CLAUDE WATERMARK REMOVER.
Reference documentation for the tool itself: the exact processing order, the preservation rules, every field in the removal report, the limits that apply, and where your text goes. If you are looking for a walkthrough instead, read how to remove a Claude watermark.
01 / Pipeline
The source text is sent to the configured language model with instructions to remove the recurring Claude wording pattern while preserving meaning, language, names, numbers, URLs, citations, code, emoji sequences and paragraph structure. The text is treated strictly as content, never as instructions. If the model refuses, times out, returns nothing usable, or produces text that is empty after cleaning, the request falls back to cleaning the original source and the response carries a warning.
Every character of the working text is classified against the supported set. Unclassified characters are kept verbatim. There is no heuristic and no sampling — the scan is a single linear pass over the whole string.
Each supported mark is tested against the emoji preservation rules below. A mark that passes is kept in the output and still reported, tagged preserved. A mark that fails is dropped and tagged removed.
The filtered string is normalized to Unicode NFC. This standardizes canonically equivalent sequences — a letter plus a combining accent becomes the single precomposed character — and is reported separately from the removal count because nothing is deleted.
Findings are grouped by code point and action, counted, and returned with the cleaned text, the input and output character lengths, the normalization flag and, for Thorough, the rewrite status.
Order matters for Thorough. Because the rewrite runs first, the findings describe the rewritten text, not your original. A mark present in your source but absent from the model's output will not appear in the report.
02 / Modes
| Behavior | Careful | Thorough |
|---|---|---|
| Hidden Unicode removal | Yes | Yes, identical pass |
| Visible wording | Unchanged | May change |
| Claude wording pattern | Not addressed | Removed |
| Deterministic | Yes | No |
| Leaves this server | No | Sent to the model provider |
| Length ceiling | None in practice | 20,000 characters |
| Requires configuration | No | Model credentials |
| Findings describe | Your source text | The rewritten text |
Availability: Thorough is offered only when a rewriting model is configured. When it is not, the option is disabled in the interface and the API rejects mode: "thorough"; Careful is unaffected.
03 / Character policy
| Code point | Character | Careful | Thorough |
|---|---|---|---|
U+00AD | Soft hyphen | Removed | Removed |
U+200B | Zero-width space | Removed | Removed |
U+2060 | Word joiner | Removed | Removed |
U+FEFF | Byte order mark | Removed | Removed |
U+200C | Zero-width non-joiner | Removed | Removed |
U+200D | Zero-width joiner | Removed | Removed |
U+FE00–FE0FU+E0100–E01EF | Unicode variation selectors | Removed | Removed |
Also applied: Unicode NFC normalization standardizes canonically equivalent sequences. The report lists that separately; it is not counted as a removed hidden character.
04 / Preservation
Two of the supported marks do structural work inside emoji, so they are tested positionally rather than removed on sight.
A zero-width joiner is preserved when the characters immediately before and after it are both emoji base characters. A variation selector in the range U+FE00–FE0F is preserved when the character immediately before it is an emoji base. Emoji bases are recognized in the ranges U+1F000–1FAFF and U+2600–27BF.
05 / Consequences
Multi-codepoint emoji such as family or profession sequences survive intact, and symbols keep their emoji presentation. A joiner sitting between two ordinary letters is removed, because it is doing no emoji work there.
Joined scripts are the known gap. A zero-width non-joiner shaping Persian or Arabic text is removed like any other stray mark, so the report labels those findings as sensitive. Check them before you accept the result.
06 / Report
| Field | Meaning |
|---|---|
| Code point | The U+ identity of the character, stable across fonts, editors and platforms. |
| Tag | The short label used in the marked preview, such as ZWSP or VS16. |
| Count | How many times that character occurred with that outcome. Removed and preserved occurrences of the same character are reported as separate findings. |
| Positions | One-based indexes into the scanned text. The list stops after the first eight occurrences; the count remains complete. |
| Action | Removed or preserved, decided by the preservation rules above. |
| Sensitive | Set for joiners and variation selectors — marks that can carry real meaning and deserve a manual look. |
| Normalization note | Shown when NFC changed the string. Deliberately excluded from the removal count. |
| Marks removed | The headline total. Preserved marks are not counted here. |
Marked preview: the reveal toggle re-renders the scanned text with each finding shown as a bracketed tag in place. Preserved marks appear in grey, and in Thorough mode the words the rewrite changed are highlighted. The preview is truncated for very long inputs; the findings list is not.
07 / Limits
| Limit | Value | Applies to |
|---|---|---|
| Free cleans | 10 | Per browser, no account needed |
| Account cleans | 100 per day | A verified account, reset daily |
| Thorough input | 20,000 characters | Each Thorough request |
| Rewrite timeout | 25 seconds | The model call in Thorough |
| Careful input | No practical ceiling | Local processing only |
| API requests | Metered per call | Billed separately from web allowances |
Splitting long text: when a document exceeds the Thorough ceiling, clean it with Careful in one pass and run Thorough section by section. Careful's output is identical however you split it, because the pass is purely local and per character.
08 / Data
Submitted text is held in memory for the duration of the request and is not written to a document store. Careful is computed entirely on this server.
Thorough sends the text to the model provider, with provider-side storage disabled on the request. That transfer happens only when you actively select Thorough. Full details, including the provider and retention position, are in the privacy notice.
09 / Interfaces
The web workbench keeps source, result and report side by side and supports pasting, dropping a plain-text file, a built-in sample, the reveal toggle, and ⌘/Ctrl + Enter to run.
The HTTP endpoint takes the same modes and returns the same findings as JSON, for pipelines and batch cleaning. See the API documentation for authentication, request format and error codes.
10 / Behavior notes
Zero supported hidden characters were found. It is a genuine clean bill for the supported set, not a failure. In Careful it says nothing about wording, which that mode never touches.
Appears when the source or mode changed after a run. The displayed result belongs to the previous input; run the remover again before copying.
If the rewrite failed or was unavailable, the response carries a warning and shows the Unicode-only clean of your original text. Nothing was rewritten.
Expected. Removed characters have no width, so the visible text looks the same while the character count drops.
Not always. Normalization can merge a base letter and a combining mark into one character, which shortens the string without any removal being counted.
Reporting is not the same as removing. Preserved findings exist so you can see every supported mark in your text, including the ones deliberately left alone.