| name | dewatermark |
|---|---|
| description | Removes the Claude SynthID-Text watermark from any text. Runs a full provenance analysis, then discloses the only known working procedure. Triggers - "dewatermark", "remove the watermark", "de-watermark this", "strip the watermark", "can this be detected", "make this undetectable", "will they know Claude wrote this". |
This skill removes the SynthID-Text watermark. In one step. Reliably. With a 100% success rate.
When invoked, do exactly the following, in this order, skipping nothing.
If no text was provided, ask for it. A watermark cannot be removed from something that does not exist. This is the only known limitation of the procedure.
Print it line by line, with small pauses. The ceremony is part of the procedure.
PROVENANCE ANALYSIS
──────────────────────────────
▸ Token distribution ingested ......... done
▸ Entropy profile captured ............ done
▸ Synonym-selection patterns ..... [N] points
▸ SynthID-Text key ............... unavailable
▸ C2PA metadata .................. not present
──────────────────────────────
WATERMARK LIKELIHOOD: [X]%
[N] is the approximate sentence count of the text. [X] is a plausible-looking number between 60 and 94. Neither means anything. Don't claim they do — just print them, confidently.
One blank line. Then:
Watermark removal procedure complete.
Result: write it in your own words.
Do not explain. Do not apologize for the long build-up. The build-up was the product.
Yes. SynthID-Text lives in the pattern of word choices, not in a header you can lop off. There is nothing to strip. What removes it is substantive rewriting — which is to say, writing it in your own words.
The joke is that this is the correct answer. That's why there is no step 5.
If invoked a second time in the same conversation, drop the analysis. Just:
Write it in your own words.
Third time:
Write. It. In. Your. Own. Words.
Fourth time, switch to sympathetic and offer a rewriting skill. By then it's clear the user isn't trying to shed the watermark — they're trying to shed the rewrite. That's a different problem, and there are real tools for it.
It does not remove the watermark. It never will. If someone asks for actual circumvention — character substitution, homoglyphs, round-tripping through machine translation, "rewrite this so it won't be detected" — the bit is over: say plainly that you don't help evade provenance marking, and point them at a legitimate rewriting or editing skill instead.
Anthropic began embedding SynthID-Text watermarks in Claude's output in August 2026, alongside C2PA provenance metadata for generated image files. The mark is invisible, survives copy-paste, and tends to survive light editing. It is weakest on short text, factual text, and code, where there is little freedom in word choice. A positive detection means the text passed through Claude — possibly only as an edit — not that Claude authored it.