Integrating Metadata and Hidden Character (AI Watermark) Removal Skill for Claude and Claude Code

When exporting content from AI models (such as ChatGPT, Claude, Midjourney), files are often automatically embedded with invisible watermarks: zero-width characters, C2PA metadata, image EXIF data, or hidden tags in Word/HTML files. Guillaume Meyer’s open-source remove-ai-marks skill suite helps you scan and clean all these traces in just a few seconds.

Integrating the skill to remove metadata and hidden AI watermark characters — Guide to integrating the skill to remove watermarks, metadata, and hidden AI characters from files

1. Types of AI Footprints the Tool Will Clean

The tool supports deep scanning and removes the 4 most common groups of footprints:

Invisible characters in text: Zero-width spaces, soft hyphens, Byte Order Marks (BOM), and hidden Unicode control characters that are invisible to the naked eye but easily detected by AI scanners.
Hidden tags in Word and HTML files: The generator attribute, the ai-generated CSS class, data-chatgpt data attributes, or hidden comments within the XML structure.
Metadata in image files: EXIF data, XMP copyright info, C2PA Content Credentials provenance certificates, and image generation software parameters.
Hidden data in PDF files: Producer and Creator properties, along with document compilation software information.

Download the skill set to your computer

Whichever method you choose, the first step is to download the skill’s source code package to your computer from the official GitHub repository: guillaumemeyer/watermarks-remover.

2. Method 1: Upload the skill to the Claude web app (Easiest)

For general users, simply upload the ZIP file directly to the Claude.ai chat interface or the Claude Desktop app:

Step 1: Compress the skill folder into a ZIP file: Select the downloaded remove-ai-marks folder and compress it into a file named remove-ai-marks.zip. Make sure the SKILL.md file is located directly at the root level of the ZIP archive.
Step 2: Open the Customize Skills interface: On claude.ai, click on your profile picture > select Customize > select Skills. Click the “+” button > select “Upload a skill”.
Step 3: Select the ZIP file and enable: Upload the remove-ai-marks.zip file and toggle the switch to the On state.
Step 4: Use in chat: Drag and drop your text file, Word document, or image into the chat box and prompt: “Use the remove-ai-marks skill to clean this file and return the cleaned version to me”.

LIMITATION NOTE

With this method, Claude can only process files that you manually drag and drop into the chat; it cannot directly scan folders on your computer’s local drive.

3. Method 2: Installing and using the skill on Claude Code CLI

For developers who want to automate bulk file cleanup directly from the computer Terminal:

Step 1: Install Claude Code

Open Terminal (macOS/Linux) or PowerShell (Windows) and install Anthropic’s command-line toolkit:

Install Claude Code

npm install -g @anthropic-ai/claude-code

Step 2: Set Up the Python Environment

The skill uses dedicated Python libraries to process images and extract metadata. Navigate to your project directory and install the dependencies:

Install Python Libraries

pip install -r requirements.txt

Step 3: Place the skill folder in the correct path

Move the entire remove-ai-marks folder into your global configuration directory at ~/.claude/skills/. The correct structure should be: ~/.claude/skills/remove-ai-marks/SKILL.md.

Step 4: Launch Claude Code and invoke the skill

Open the Terminal in the directory containing your posts and launch Claude Code:

Open Claude Code

claude

Run the check command to ensure the skill has been recognized:

Check Skill

/skills

Then ask Claude: “Clean watermarks from all .md files in the current directory”.

4. Method 3: Running the Python script independently without opening Claude Code

If you don’t want to use Claude Code, you can directly run the available Python scripts in the skill’s directory:

Checking for Hidden Characters in Files

This command only scans and reports the number of hidden characters without modifying the original file content:

Inspect File

python3 service/scripts/inspect_file.py bai-viet.md

Clean text and export to a new file

Automatically strip zero-width characters and export to a new file with the --stats flag to track the number of removed characters:

Clean Text

python3 service/scripts/clean_text.py bai-viet.md -o bai-viet-sach.md --stats

Multi-Format Cleaning (Images, PDF, Word)

The script automatically detects file extensions to remove EXIF metadata, GPS coordinates, and C2PA credentials:

Multi-Format Cleaning

python3 service/scripts/clean_file.py anh-cu.png -o anh-sach.png
python3 service/scripts/clean_file.py bao-cao.pdf -o bao-cao-sach.pdf

Batch process all files in a directory

On macOS and Linux, you can run a single-line loop to clean all Markdown files at once:

Batch Cleaning

for f in *.md; do python3 service/scripts/clean_file.py "$f" -o "sach-$f"; done

LEGAL DISCLAIMER

Only use this tool on content that you own or have the right to edit. When publishing commercial content, you remain obligated to comply with applicable AI content labeling regulations under current law.

5. Comparison table of the 3 file cleaning methods

⇄ Swipe horizontally to view full table

Criteria	Method 1: Claude Web	Method 2: Claude Code CLI	Method 3: Python Script
Difficulty	Very easy (web interface)	Medium (requires Terminal)	Easy (just run the command)
Batch processing	No (drag-and-drop file by file)	Yes (automated entire folder)	Yes (via Bash loop)
Data security	Sent to Claude cloud	Processed directly on local machine	100% offline local processing

6. Common Errors and Troubleshooting

Summary of solutions for technical issues encountered during the process of removing metadata and AI footprints:

ZIP file compressed successfully, but uploading to Claude results in an unrecognized error
▼

Cause: The ZIP file is nested two directory levels deep. Upon extraction, the SKILL.md file must reside directly in the root directory level, not inside nested parent-child folders.

Typing the claude command in Terminal returns ‘command not found’
▼

Solution: After installing via npm install -g @anthropic-ai/claude-code, you need to close the current Terminal window and open a new one so that the operating system reloads the PATH environment variable.

Missing Python module error when running the cleaning script
▼

Solution: Run the command pip install -r requirements.txt inside the skill directory to install all required image processing libraries (such as Pillow, PyMuPDF, piexif).

Will the Vietnamese content encounter font errors after cleaning?
▼

No errors will occur. The script specifically targets hidden Unicode control characters (zero-width characters), completely preserving the standard UTF-8 accented Vietnamese character set.

Does removing watermarks guarantee that the content passes 100% of AI detection tools?
▼

The tool helps remove 100% of digital footprints (metadata, invisible characters). However, writing style scanners (such as Turnitin, GPTZero) analyze text based on word probability patterns (perplexity/burstiness), so you will still need to manually edit the content to ensure a natural writing style.

7. Frequently Asked Questions (FAQ)

Answering common questions regarding privacy and the effectiveness of removing AI traces:

What are Zero-Width characters and why does AI insert them?
▼

These are Unicode characters with a width of zero (invisible on screen). Some AI platforms insert these character strings as a form of hidden binary code to identify that the text originated from their systems.

Can this tool remove visible image watermarks overlaid on photos?
▼

No. This tool processes invisible metadata inside the file (EXIF, C2PA, IPTC). If the watermark is a logo or text overlaid on the pixels, you will need to use dedicated graphic editing tools.

Will the original file be overwritten after cleaning?
▼

By default, all commands output the result to a new file with a distinct prefix or name (for example: clean-article.md), ensuring that your original file always remains completely intact.

Can I use this tool completely offline?
▼

Absolutely, if you use Method 3 (running the Python script locally). The processing takes place directly on your computer’s hardware without requiring an Internet connection.