Home - m4bwav/everwrite GitHub Wiki

everwrite

everwrite is an agent skill that makes an AI agent's prose read as if a person wrote it. The agent follows its rules while drafting docs, READMEs, reports, emails and release notes, then runs a small checker, tells.py, on the result. The checker is a standard-library Python script. It lists each machine-writing tell it finds and exits 1 while a strong one remains. It does not detect who wrote a text, and it does not touch code or config.

It runs in Claude Code as a plugin, and in GitHub Copilot, Codex, Cursor and other agents that read SKILL.md skills.

Install

In Claude Code:

/plugin marketplace add m4bwav/everwrite
/plugin install everwrite@everwrite

Other agents, the Claude directory and gh skill are on Getting started.

The smallest example

Take this draft.md:

# Release notes

We're excited to announce version 2.0, which marks a pivotal shift for the project.
It's not just faster, it's smarter. Let me know if you have questions!

From a clone of the repository, the checker reports:

$ python skills/everwrite/scripts/tells.py draft.md
draft.md:3: strong inflation: 'marks a pivotal shift' (state the fact)
draft.md:3: strong residue: 'excited to announce' (say it directly)
draft.md:3: weak vocabulary: 'pivotal' (plain word or cut; never swap for another listed word)
draft.md:4: strong contrast: "It's not just faster, it's" (state the second half directly)
draft.md:4: strong residue: 'Let me know if' (delete the chat wrapper)
tells: 4 strong, 1 weak, 28 words, 1 file(s)
$ echo $?
1

After a rewrite that states the facts:

# Release notes

Version 2.0 parses large files about twice as fast as 1.4.
Questions go in the issue tracker.

$ python skills/everwrite/scripts/tells.py draft.md
tells: 0 strong, 0 weak, 19 words, 1 file(s)
$ echo $?
0

With the skill installed you rarely type this yourself. The agent runs the checker on its own drafts and fixes what it reports.

Pages

  • Getting started: read first. Every install route, the Python it needs, and a first check.
  • How the skill works: what the agent does when the skill loads, the rules it writes by, and its edit and detect modes.
  • What the checker flags: every category, which hits are strong, and what the checker skips.
  • Commands: every option of tells.py, its exit codes and where each message goes.
  • Recipes: a pre-commit hook, a CI step, JSON output, stdin and wiki pages.
  • Versions and upgrading: what changed since 1.0.0 and how to update each install.
  • FAQ: watermarks, privacy, false positives and install errors.
  • Development: the self-test, the repository's own prose check, CI and the evergreen files.

Elsewhere

The current version is 1.0.0, first published on 2026-09-25. Later changes are on master under the same version number (see Versions and upgrading).