. The same shape appears in every language, and a reader recognises it before they can name it.
A contrast that corrects a belief the reader actually holds earns its place. The rest is decoration.
Sign 3: a saying in the place of a claim
“At its core, this is about people” hides the absence of an argument behind a sentence that sounds deep. The fix is to write the specific claim: who does what, with which consequence.
There is also a test of humility here. If the sentence fits any company in any industry, it is not about this one.
Sign 4: a rhythm built by rule
Three items in every list, a dash in every paragraph, the same sentence length from start to finish. The pattern announces a generator even when every word is right.
Real writing alternates: long sentence, short one, a list of two, a paragraph that stops early. The point is not variety for its own sake but variety that follows the meaning.
Sign 5: the closer that repeats the point
Two weeks of silent schedule is how a business discovers it has no monitoring.
That is the real lesson.
The second line adds nothing, and a text with one after each section reads like a presentation. Removing them is the cheapest improvement available.
Why one marker is not a marker
A person writes “not X but Y” too, and one dash in a long text is a style. The skill’s scanner does not penalise a weak signal when the paragraph has no other evidence, and that rule is the difference between an editor and a search-and-replace.
Density and automation are the tell, which is why the output is three scales rather than one verdict.
What the tool refuses to do
It does not claim a text will pass a detector, does not insert typos, and does not promise invisibility. It edits: removes the wrapper, keeps the fact, and says where the fact-risk is.
A scanner that checks 33 of 80 markers is useful because it names which 33. The rest stays a judgement, and the readme says which part is judgement.
Where it works
In a personal blog where the text is written once a month and the edit takes an evening. On a studio site where the Ukrainian version was translated from Russian and carries Russian word order. And in someone else’s project where the editor asks not for a rewrite but for a list of what gives the machine away.
The last case is the most useful: a list of findings explains to the author why the text sounds off, and that persuades better than the note rewrite it.
Where the list comes from
The catalogue follows the Wikipedia project on signs of AI writing, with Ukrainian-specific additions: Russian syntax in Ukrainian sentences, calques from Russian, surzhyk, and the apostrophe and dash rules that make a translated page look foreign before anyone reads a word.
The skill is MIT, installs into any agent folder, and runs offline.
Each marker comes with a counter-example, so the fix is a rewrite of one sentence, not a ban on the whole draft.
Related: ai writing patterns for the English list, and skills for ai agents for why a reference file beats a promise.