THE IDEA
Here’s what you need to know.
Anthropic says supported Claude models can place an invisible statistical watermark inside their word choices. It is not a hidden character you can reveal or delete. Copying normally preserves it, light editing may leave enough of the pattern behind, and a complete rewrite can disrupt the original signal. But there is a catch: if a supported Claude model performs that rewrite, the new response can receive a fresh Claude watermark. Use this guide to understand the technology, edit responsibly, and inspect any community skill before installing it.
See the visual breakdown
- 1
Claude selects low-stakes words
- 2
Those choices form a statistical pattern
- 3
Copying preserves the words
- 4
A full rewrite can disrupt the old pattern
- 5
A supported Claude rewrite can add a new one
Model
Check whether the Claude model and surface support watermarking.
Method
Word-choice patterns are not invisible characters or removable metadata.
Tool
Inspect the complete third-party package, service, and security audits.
Disclosure
Use editing honestly and follow the rules that apply to the work.
Official references and the community tool
- Open
Anthropic's text-watermark announcement
The official explanation of how Claude's word-choice watermark works, what it can survive, and its limitations.
- Open
Current Claude model coverage
Anthropic's maintained list of supported models and surfaces, plus guidance on text watermarks and Content Credentials.
- Open
Anthropic's custom Skills guide
The official upload workflow. Claude chat requires a correctly packaged ZIP containing the complete skill folder.
- Open
Remove AI Marks on Skills.sh — third party
The community skill mentioned in the video. Its current audit results are mixed, so review the warnings and every referenced file before installing it.
- Open
Remove AI Marks source repository — third party
Inspect the full repository, dependencies, scripts, service calls, permissions, and version history rather than downloading only SKILL.md.
PUT IT INTO PRACTICE
Follow the steps.
Tick off each step as you go. Your progress is saved on this browser.
- STEP 01
Open Anthropic's maintained model-coverage page before making assumptions. Watermark support varies by model, date, first-party surface, and cloud provider. Anthropic says the mark is applied at the model level, so it can appear across Claude, the API, Claude Code, and Cowork when a supported model is used.
- STEP 02
The watermark is a statistical pattern created through low-stakes word choices. It is not an invisible Unicode character, metadata tag, or secret line you can delete. Copying and pasting usually carries the word pattern with it, while light edits may still leave enough evidence for detection.
- STEP 03
Anthropic says a complete rewrite that replaces every word can remove the original pattern. That does not prove human authorship, and there is no public detector or verification key available for you to check the result yourself. Heavy editing, translation, or mixing text can also reduce detectability, so the absence of a detected mark is not proof that Claude was never involved.
- STEP 04
A supported Claude model can add a fresh watermark to the rewritten output. A community prompt or skill may disrupt the pattern from the original passage, but running the rewrite through Claude is not a reliable path to unwatermarked text. Treat claims of guaranteed or undetectable removal with caution.
- STEP 05
The current Remove AI Marks listing is not an Anthropic product and shows mixed security-audit results. Its SKILL.md references additional files and a separate service, so downloading only the raw SKILL.md is incomplete. Read the whole repository and audits, confirm what leaves your device, and do not submit confidential text to a service you have not vetted.
- STEP 06
For Claude Code and other agents supported by the Skills CLI, Skills.sh currently provides the command below. For Claude chat or Cowork, Anthropic documents uploading the complete reviewed skill folder as a ZIP through Customize > Skills—not a lone raw file. Installation does not make the output guaranteed undetectable.
npx skills add https://github.com/guillaumemeyer/watermarks-remover --skill remove-ai-marks
- STEP 07
Use rewriting to make the text clearer, more accurate, and genuinely aligned with your voice. Keep records of your sources and edits, and follow any AI-disclosure rules that apply to school, work, publishing, clients, competitions, or regulated communications. If you run this prompt through a supported Claude model, remember that the new output may carry a new Claude watermark.
Rewrite the following text from its underlying facts and outline, rather than swapping a few words sentence by sentence. Requirements: - preserve every verified fact, number, qualification, and source - remove repetition, filler, clichés, and robotic transitions - use clear, natural language that matches this voice: [DESCRIBE YOUR VOICE] - vary sentence length only where it improves readability - do not invent facts, quotes, sources, or personal experiences - flag any claim that needs verification instead of strengthening it - disclose AI assistance wherever my school, workplace, publisher, client, or platform requires it Text: [PASTE TEXT HERE]
A FEW THINGS THAT HELP
Get more out of it.
- Claude's text watermark lives in word-choice patterns; it is different from C2PA Content Credentials attached to supported generated files.
- Anthropic's watermark detector is in private preview, not a public tool anyone can use to certify a passage.
- A missing or weakened signal does not prove that a person wrote the text without AI assistance.
- Third-party skills can reference scripts, services, network requests, and other files. Audit the complete package, not just its main instructions.
- Do not strip provenance to misrepresent authorship or bypass school, workplace, publisher, client, platform, or legal disclosure requirements.