Repository navigation
FAQ
No. Claude and Gemini watermark their text through which words the model picks, and removing tells leaves that mark mostly in place. The skill says so and never promises undetectable text. What it does is make the text better to read and free of the patterns people notice. How the skill works has more.
No. The skill is instructions for the agent, and tells.py reads the files you name on your own machine with the Python standard library. It makes no network calls and keeps no copy. What your AI app does with the conversation is covered by that app's own privacy policy.
No. Detect mode lists the patterns it finds, each with the line, a quote and a fix. It gives no score and no verdict, because detectors and people both guess badly at authorship.
One listed word alone is a weak hit and does not fail the check. Two different listed words in the same paragraph are strong. If both words carry meaning, rewrite one of them in plain terms; if the text quotes them on purpose, wrap it in <!-- tells: off --> and <!-- tells: on --> lines (Recipes).
The skill's first hard rule bans em and en dashes in prose. Its research notes say dashes are fading as a general tell, since newer GPT models suppress them. A July 2026 study still found Claude using em dashes more than professional writers, and the skill mostly runs on Claude. Dashes in code, paths and URLs are skipped. If you write with dashes yourself, --allow-dashes turns the rule off (Commands).
No. Fenced and inline code are skipped, and the skill does not rewrite code or config. The one exception is invisible characters, which the checker looks for on every line, code included.
3.9 or later, with nothing else installed. The examples on this wiki printed the same output on Python 3.9.25 and 3.14.6. The repository's CI runs 3.9 and 3.13 on Ubuntu, macOS and Windows. Use python3 where python is missing.
On 2026-10-03, with Claude Code 2.1.281, claude plugin marketplace add m4bwav/everwrite cloned the repository over SSH. On a machine without GitHub's SSH host key it failed:
✘ Failed to add marketplace: Failed to clone marketplace repository: SSH host key is not in your known_hosts file.
Add the marketplace by its HTTPS URL instead, which needs no SSH key: /plugin marketplace add https://github.com/m4bwav/everwrite.git. That worked on the same machine (Getting started).
The repository holds a file in ai-docs/decisions/ whose name is 74 characters long. When Claude Code's config folder sits deep in the file system, the clone path passes Windows' 260-character limit and the checkout fails with unable to create file ... Filename too long. This happened on 2026-10-03 with a config folder 119 characters deep, and did not happen with the usual ~/.claude. Turning on long paths in Git fixed it: git config --global core.longpaths true, then add the marketplace again.
The checker writes UTF-8. Windows PowerShell decodes a native command's piped output with the console code page, so a dash in a match can show as three odd characters after | Select-String. The checker's output is correct. Read it directly, or use --json.
gh skill install adds a metadata block to the installed copy's frontmatter, naming the repository, path, ref and tree SHA. gh skill update uses it to find newer versions. The source file in the repository has no such block.
The version number has not been raised since the first release, though the checker gained new rules after it. Versions and upgrading lists what changed and how to update.
The skill follows the Evergreen Protocol: it re-researches the signs of AI writing on a schedule, logs each finding and change, and carries tests that prove it loads and runs the checker. The credits for the patterns are in the README.
This wiki describes everwrite 1.0.0 (master at 0673975) and was last updated on 2026-10-03. The skill is MIT licensed. Report problems in the issues. Listed in the Claude directory as Everwrite.
Using it
The releases
Contributing
Elsewhere