Remove
Remove Claude Watermark From Code
Remove Claude watermark from code is a much easier problem than the same question about prose, for a structural reason rather than a tooling one.
Code gives the model very little room to make free word choices, and the free choices it does make sit almost entirely in comments and names.
Reformat, rename and rewrite comments. Ordinary refactoring removes most of what a code watermark could carry.
Runs in your browser. No signup, no upload, nothing stored.
Why code is a poor host
A statistical watermark needs choice. It works by nudging the model between roughly equivalent options, thousands of times over, until the aggregate pattern becomes measurable.
Source code removes most of that freedom. Identifiers must match their declarations, syntax is rigid, and the compiler rejects near-misses that would be perfectly acceptable variations in prose.
The entropy available per token is far lower than in natural language, so there is much less room to carry a signal in the first place.
Comments and names are the exception
Natural-language comments behave exactly like prose and carry whatever signal exists. Identifier names are the other place where genuinely free choice happens.
That gives you a short, concrete list of things to change, and all of them are things you might change anyway during review.
- Rewrite comments and docstrings in your own words
- Rename variables, functions and helpers
- Run the formatter and the linter with autofix enabled
- Apply an automated refactor pass
- Rewrite string literals and error messages
Your existing pipeline already does most of this
Formatters normalise whitespace and line breaks. Linters rewrite constructs. Refactoring tools rename symbols across an entire file in one action.
Code that has been through review, formatting and a rename pass has had most of its remaining discretionary choices overwritten as a side effect of ordinary work, without anyone setting out to do it.
Claude Code is in scope
Anthropic's coverage list explicitly includes Claude Code alongside the API, Claude, Claude Cowork and Claude Tag, applied worldwide with no opt-out.
So marking applies here. The point is that the medium resists carrying it, not that developers were granted an exemption.
The commit message is the softer target
Commit messages, pull request descriptions and changelog entries are prose, and they are frequently generated without much thought.
If you care about this at all, they carry more signal per line than the diff does — and unlike the diff, they are rarely rewritten by tooling afterwards.
One thing not worth doing
Deliberately mangling formatting or introducing odd naming to defeat a mark costs you readability and gains you very little.
Anything irregular enough to matter statistically is irregular enough to fail review. Run the standard tools instead; they do the same job as a by-product of making the code better.
Source
The claims on this page are drawn from primary documentation and reporting rather than from other tools’ marketing copy.
Anthropic: How Claude marks AI-generated content →Related
Check what your text and files actually carry
Runs in your browser. No signup, no upload, nothing stored.