Claude zero-width space remover, codepoints only

Exactly one class of edit. Fourteen codepoints are matched and handled: eleven removed, three replaced with U+0020. No other byte in the input is altered. Whitespace handling is switched off on this route, punctuation normalisation is off, blank-line collapsing is off, and the output is not trimmed at either end.

That narrowness is the specification, not an omission. A transformation that touches one known set is auditable: run the result through a scanner and the count should be zero, and diff the input against the output and every change should be accounted for. A cleaner that also reflows whitespace produces a diff nobody can review.

Comparable tools delete every codepoint on their list, iterating the string and dropping anything matched. Applied to no-break space that is wrong. U+00A0 occupies the semantic role of a word separator, so removing it concatenates the two words around it and produces a corruption that reads as a typo rather than as an encoding artefact.

Three of the fourteen therefore carry a substitution rather than a deletion: no-break space U+00A0, narrow no-break space U+202F and hair space U+200A, each replaced with an ordinary space. The other eleven have no visual role, so deletion is correct for them. This distinction is not cosmetic. It is the difference between output you can publish and output you have to proofread.

Operate on the received artefact. Passing text through an editor or a formatter first frequently strips these codepoints incidentally, which destroys the evidence and the measurement at the same time.

Step 1

Paste the unmodified text. No length cap is imposed. A sample loader is available for verifying behaviour against known contaminated input.

Step 2

Execute. The action is labelled Remove hidden characters and takes no parameters, since this operation exposes no options.

Step 3

Copy the output. The input is retained in place, so the two can be diffed directly.

Fixed in this codebase, not configurable, and deliberately narrower than the widest lists published elsewhere.

Zero-width group, four

U+200B, U+200C, U+200D, U+2060. Each consumes a code unit while rendering nothing, so byte-level equality diverges from visual equality.

Byte-order mark

U+FEFF. Meaningful only at stream head. Mid-document instances surface as parser faults, feed validation errors and stray glyphs in some editors.

Soft hyphen

U+00AD. Conditional, inert until wrapping, at which point a hyphen appears inside a word that never contained one.

Bidirectional controls, five

U+202A through U+202E. The two overrides decouple display order from storage order, which is the mechanism behind filenames that read as something other than what they are.

Substituted three

U+00A0, U+202F, U+200A. Replaced with U+0020, since each occupies the visual position of a space.

Outside the set

Variation selectors, U+200E and U+200F, the isolate controls, the invisible math operators and U+180E are not handled here. Tools listing them are describing a wider operation than this one.

Because whitespace carries meaning in a large fraction of the text people paste here. Indentation is syntax in Python and YAML. Trailing spaces are a hard line break in Markdown. Blank-line counts separate blocks in half a dozen formats. A cleaner that normalises spacing is making a semantic edit while presenting itself as hygiene.

The sibling route that also handles whitespace exists, and it is the right choice for prose. This one is the right choice when the artefact is structured and the only defect is a codepoint that should not be there.

U+200B was defined to indicate word boundaries without rendering a visible space, and to mark a permitted line break without a hyphen. It is used for exactly that in scripts that do not separate words visibly, including Thai, Myanmar, Khmer and Japanese. In HTML the wbr element serves the same function. Removing it from such text is data loss, not cleanup.

U+200C is likewise meaningful in Persian and in several Indic scripts, where it governs letter joining. The security case runs the other way: ICANN prohibits zero-width spaces in domain names because they support homograph attacks, in which a malicious URL is visually indistinguishable from a legitimate one. The right posture is to know which case you are in before running anything.

It does not remove Claude's text watermark. Anthropic's published account states that nothing is added to the text and that there are no hidden characters, which places the mark categorically outside any codepoint operation. This is text hygiene, and it is not a method of concealing authorship or avoiding a disclosure requirement.

It produces no report, no per-character explanation and no analysis, since it is a transformer rather than a scanner. It offers no options, because this operation exposes none. It performs no punctuation or Markdown handling, accepts no files, exposes no API and makes no prediction about any detector's output.

The figure reported after a run is a length delta: input length minus output length. It is not a count of codepoints handled, and the difference is not a rounding issue. A substituted no-break space contributes zero to it, because one character was exchanged for one character.

So a run that reports zero may still have made three substitutions. If you need the count rather than the delta, scan the input before running this, and scan the output afterwards to confirm the pass completed.