CLAUDE WATERMARK REMOVER.
Two things travel with text copied out of Claude: invisible Unicode characters and a recognizable way of writing. Careful removes the first, Thorough removes both. This tutorial covers each mode, shows a before-and-after example, and explains how to verify the result yourself.
01 / Definition
There is no visible badge on Claude output. When people say “Claude watermark” they usually mean one of two very different things, and they need very different treatment.
The first is hidden Unicode: real characters in the string that render as nothing. The second is the wording pattern — the vocabulary, rhythm and structure that make prose read as machine-written. Careful removes the first; Thorough removes both.
02 / Scope
If a CMS field breaks, a diff shows changes in an apparently identical line, or a search fails on a word you can plainly see, you have hidden characters. Careful mode fixes that in one pass.
If a colleague says the text “sounds like AI,” there are no characters to delete — the phrasing itself is the mark. That is what Thorough rewrites away. For the background, see what a Claude watermark is.
03 / Mechanics
Zero-width and formatting characters sit between the letters you can read. They occupy positions in the string, so they can be counted, located and deleted exactly. They survive plain-text copying, which is why they turn up in a CMS or a Git diff long after the original paste. This layer is deterministic: the same input always produces the same clean output.
A statistical watermark biases which words a model picks, spreading a signal across hundreds of tokens rather than storing it in any single character. Nothing is hidden in the string — the visible words are the signal. A verifier needs the original key to test for it, and no amount of Unicode stripping touches it.
The two layers need different treatment, which is why there are two modes. Careful clears the character layer and leaves your sentences exactly as written. Thorough also rewrites the wording, removing the pattern that lives in the visible words. Either way the report tells you exactly what changed, instead of a bare “done” message.
04 / Careful removal
Open the remover and paste the Claude response, drop a .txt file, or load the built-in sample. Paste directly from Claude rather than by way of a rich text editor — a round trip through Word or Docs can add its own marks on top of the ones you are trying to remove.
Careful scans every character, removes the supported hidden marks outside emoji sequences, and applies NFC normalization. Your sentences come out identical, character for character, apart from the deletions listed in the report.
Each finding names the character, its U+ code point, how many times it occurred and where. Marks flagged as sensitive — joiners and variation selectors — are worth a second look if your text contains emoji or a joined script such as Arabic or Persian.
Copy the result, or download it as a text file. Paste it into its destination as plain text so nothing is reintroduced on the way.
05 / Thorough rewriting
The rewrite reworks the phrasing that carries the pattern, while preserving names, numbers, URLs, citations, code, emoji and paragraph structure. Your facts and structure survive; the Claude-sounding wording does not.
Generative rewriting is nondeterministic: run it twice and you get two different outputs. Check facts, names, quotations and formatting every time before you use the text.
Thorough removes the Claude wording pattern from your text. That is not the same as predicting how any particular AI-detection classifier will score the result, which no tool can promise. If accuracy matters more than style, use Careful and edit the wording yourself.
06 / Example
A tidy draft[ZWSP] can still carry invisible edits. This soft[SHY]hyphen and the stray byte mark here[BOM] will be removed, along with the plane emoji ✈[VS16] left intact.
A tidy draft can still carry invisible edits. This softhyphen and the stray byte mark here will be removed, along with the plane emoji ✈️ left intact.
The wording is identical. Careful deleted three characters and changed nothing you can read — “softhyphen” closes up because the soft hyphen between the words is gone. The fourth mark, the variation selector after the plane, is shown in grey: it belongs to an emoji sequence, so it was reported and kept rather than removed.
07 / Coverage
| Property of your text | Careful | Thorough |
|---|---|---|
| Hidden Unicode outside emoji | Removed | Removed |
| Canonically equivalent sequences | Normalized | Normalized |
| Emoji joiners and presentation | Preserved | Preserved |
| Your visible wording | Untouched | Rewritten |
| Line breaks and paragraphs | Untouched | Preserved where possible |
| The Claude wording pattern | Not addressed | Removed |
| Authorship or originality | Unchanged | Unchanged |
Character-level detail: every supported code point, with its behavior in both modes, is listed in the Claude hidden Unicode character reference.
