What prompt makes Claude sound human?
The prompts that work give Claude your own writing to copy, a named reader, and a few hard rules; a bare "write like a human" line does far less, because it tells Claude nothing about whose voice to use.
- Two or three samples of your own writing. Paste them in and ask Claude to match their sentence length, word choice and level of formality.
- The reader. "A tired manager reading on a phone" produces different prose from "a professor grading a seminar paper".
- Banned phrases. Anthropic itself names two habits AI models lean on: the "this isn't X, it's Y" construction and the word "quietly". Ban those, plus long dashes and any filler you dislike.
- Facts locked in. Tell Claude to keep every name, number and claim exactly as given, so style work does not drift into new content.
- A length cap. A word limit stops the padding that makes drafts sound machine-made.
- Edit, do not draft. Hand Claude your rough version and ask for edits. Your ideas and order survive, and the voice stays closer to yours.
A short prompt you can copy and fill in:
Edit my draft below for [reader]. Match the voice of my two samples: their sentence length, plain words and tone. Keep every fact, name and number unchanged. Vary sentence length, cut filler, stay under [N] words. Do not use long dashes, the word "quietly", or "this isn't X, it's Y". Return only the edited draft.
Source: Anthropic explainer, FAQ on AI detection software (read 26 Oct 2026). The prompt is our own wording.
Why does a style prompt make Claude's text read as more human?
Because it strips out the phrasing habits that readers, and wording-based checkers, notice first.
Those habits are surface patterns: favourite constructions, a fondness for certain words, sentences of the same length in a row. A prompt that bans them and supplies a real person's rhythm removes most of what makes a draft feel generated.
Anthropic separates this from its own mark. Spotting phrasing patterns, it writes, is "fundamentally different from checking for a watermark." A prompt can work on the first. The second is a different layer, explained in how style detectors differ from a watermark check.
The clues people look for by hand are covered in spotting Claude writing by eye.
What does a prompt change, and what does it leave alone?
A prompt changes which words are likely and how the sentences flow; Claude still makes every choice among near-equal words, and Anthropic's key settles those choices.
Anthropic describes the mechanism in one line: the watermark "uses the key and a few words that come before" to settle which word the model picks. Nothing is added to the text and there are no hidden characters. To a reader, marked and unmarked answers look the same.
Put those two facts together. Your prompt reshapes the list of good candidate words at each step. The key then decides among them, as it does for any prompt. Anthropic does not publish a sentence saying a prompt cannot remove the mark, so this is our reading of its description: on that description, a prompt has no route to the key. The pages we read today list no setting or instruction that turns marking off. The scoring side is laid out in how SynthID-style marks work.
Sources: Anthropic explainer · Claude help center (both read 26 Oct 2026)
Can a prompt reduce how much watermark a text carries?
It can shift the amount as a side effect, for example by forcing fixed wording or a very short answer, but that is a smaller signal, not a way to switch the mark off.
The method behind Claude's mark is a version of SynthID-Text, published by Google DeepMind in Nature on 23 October 2024. That paper ties the strength of the mark to how much freedom the model has at each word, and it lists "the prompts, the temperature and other decoding settings" among the things that change that freedom. More freedom means more room for the mark; less freedom means less.
Anthropic says the same in plain terms. The more Claude writes, the more decisions it makes and the more space there is for a watermark, while small samples are hard to judge and factual passages carry a sparser mark. A prompt that demands a two-line answer or fixed wording therefore leaves fewer marked choices. Three limits matter here: the paper measures the open version, not Claude's undisclosed setup; neither source gives a figure for any prompt; and a long piece written by Claude under any prompt still holds hundreds of Claude-picked words.
Sources: Dathathri et al., Nature 2024 · Anthropic explainer
Does it matter which Claude model or app I use?
No: Anthropic applies the watermark at the model level, so it is present whichever Claude product or surface the text came from.
The help center says this directly, and Anthropic adds that it is also adding watermarks to outputs from models released before 2 August 2026. Switching from the app to the API, or to an older model, is a change of door, not a way around the mark. Anthropic also applies it worldwide, since it says it has no durable way yet to scope it by region.
Source: Claude help center, how Claude marks AI-generated content (read 26 Oct 2026) · Anthropic explainer
Can I check whether my prompted draft still carries the mark?
Not yourself: Anthropic's text check is a detection API in private preview for eligible organizations, and its free Content Checker reads files, not text.
That has a practical consequence. When someone online says a certain prompt "beat" Claude's watermark, there is no public way to confirm it, because nobody outside the key holders can read the mark in text. A style checker scoring the draft as human says something about phrasing, not about the key.
It is also why this page makes no claim about any prompt's effect on the mark. What can be measured is how far the wording moved, which is what our receipt reports after a rewrite.
Sources: Anthropic explainer · Claude Content Checker · Claude help center
What actually changes Claude's wording for real?
New wording does: Anthropic's own answer is that light editing probably won't remove the watermark completely, while "a complete rewrite where every word is replaced" removes it.
Write first, then let Claude edit. When Claude only proofreads text a person wrote, it makes few word choices of its own, so there is little for a mark to attach to. This is also the most honest route when the ideas are yours.
Rewrite with a separate engine. The rewrite tab of our Claude watermark remover hands your text to our own rewriting engine, which produces fresh wording, and gives you a receipt showing words in and out and the share of your original three-word runs it kept. It needs an email sign-in and at least 20 words per run.
Rewrite by hand. A person putting the draft into their own words works too; it just takes longer.
The free hidden-character cleaner above needs no sign-in, but Claude's text has no hidden characters to strip, so on its own it leaves the wording as it was.
Source: Anthropic explainer, FAQ on editing (read 26 Oct 2026)
A style prompt is the right tool when you want Claude to sound like you. When you want the words themselves to be new, rewrite. Your first four rewrites of up to 250 words are free after sign-in, and yearly billing brings plans down to $12 a month. Compare rewrite plans