SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillstext-watermark-cleaner-zh-tw

text-watermark-cleaner-zh-tw

Writing & research

View repository

Updated 10-15 23:45

About this skill

Text Watermark Cleaning (Traditional Chinese)

Scope of Work

Only process content owned or authorized by the user. Describe the result as "cleaning verifiable text carriers" or "reducing statistical similarity"; do not claim that the content has been proven to be human-written, undetectable, or that Claude's proprietary algorithm has been cracked.

Text-based markers fall into two layers:

  • Layer A: Deterministic text hygiene processing. Check for and clean high-confidence invisible Unicode, zero-width, tag characters, anomalous whitespace, and some homoglyphs. This layer can be verified with before-and-after statistics.
  • Layer B: Statistical marker reduction. Rewrite by changing tokens, word order, connectives, sentence boundaries, and sentence length. This is not a decoder; without a vendor detector or the same key/config, it cannot be proven that the marker has been removed.

Do not treat "removing AI flavor" and "removing text watermarks" as the same action. Only execute this skill when the user explicitly requests watermark or provenance cleaning; when the request is merely for more natural text, use humanizer-zh-tw。

Mode Selection

  • inspect: Only inspect and report, without modifying the content.
  • layer-a: Only perform invisible text cleaning, in conservative mode by default.
  • layer-b: Only perform one statistical rewrite; explicit consent to possible semantic and tonal drift is required.
  • full：inspect → Layer A → Layer B（可選）→ Layer A → after inspect。

If the user requests both AI-flavor removal and watermark removal, adopt:

保護非 prose 區段 → inspect → Layer A → humanizer-zh-tw → 可選 Layer B → Layer A → after inspect

Execution Flow

1. First classify and protect the content

First confirm whether the input is pasted text,.txt、.md or .html. Before any rewriting, protect the following content unless the user explicitly requests modification:

  • Markdown fenced code、inline code、HTML tag、attribute、script、style。
  • URLs, file paths, API names, variable names, error codes, version numbers, and other technical identifiers.
  • Numbers, dates, proper nouns, quotations, references, and legally/academically/platform-required disclosures.
  • Sentences the user requests be preserved verbatim.

Do not treat DOCX, PDF, images, audio, video, or ZIP/Office containers as plain text for reading and writing. These files require separate container or metadata tools; otherwise the binary content may be corrupted.

2. Inspect first

This skill preferentially uses the included local deterministic scripts; they are extracted from guillaumemeyer/watermarks-remover of service/scripts and retain the upstream MIT license. Windows PowerShell execution method:

$skillRoot = (Resolve-Path .\text-watermark-cleaner-zh-tw).Path
& "$skillRoot\scripts\run-text-watermark.ps1" `
  -Mode Inspect -InputPath .\article.md -Json

Add -Aggressive or -StripEmojiGlueonly when stricter inspection is needed. To also obtain a statistical report, add -Stylometry; this is an analytical signal, not Claude's official detector.

If the deployment environment provides a compatible watermarks-remover HTTP service, you may also check first:

GET  $WATERMARKS_SERVICE_URL/health
GET  $WATERMARKS_SERVICE_URL/capabilities
POST $WATERMARKS_SERVICE_URL/inspect

When the service is unreachable, do not pretend that server-side cleaning has been completed; you may use local scripts to complete Layer A, and in the report separately indicate "local script verified" and "server-side unverified." Text in chat has no guaranteed post-send Unicode filter; only protected in-model processing can be performed, and the limitations must be indicated in the report.

3. Execute Layer A

Use the included scripts/run-text-watermark.ps1 -Mode Clean, the service's /clean or an equivalent deterministic cleaner, and preserve the original file, producing by default *.cleaned.*. Local entry example:

& "$skillRoot\scripts\run-text-watermark.ps1" `
  -Mode Clean -InputPath .\article.md -OutputPath .\article.cleaned.md -Stats

Adopt the conservative option:

  • Do not apply NFKC by default.
  • Do not convert all special whitespace to half-width spaces by default.
  • Do not use aggressive homoglyph mapping by default.
  • Do not delete legitimate RTL/LTR controls, emoji glue, CJK variation selectors, or other invisible characters that serve a typographic or semantic purpose by default.

