The AI Text Humanizer skill for Claude

You probably recognise the problem straight away. You let AI write a first draft, or you write it yourself and ask for an editing pass, and the result is technically fine but reads like a brochure. Every list has three items. Every paragraph opens with Furthermore or Moreover. Something is called essential twice a page and a gamechanger once. Your readers rarely put it into words, but they switch off, and anyone who does recognise it draws a harsher conclusion: this person did not put in the effort.

That is tackled here the way a good copy editor would. Count first, judge afterwards. It scans the text against a fixed catalogue of 18 Dutch AI tells, reports what it found with location and count, rewrites in a second pass, and checks itself in a third. Facts, figures, names and quotes stay exactly as they are. The writer's tone stays, or is restored if the AI version had ironed it flat. And every change is traceable to a pattern code, so you do not have to take it on trust that it got better: you can check it yourself.

This file is written for anyone who publishes Dutch text: business owners who write their own newsletter, marketers editing AI drafts, agencies delivering client copy, and writers who want to know which habits have crept into their own work. It is part of the skill library NL we make freely available through the AI and automation service, with no account and no sales email afterwards.

For English this is thoroughly documented. Wikipedia's AI Cleanup project maintains a list of recognisable AI writing patterns, and the best-known writing skill on GitHub built a catalogue of 33 patterns from it, with a two-round check. Anyone working in English can use that. But running Dutch text through an English pattern list only gets you halfway there, and that is exactly the gap this skill was built for.

Some of the patterns simply translate across. The triple, the negative parallel, the inflated importance: those show up in Dutch AI text too. But some behave differently. Em dashes are more common in English than in Dutch, so the same dash density stands out far more here. And some are typically Dutch AI phrasing that appears in no English list: loan translations like unlock and seamless that no one uses in conversation, the whether-you-are construction that wants to address everyone and reaches no one, and the drift in the form of address, where a text slides halfway through from formal to informal because the model saw both forms in its training data.

That is why this skill works with its own catalogue of 18 Dutch tells, spread across four groups: words and phrasing, sentence structure, structure, and content and tone. The full list is below in FIG.02, with an example and the matching fix for each pattern. How Google itself views AI text, and why recognisability plays a different role there than most people think, you can read in the article Google knows your article was written with AI.

One nuance the skill itself also makes: a pattern is only a problem when it is unintentional. A triple can be exactly right rhetorically. An em dash is correct Dutch. It is about the stacking and the predictability, not isolated occurrences. That is why the skill counts first, and why its report has a separate section for what was deliberately left standing.

And the frontmatter states just as clearly when the skill is not used: not to dodge detectors, not to disguise plagiarism, not to make a text more persuasive than the facts allow.

Then comes the core: the catalogue of 18 Dutch tells you saw above in FIG.02, each with a code, a description, an example and a fix. The method around it consists of three rounds. Pass 1 is the scan: read the whole text and label every occurrence with a code, without judging much yet. Pass 2 is the rewrite, in order of damage, with the hard rule that meaning, facts, figures, names and quotes stay exactly as they are.

Pass 3 is the self-check: reading its own result again with the question of what still gives away a language model, and checking line by line that nothing has been invented and nothing essential has dropped out. That third round comes straight from the best international humaniser skills, and it is not there for show: even a rewrite can slip back into patterns, and a check pass catches that.

Before anything happens, the skill runs through an input checklist: the full text, what it is for, who the reader is, whether it is formal or informal, the tone you want, ideally with two or three sentences of your own writing as a style reference, and what absolutely has to stay, such as legal wording or brand names. It only asks about what is missing.

The output format is fixed and consists of five blocks: the scan report with the number of occurrences per code and a sample line, the rewritten text, the before-and-after overview with a pattern code and one line of explanation per change, the section for deliberately kept for anything that stayed despite a match, and the open questions where the skill could not make a choice without you. Those last two sections are the difference between a black box and an editing pass you can check.

The file closes with eight refusals and a list of sources, from the Wikipedia pattern list to Renkema's Schrijfwijzer. More on that below.

The pattern approach comes from Wikipedia. The AI Cleanup project maintains the Signs of AI writing page, a collection of recognisable patterns built up from thousands of cases of AI text on Wikipedia itself. The core insight from there appears almost word for word in the skill: a language model pulls towards the statistically most likely result that fits the most cases at once, and that is exactly what makes the output recognisable.

The method comes from the viral humaniser skill. The best-known English version on GitHub combines a numbered pattern list with a two-round check: rewrite first, then attack your own result with the question of what still looks like AI. This skill takes that self-check pass over wholesale, because it demonstrably makes the difference between a surface polish and a real rewrite. What this skill adds is the catalogue itself: 18 tells written for Dutch, not translated from English.

The language standard comes from Renkema and Onze Taal. Jan Renkema's Schrijfwijzer has been the standard for clear business Dutch for decades, and Onze Taal's advice on em dashes and forms of address underpins two of the most Dutch tells in the catalogue: T07 and T18. So copywriters already knew that a text without an excess of dashes and without formal/informal drift reads more pleasantly, long before any language model existed.

And the oldest source is Orwell. Politics and the English Language from 1946 describes how ready-made phrases replace thought: write in prefabricated formulations and you no longer have to think about what you are actually saying. That essay was about political language, but forty years before the first language model it already describes precisely what happens to text when the most likely phrasing wins over the truest one.

Even without installing the skill you can check your own text against the catalogue in FIG.02: count your triples, cross out your filler words, and read your closing paragraph out loud. More background sits in the knowledge base on writing and visibility, and the blog on AI recognition by Google gives the search engine perspective on this same subject.

After that, using it is simple: paste your text and say what it is and who it is for. If anything is missing, it asks for it itself, and it is best to include two or three sentences of your own writing as a style reference. Want to learn how an instruction file like this is put together yourself? Read writing a SKILL.md yourself NL: this humaniser is a direct example of it.

There are also situations where you are better off leaving it be. The biggest one is already in the refusals: any scenario where the rewrite has to disguise rather than improve. Detection evasion, plagiarism, schoolwork that has to pass as your own work: the skill does not take part in that, and if you are unsure whether your situation falls under it, that is usually the answer already. It is also not a fact-checker: it guards against facts changing, but it does not check whether they are correct. A text with the wrong figures comes out just as wrong, only more readable. And it is not a translator from a poor offer into a good one: if nobody cares about the content, the form will not fix that.

A perhaps more honest limit: the skill makes your text more human, but your voice comes from you. You get the best results if you give it a few sentences of your own writing and genuinely answer the open questions in the report. In the example above, the skill ends by asking a question back: which three numbers do you actually check first? As long as you answer that question, it stays your newsletter. Would you rather write entirely new text around a keyword than repair existing text, then the AI SEO Copywriter skill NL is the more logical place to start.

A human text keeps readers, but readers have to arrive first. If your site is poorly found, the real work sits before this skill: the free SEO scan shows in seconds where your pages stand, and the improving SEO service picks that up structurally. And if you would like to talk further about what else AI can do for your text and your business, from separate skills to full automation, we simply do that in a conversation.