08 / Things people try first
For hidden Unicode, only if you retype the text by hand. A paraphrasing tool that ingests your string and returns a new one usually drops the invisible characters as a side effect, but you have no report and no way to confirm it — and some tools happily copy the marks straight through. For a statistical watermark, heavy paraphrasing does weaken the signal, because the signal lives in the word choices you are replacing. Light paraphrasing generally does not.
No. Both preserve zero-width characters, soft hyphens and joiners, and a round trip can add marks of its own such as byte order marks and smart-quote sequences. “Paste as plain text” strips formatting, not code points. If you want to use Word as a cleaning step, save as .txt and inspect the result — do not assume it worked.
Partly, and less than people expect. NFC and NFD only reorganize canonically equivalent sequences; a zero-width space stays a zero-width space in all four normalization forms. NFKC is more aggressive and does strip several formatting characters, but it also rewrites text you probably wanted to keep — ligatures, superscripts, full-width characters and non-breaking spaces all change. Targeted removal plus NFC, which is what Careful applies, is the predictable option.
It works, if you already know exactly which characters to look for and your editor accepts code point escapes. That is the manual version of what Careful does across the whole supported set in one pass, and it is easy to miss the variation selectors.
Yes for hidden Unicode — an image or freshly typed text carries no invisible characters at all. It also discards every advantage of having text, and it does nothing about wording patterns.
Careful alone will not: it clears the invisible characters, and AI-detection classifiers work on phrasing, which they never looked at. Thorough removes the wording pattern as well, which is the layer those classifiers read. No tool can promise a particular score from a particular detector, so treat any site that guarantees one with suspicion.
09 / Verification
Run Careful in the remover and switch on “Reveal Unicode input.” Every invisible character appears in brackets at its exact position, in context. This is the fastest inspection, and the Claude watermark detector page walks through reading the output.
Paste the text into any character counter, then into a plain .txt file. If the counts disagree, or the count is higher than the visible text can account for, something invisible is in there.
VS Code renders unusual invisible characters with a highlighted box by default and flags them in the problems panel. In Sublime Text, Vim and most programmers' editors you can display or search for the raw code points directly.
A few lines of Python list every hidden character in a file with its position and Unicode name:
python3 - file.txt <<'EOF'
import sys, unicodedata as u
text = open(sys.argv[1], encoding='utf-8').read()
for i, c in enumerate(text):
if u.category(c) == 'Cf' or 0xFE00 <= ord(c) <= 0xFE0F:
print(i, hex(ord(c)), u.name(c, '?'))
EOFNo output means the file is clean. Category Cf covers the zero-width and formatting characters; the range check adds the variation selectors.
Hidden characters come back whenever text passes through another system. Verify the final copy in its destination, not the version sitting in your clipboard.
10 / Questions
Cleaning invisible characters out of text you are entitled to use is ordinary text processing, the same as stripping formatting. What you may do with the resulting text is governed by whatever rules apply to it — your school's policy, a publisher's contract, Anthropic's terms. A tool cannot grant permission you do not have.
In Careful mode, no. The characters removed have no visual representation, so the rendered text is identical apart from a soft hyphen no longer offering a break point. Emoji keep their joiners and stay intact.
Careful handles long documents without a practical ceiling. Thorough has a character limit per request because it calls a language model; split very long pieces into sections and review each one.
Yes. Removal is driven by code points, not by which assistant produced the text, so output from any tool — or from a PDF, a website or a word processor — is handled identically.
Text is processed in memory for the length of the request. Careful never leaves this server; Thorough sends the text to OpenAI only when you choose that mode. The privacy notice has the details.
The exact processing order, preservation rules, report fields and limits are documented in the tool documentation. Shorter answers live in the Claude watermark FAQ.