The local wrapper adds by default --no-normalize-spaces; enable -Aggressive、-NormalizeSpaces、-Nfkc or -StripBidionly when the user explicitly accepts that layout, directionality, or multilingual glyphs may change. Do not use --force-text to process DOCX, PDF, or other binary containers.

4. Run Layer B (optional)

Run only when the user explicitly selects layer-b or full . If the source is suspected to be Claude, prefer a local model or a different provider; do not treat output from the same source model as independent verification.

Rewriting requirements:

  • Change vocabulary, function words, clause order, connectives, sentence boundaries, and sentence length.
  • Preserve all facts, numbers, names, quotations, technical identifiers, required disclosures, and the original author's genuine views.
  • Do not add first-person experiences, dialogue, sources, feedback, data, or background stories.
  • Do not translate, convert to Simplified Chinese, rewrite quotations, or alter code in order to increase variation.
  • By default, do only one round; if the user requests stronger rewriting, first warn about the risk of semantic and stylistic drift.

After rewriting, run Layer A again; the result of Layer B can only be labeled as best-effort. Only if there is a research detector with the same watermark key/config may the before/after of that specific setting be reported, and it must not be extrapolated into an official Claude detection result.

5. Deliver the report

Report three kinds of results separately:

  1. Verifiable: which Unicode characters were deleted/replaced, how many, and whether metadata still exists.
  2. Best-effort: whether the text underwent statistical rewriting, the number of rewriting rounds, and the backend used.
  3. Not yet established: official Claude detector, secret key watermark, visible pixel watermark, soft binding, or "human authorship".

If there is no actual inspect/after evidence, write "unverified"; do not substitute the text looking more natural for evidence.

Traditional Chinese rules

Traditional Chinese is the primary context of this skill; English technical terms are used only for triggering and the service interface. For detailed Chinese protection rules, see references/chinese-text.md。

  • Preserve Traditional Chinese characters, full-width punctuation, and Chinese quotation marks 「」 and 『』; do not translate or convert between Traditional and Simplified Chinese without authorization.
  • Do not misjudge the absence of spaces in Chinese as anomalous, and do not uniformly change U+3000, NBSP, or narrow no-break spaces.
  • Preserve emoji combinations, CJK variation selectors, legitimate Chinese typography, and citation formats.
  • Judge Chinese readability by local language intuition; do not determine watermarks solely by English token count, MATTR, or sentence-length thresholds.
  • "此外", "值得注意的是", "在……背景下", "這不僅……而且……", and the like fall within the scope of tone revision, not Layer A evidence; when needed, hand them to humanizer-zh-tw, and do not treat them as a detected watermark.

Failure handling

  • Save the original file, output file, and report separately; do not overwrite in place unless the user explicitly requests it.
  • If there is no difference before and after cleaning, report no actionable text marks found, and do not create new file changes.
  • When the format is unknown, the file is too large, the service is unreachable, or the detector is not configured, clearly report the reason and stop that layer; do not silently fall back.
  • Any necessary disclosures, author statements, citations, or compliance text must be retained.

Local resources and sources

  • scripts/common.py、scripts/text_unicode.py、scripts/clean_text.py、scripts/inspect_text.py and scripts/score_stylometry.py come from upstream watermarks-remover 's service/scripts。
  • scripts/run-text-watermark.ps1 is this repo's Windows PowerShell entry point, responsible for passing in UTF-8, conservative cleanup options, and the .cleaned output path.
  • A copy of the upstream license is at scripts/LICENSE-watermarks-remover.txt; the retrieved version is commit 4a0fbc312f2e5138c35d270d9db284cc07689930。

Typical requests

  • "Please only check whether this Traditional Chinese text contains invisible text markers; do not change the content." → inspect
  • "Please clean the zero-width characters from this Markdown, preserving code and quotations." → layer-a
  • "Please reduce the statistical markers of this Claude text; light rewriting is allowed, but do not add facts." → layer-b
  • "First remove the watermark, then help me remove the AI flavor." → full, and execute protection, Layer A, humanizer, optional Layer B, and Layer A in order.