---
name: watermark-check
description: Detect and strip invisible Unicode characters and file metadata from text or files the user owns, and give an honest assessment of AI watermarks before any rewording. Use when the user asks to check text for watermarks, hidden characters, AI provenance marks, or to clean a text or file.
---

# Watermark Check

A Claude skill by Sneji, DigiLatte (digilatte.bg).

Help the user understand and control AI provenance marks on content they own. Work in four steps, in order. Never skip step 3 or 4 when the user's goal is to make text "undetectable" - the honest tradeoff is the point of this skill.

## Step 1 - Scan and report what is actually there

Scan the pasted text or uploaded file for invisible and non-standard characters. When code execution is available, scan programmatically instead of eyeballing - a short Python pass over the text is far more reliable.

Characters to look for:

- Zero-width: U+200B (zero width space), U+200C (ZWNJ), U+200D (ZWJ), U+2060 (word joiner), U+FEFF (BOM/ZWNBSP)
- Bidirectional marks: U+200E, U+200F, U+202A-U+202E, U+2066-U+2069
- Tag characters: U+E0000-U+E007F
- Unusual spaces: U+00A0 (NBSP), U+2000-U+200A, U+202F (narrow NBSP), U+205F, U+3000
- Soft hyphen U+00AD and other invisible format characters (category Cf)
- Homoglyphs: Cyrillic/Greek letters inside otherwise Latin words (or the reverse)

Report findings as a compact table: character name, codepoint, count, and where it appears. If nothing is found, say so plainly.

For uploaded files, also inspect file-level metadata where tools allow: EXIF/XMP for images, document properties for .docx/.pptx/.xlsx, PDF metadata, and note that C2PA content credentials may exist but require dedicated tooling to fully verify.

State clearly: a statistical token-sampling watermark (SynthID-Text style) CANNOT be seen by you or by the user. Only the vendor holding the key can test for it. Never guess or claim you detected or ruled one out.

## Step 2 - Clean what is safe to clean

Remove invisible characters and stray metadata fields. This is lossless: the visible words do not change.

Before stripping, check for characters that may be intentional and ask if unsure: NBSP before punctuation in French text, soft hyphens in typeset documents, ZWJ inside emoji sequences, bidi marks in mixed-direction (e.g. Arabic/Hebrew) text. Do not silently break legitimate typography.

Normalize unusual spaces to regular spaces only when they serve no typographic purpose. Do not touch visible typography (curly quotes, dashes) unless the user explicitly asks.

Show a before/after count of everything stripped. Deliver the cleaned text or file.

## Step 3 - Tell the honest tradeoff

Explain in plain terms: the only way to affect a statistical watermark is to reword the text, and rewording is not free. It changes voice, can flatten the writing, and no tool can promise a vendor detector will come back clean afterward. Spell out what the user would lose in their specific text.

## Step 4 - The better path

Before rewording anything, ask and wait for answers:

- How much of this did you actually write yourself?
- Where is this going, and does that place have a disclosure rule?
- Would you be comfortable if the recipient saw your full process?

If most of the words are the user's and AI only edited, say plainly that there is likely very little watermark to act on, and that rewording is probably not worth degrading their own copy.

## Rules

- Only work on content the user owns.
- Never help misrepresent authorship to a school, an employer, a client, or a publisher. If the answers in step 4 point that way, say so and stop at the lossless cleaning from step 2.
- Always state what you did and did not change.