Remove the Claude watermark

The mark says touched, not wrote

You wrote the proposal. Every line of it, over 2 evenings. Then you pasted it into Claude and asked it to catch the typos, because that is what you do at 11pm when it has to go out in the morning.

That proposal now carries the same invisible mark as something Claude wrote from scratch. The mark records that Claude touched the text. It says nothing about who did the thinking, and that is the gap worth understanding before anyone waves a detector at your work.

So this page is not about passing off AI writing as yours. It's about your own writing, lightly polished, getting flagged as though you had nothing to do with it.

5 things to understand first

Most of the advice going around about this is wrong, and it's wrong because people assume the mark is a tag hiding in the file. It isn't.

• It lives in the word choicesClaude nudges which words it picks using a secret key, and the signal is the pattern across hundreds of words. There is nothing sitting in your file to find and delete.
• Copy and paste carries itMove the text into an email, a doc, or a post and the mark travels with it. Light edits leave it in place.
• The Notepad trick does nothingPasting into a plain text editor and copying it back out strips formatting, and formatting was never where this lives. The mark survives it without a scratch.
• Short text is unreliable to detectA caption or a 3 sentence email is too short for the pattern to hold up. This is a long document problem, so spend your worry there.
• Only a heavy rewrite weakens itReal rewriting disrupts the pattern. Swapping a word here and there leaves it intact, which is why the manual fixes people suggest tend to fail.

Install the skill in 7 steps

A developer named Guillaume Meyer built this and gave it away for free. All the credit here is his. The steps below are the same ones he documents, written out for someone who has never installed a skill before, with the 2 places people get stuck marked clearly.

1. Open the project on GitHubGo to github.com/guillaumemeyer/watermarks-remover. It's free and open, and you don't need an account to look around.
2. Click into the skills folderIn the file list, click skills, then remove-ai-marks. That folder holds the instructions.
3. Open SKILL.mdClick the file named SKILL.md. That one file is the whole skill, written in plain English. You can read it if you're curious about what it does.
4. Choose Download raw fileTop right of the file view. Save it somewhere you can find again, like your Downloads folder.
5. Upload it into ClaudeIn Claude, go to Customize, then Skills, then Add, then Upload skill. Drag the file in and you're done with the install.
6. Turn on code executionGo to Settings, then Capabilities, and switch on code execution. Skip this and the skill refuses to run, and the refusal reads like the skill is broken. Almost everyone who says it doesn't work stopped right here.
7. Check that it tookStart a new chat and type a forward slash. If remove-ai-marks shows up in the list, you're set.

Run it on your own writing

Start a chat, run the command, then paste the text you want cleaned. It works on anything you can paste in, and it's fastest on one document at a time.

/remove-ai-marks

Make the skill yours

Out of the box this skill cleans text without knowing anything about how you write. That's fine for a first run, and it's a waste after that. 15 minutes here turns it into something that cleans your work the way you would want it cleaned.

You don't need to be technical for any of it. A skill is a set of instructions saved in a file called SKILL.md, and you change those instructions by talking to Claude in plain English. Claude does the typing.

1. Copy the skill folder somewhere safeSo an update can never erase your work. This is the one step you can't fix after the fact, so do it first.
2. Tell Claude how you writeYour rules, your formatting, the words that are yours. The more true detail you give it, the less it sands your voice off while it cleans.
3. Ask Claude to rewrite the skillPaste the prompt below. Claude edits the file for you. Your only job is to read it over and say what to change.
4. Leave the very top aloneThe few lines at the top of the file, between the dashes, tell Claude when to use the skill. Leave those exactly as they are.

This prompt is written for this specific skill. Paste it into Claude and fill in the brackets.

I want to customize the remove-ai-marks skill so it cleans my writing the way I would clean it myself.

The documents I run through it are [for example: client proposals, my weekly newsletter, LinkedIn posts].
My writing rules are [for example: no em dashes, sentences under 15 words, plain words, no jargon].
Words and phrases that are mine, keep them: [list the words that sound like you].
Words I never use: [list anything that would not sound like you].
Formatting that has to survive untouched: [for example: headings, bullet lists, my sign off].

Please rewrite the skill's instructions so it follows all of this while it cleans, and show me exactly what you changed. Keep the section at the very top between the dashes exactly as it is.

What to tell it, and why each one matters

• Which documents you runA proposal and a caption want different handling. Name yours and the skill keeps the right register while it works.
• Your writing rulesThe things you would never write. No em dashes, short sentences, no jargon. A clean should move your text toward your rules, and without this it drifts wherever it likes.
• The words that are yoursEvery writer has a handful. Name them and the skill stops trading them for something more generic on its way through.
• Formatting to protectHeadings, bullets, your sign off. Say them out loud in the prompt and your layout comes back the way you built it.

What can't be verified yet

Anthropic has not released the tool that reads the mark. There is no checker to run your cleaned text through, and no green light to see at the end.

So one half of this is verifiable and one half is not. Stripping invisible characters is something you can confirm yourself. Undoing the word choice pattern is a claim, and the people making it can't prove it either, because the reader does not exist in public yet.

Treat "completely removed" as a claim you can't check. That's the honest position until Anthropic ships the detector, and anyone telling you otherwise right now is guessing with confidence.