AI prose often keeps the same structure even after the obvious buzzwords are removed. Humanizer edits the structure first, then the language.
Install with Codex
$skill-installer Install the humanizer skill from
https://github.com/andreaskonopka/humanizer/tree/main/skills/humanizer
Try it
$humanizer Revise this text without changing its facts or level of certainty:
<paste text>
Release v0.1.1 · MIT license · Security · Submit a difficult text
Humanizer is an Agent Skill for editing German and English prose. It targets repeated structures, inflated claims, vague attribution, canned transitions, and language-specific typography while preserving the source's meaning and factual uncertainty.
It is an editorial tool, not an authorship classifier. A phrase is not removed just because it appears on a watchlist; frequency, context, and document type matter.
Humanizer reviews structure before vocabulary. It looks for patterns such as repeated rules of three, punchy punctuation pivots, unsupported interpretive tails, generic concession-and-reassurance arcs, and tidy conclusions that add nothing.
The skill then applies a language-specific catalog. German and English have different punctuation, quotation-mark, heading, and register conventions, so they are not processed as translations of one another.
The final pass checks that the revision has not introduced facts, numbers, sources, quotations, scenes, or certainty that were absent from the input.
Input:
In today's dynamic landscape, the new process marks a pivotal moment, highlighting our commitment to efficiency and setting the stage for future success.
Possible revision:
The new process shortens the approval path.
That revision is valid only if the source or supplied context establishes the shorter approval path. Without that support, the safer edit is more modest:
We introduced a new approval process.
Humanizer selects and clarifies available information. It does not manufacture the concrete detail that good prose needs.
Copy skills/humanizer into a skills directory supported
by your agent. Codex and Claude Code both recognize the cross-platform
~/.agents/skills/ location:
git clone https://github.com/andreaskonopka/humanizer.git
mkdir -p ~/.agents/skills
cp -R humanizer/skills/humanizer ~/.agents/skills/humanizerIf the skill does not appear immediately, start a new agent session.
Codex users can also ask the built-in installer:
$skill-installer Install the humanizer skill from
https://github.com/andreaskonopka/humanizer/tree/main/skills/humanizer
Or use the cross-platform skills CLI:
npx skills add andreaskonopka/humanizer --skill humanizerThe repository includes a valid skills-only Codex plugin manifest. Clone the repository, then ask Codex to add that folder to your personal marketplace:
$plugin-creator Add the plugin at /absolute/path/to/humanizer to my personal
marketplace and validate it.
Install the registered plugin, then restart the ChatGPT desktop app:
codex plugin add humanizer@personalYou can also open Plugins, choose the Personal source, and install Humanizer there.
The manifest prepares the project for Codex plugin installation and future submission. Publishing this repository on GitHub does not by itself add it to the universal public plugin directory.
Invoke the skill directly:
$humanizer Revise this support reply without changing its facts.
It can also trigger from requests such as:
Make this paragraph sound natural and less formulaic.
The skill intentionally stays inactive for code, commit messages, API references, short UI strings, and legal text.
- German, including German typography and common translated AI-writing patterns
- English, with its own punctuation and editorial conventions
Mixed-language documents are supported, but each passage must be evaluated under its own language rules.
.codex-plugin/plugin.json Codex plugin metadata
skills/humanizer/SKILL.md Shared workflow
skills/humanizer/references/ German and English pattern catalogs
skills/humanizer/agents/ Optional Codex UI metadata
evals/cases.yaml Activation and output-quality cases
The skill follows the Agent Skills specification and OpenAI's guidance for skills and plugin packaging.
evals/cases.yaml contains direct and indirect activation
prompts, requests that should not trigger the skill, and output constraints for
factual preservation. The cases describe observable expectations rather than
one canonical rewrite.
Humanizer cannot supply missing reporting, sources, or subject knowledge. Removing stylistic signals does not correct factual errors already present in the source. Editorial judgment remains necessary, especially for technical, medical, financial, or regulated material.
The pattern catalogs build on observations documented by Wikipedia's WikiProject AI Cleanup and the German page on Anzeichen für KI-generierte Inhalte, combined with practical baseline testing.
Contributions are welcome. Read CONTRIBUTING.md before adding a pattern or changing a language rule.