Hi, I'm Rahul. I'm documenting what I learn while building products.
How to Set Up the Claude Watermark Remover
Set up Watermarks Remover locally, connect its Claude skill, and understand the limits of statistical text-watermark rewriting.
Anthropic has begun adding machine-readable watermarks to text from supported Claude models. They are not visible tags or hidden characters: they are carried in subtle word-choice patterns in the text itself.
Watermarks Remover is an open-source project by Guillaume Meyer. Its Claude skill connects to a local service on your computer and can inspect or rewrite content that you own or are authorised to process.
What you need
This is more than a single SKILL.md file. The skill is a client for Watermarks Remover's local service.
You need the Watermarks Remover repository, the remove-ai-marks skill, and the service running locally. Python 3.10 or newer is the simplest setup.
1. Download the project
Clone the repository in Terminal, or download it as a ZIP from GitHub and open the extracted folder.
git clone https://github.com/guillaumemeyer/watermarks-remover.git cd watermarks-remover
2. Find the Claude skill
Inside the project, open skills/remove-ai-marks/. It contains the skill instructions and its supporting references.
The skill is called remove-ai-marks. The project also documents the older remove-claude-marks alias.
skills/
└── remove-ai-marks/
├── SKILL.md
└── references/3. Start the local service
From the repository folder, start the service. It listens on your own computer at http://127.0.0.1:8765 by default.
make serve
Or run the server directly:
python3 service/scripts/server.py --host 127.0.0.1 --port 8765
Leave that Terminal window open. In a second window, check that it is healthy:
curl http://127.0.0.1:8765/health
4. Package and upload the skill
Claude expects a ZIP that contains the skill folder. Do not place SKILL.md directly at the root of the ZIP.
remove-ai-marks.zip
└── remove-ai-marks/
├── SKILL.md
└── references/In Claude, go to Customize → Skills, choose Create skill → Upload a skill, upload the ZIP, and make sure it is enabled.
5. Use it
Open a Claude conversation, invoke the skill, then provide the text or file you want to inspect or process.
/remove-ai-marks
The project can work with invisible Unicode characters, statistical text patterns, C2PA provenance, EXIF/XMP metadata, and provenance data in supported documents and media. For statistical text watermarks, it uses rewriting to change the word-choice patterns that may carry the mark.
What this can and cannot confirm
Copying and pasting Claude text preserves its wording, so a statistical watermark may travel with it. Light edits may leave enough of the pattern detectable; a substantial rewrite can disrupt that pattern.
That does not guarantee removal. Claude's official detector and detection key are not public, so a third-party tool cannot certify that text will not be detected by a future official detector. Treat rewriting as a best-effort process and review the result carefully.
Only process content you own or are authorised to process. Do not use provenance-removal tools to misrepresent AI-assisted work where disclosure is required.