Two skills, one shared knowledge base, MIT licensed. Plus the watermark side effect nobody in this category is discussing.
Search volume for "claude humanizer skill" spiked this year and most of the results treat it as an open question or a DIY project. It isn't either. A working, open-source implementation has been published, and you can read every instruction it gives Claude before you install it.
Table of Contents
- Yes, a Claude humanizer skill already exists
- What it's actually doing under the hood
- Why humanizers like this also strip the watermark
- Watermark removal is not detector evasion
- Where this leaves you
1. Yes, a Claude humanizer skill already exists
The repo is walterwritesai/no-slop-ai-humanizer-rewriter, MIT licensed, and it ships two skills rather than one.
Writer (writer/) is instructions for generating prose that avoids AI tells from the start. Use it when drafting articles, emails, or social posts where you'd rather not need a second-pass rewrite at all.
Humanizer (humanizer/) is instructions for taking existing AI output and rewriting it so it reads like a person wrote it.
Both draw from the same reference files, which is the design decision that matters most here. Same word lists, same structural patterns, same before-and-after examples. Whether you're preventing the problem or fixing it, the definition of the problem is identical.
The structure:
├── README.md
├── LICENSE
├── writer/
│ ├── SKILL.md
│ └── references/
│ ├── phrases.md
│ ├── structures.md
│ └── examples.md
└── humanizer/
├── SKILL.md
└── references/
├── phrases.md
├── structures.md
└── examples.md
Setup is manual, and worth being upfront about that. There's no install command. You clone or download the repo, open the SKILL.md for whichever skill you want, and paste its contents into your Claude project instructions along with the reference files. That's more friction than a package install, though it has a compensating benefit: you read exactly what you're loading before it runs, which is not the default experience with most agent tooling.
2. What it's actually doing under the hood
The skill works by naming specific patterns rather than issuing a general instruction to sound human. That distinction is the whole reason it produces different output from a "rewrite this to sound natural" prompt, which mostly gets you a thesaurus pass.
Three reference files carry the actual knowledge.
phrases.md catalogs words and phrases that appear in AI output at significantly higher rates than in human writing, grouped by function: verbs, adjectives, nouns, transitions, and full phrases. Each entry has human alternatives attached, so it's a substitution table rather than a blocklist.
structures.md is the more interesting file, and it's the one that separates this from vocabulary-level tools. It catalogs the sentence and paragraph shapes models default to: setup-payoff pairs, three-item list sentences, hedge chains, reframe constructions, summary closers, and identical paragraph opening patterns. Those are structural tells, and they survive any amount of synonym swapping.
examples.md holds before-and-after rewrites showing what each replacement looks like applied to real prose, which is what keeps the instructions from being interpreted loosely.
The word lists and patterns are derived from Walter Writes' research on common ChatGPT vocabulary, which itself aggregates observations from the writing community. The repo credits stop-slop as inspiration.
For comparison on how this category generally works: Mehul Gupta's publicly documented approach runs a three-pass system across roughly 29 categorized patterns with before-and-after examples for each. Different implementation, same underlying insight, which is that effective humanization is pattern-specific rather than a vibe you can prompt for.
The practical mechanism in all of these is a full rewrite rather than an edit. The skill isn't nudging individual words. It's regenerating prose against a set of constraints about what the output must not contain structurally. That distinction becomes load-bearing in the next section.
3. Why humanizers like this also strip the watermark
Any tool that fully rewrites text clears Anthropic's watermark as a side effect, because the watermark is a statistical pattern across a long sequence of token choices rather than a tag attached to the file. Replace the sequence and there's nothing left to detect. This applies to this repo, to Walter Writes as a maintained hosted implementation of the same mechanism, and to any other tool doing genuine rewriting.
The mechanism, briefly, since it explains the behavior. Claude's watermark is applied at generation time by seeding the sampler with a secret key plus a window of the preceding tokens rather than a general random number. At positions where several next-tokens are roughly equally good, the keyed value biases the choice. One biased choice proves nothing; hundreds across a passage produce a correlation that a verifier holding the key can detect. Nothing is inserted into the text, there are no special characters, and it survives copy-paste because the mark is the word sequence itself.
Two properties follow, and both are relevant to what a humanizer does to it.
The mark is distributed, not located. There is no token you could point at and call the watermarked one. It exists only as a property of the sequence.
Removal scales with how much of the sequence you replace. Anthropic's own documentation says light editing probably won't remove the watermark completely, while a complete rewrite in which every word is replaced will. That's the mechanical reason a full-rewrite humanizer has this effect while a grammar checker doesn't.
Worth stating plainly, because it's the honest framing: this is a byproduct, not a feature. Neither this repo nor any hosted humanizer is built to target the watermark, and describing any of them as a watermark remover overstates what they are. A rewrite clears it for the same reason that rewriting a paragraph clears any property that depended on the original wording.
Independent work supports the scale of the effect. Testing by ETH Zürich's SRI Lab found a baseline paraphraser scrubbed SynthID-Text watermarks from the large majority of passages, characterizing the scheme as harder to spoof but easier to scrub than earlier designs. Forging a mark without the key is very hard. Destroying one is not.
4. Watermark removal is not detector evasion
These are two separate outcomes with two separate mechanisms, and the distinction is the first thing a technical reader should get straight.
An AI detector reads the writing itself, inferring from stylistic patterns like word predictability and sentence-length variance. It holds no key and no privileged information. A watermark check looks for a specific key-derived statistical correlation and isn't inferring from style at all.
Which means a humanizer skill addresses them through different paths. It targets detectors directly, since detectors read phrasing and the skill rewrites phrasing. It affects the watermark only incidentally, as a consequence of replacing the token sequence.
Every tool in this category, including the four most commonly cited implementations, is built and marketed around the first case. None mentions the second.
One practical note on scope: because the watermark only attaches to tokens the model chose, a document's watermark density is proportional to how much Claude actually wrote versus lightly touched. Hand Claude a finished draft for grammar fixes and there may be very little to detect. Ask for 2,000 words from a prompt and the mark is dense throughout.
5. Where this leaves you
If you want a Claude humanizer skill, clone the repo, read the three reference files, and paste the SKILL.md you need into your project instructions. Twenty minutes, no dependencies, and you can audit every rule before it touches your writing.
If you want the same mechanism without the manual setup, hosted tools do the same category of rewrite with a UI attached.
And on the watermark, the accurate summary is narrower than the discussion around it suggests: a genuine rewrite clears it because the original sequence is gone, that's a consequence rather than a capability, and detection access is currently limited to organizations with specific legal obligations under EU law rather than being broadly available. Useful to understand. Not urgent to act on.
Top comments (0)