What does this skill do?

The Watermarks Remover Skill removes AI-generated watermarks from your own content in text and files. It operates on three levels: deterministic cleaning of invisible Unicode and bidirectional characters, rewriting of statistical watermarks (token-sampling) and removal of C2PA/EXIF/XMP metadata in over 20 formats, including PNG, JPEG, PDF, DOCX, SVG, HTML, and Markdown.

Invisible Unicode Hygiene
Detects and removes invisible Unicode characters, exotic spaces, bidirectional marks, and tag characters from any text.
Cleaning Up Metadata in Images
Removes C2PA, EXIF, and XMP metadata from PNG, JPEG, WebP, AVIF, HEIC, BMP, GIF, TIFF, and SVG files.
Rewriting Statistical Watermarks
Rewrite text marked with statistical watermarks (Kirchenbauer, Gumbel/EXP) using an optional rewrite hook.
Batch Processing via HTTP
Clean or inspect multiple files in a single HTTP request using batch endpoints.

How to install

$ Quick install (one-liner) 
git clone --depth 1 https://github.com/guillaumemeyer/watermarks-remover.git ~/.claude/skills/watermarks-remover

Clone or download the repository into ~/.claude/skills/ and restart Claude. That's all. Full installation guide →

Usage examples

🔤 Clear text
Remove the AI source marks from this Markdown file and give me the clean result.
🖼️ Clean up images
Remove the C2PA and EXIF metadata from all PNG images in this folder.
📦 Batch cleaning
Scan and clean all DOCX and PDF files in this directory in batches.
🔍 Detect watermarks
Check whether this text contains statistical AI watermarks before cleaning it.

Features

More than 20 supported formats It processes PNG, JPEG, WebP, AVIF, HEIC, BMP, GIF, TIFF, SVG, PDF, DOCX, XLSX, PPTX, EPUB, ODT, HTML, Markdown, MP4, MOV, WAV, and MP3.
Dependency-free HTTP service The core runs exclusively on Python 3.10+ stdlib, with no Docker or required external packages.
Pre- and post-detection Measure the watermarks before and after cleaning using text detectors (MarkLLM, Gumbel) and the SynthID score for images.
Batch processing Clean or inspect up to 50 files per request using the HTTP service's batch endpoints.
Multi-vendor It supports watermarks from Claude, Gemini/SynthID-Text, OpenAI, and open LLMs (Kirchenbauer, Gumbel/EXP).

Frequently asked questions

You need to clone the repository and start the HTTP service using Python 3.10+. The skill is a Markdown client that calls the service via HTTP, without requiring Python on the agent host.
It supports more than 20 file formats: PNG, JPEG, WebP, AVIF, HEIC, BMP, GIF, TIFF, SVG, PDF, DOCX, XLSX, PPTX, EPUB, ODT, HTML, Markdown, MP4, MOV, WAV, and MP3.
Yes. The core service uses only the Python 3.10+ standard library, with no dependencies. Docker is optional and is only required for resource-intensive backends such as SynthID or MarkLLM.
This skill is designed to ensure the integrity and privacy of users' own content. Users are responsible for complying with the applicable regulations in their jurisdiction regarding the removal of source markings.
Watermarks Remover — Removes AI-generated watermarks

¿Prefieres escuchar el contenido? Genera la narración de audio con un clic.