Text Naturalizer
ripplek/text-naturalizer
Remove marcas de IA de textos em chinês, inglês e bilíngues
Ver no GitHubSkill para Claude Code e Codex que remove marcas típicas de texto gerado por IA em escritas em chinês, inglês e bilíngues. Preserva significado, fatos, terminologia e a voz do autor enquanto naturaliza a prosa, sendo ideal para READMEs, descrições de PRs, emails e documentação técnica.
Instalação
npx convez-ai ripplek/text-naturalizer1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Polir READMEs e documentação técnica
- Remover traços de ChatGPT/LLM de rascunhos
- Reescrever emails, essays e copy de produtos
- Reescrever texto seguindo amostra de estilo do autor
Veredito

Ferramenta especializada e bem documentada que aborda um problema específico e real: naturalizar prosa gerada por IA sem perder informação técnica ou voz do autor. Implementação modular com regras de padrão por idioma mostra design sólido, mas utilidade limitada a quem escreve em chinês/inglês ou bilíngue.
README
A tradução para o português deste README ainda está na fila.
text-naturalizer
English | 中文
text-naturalizer is a Claude Code and Codex skill for making Chinese, English, and mixed Chinese-English writing sound more natural. It removes common AI-generated prose tells while preserving meaning, facts, terminology, and the author's intended voice.
Use it for:
- Chinese, English, or bilingual polishing
- Removing ChatGPT / LLM traces from drafts
- README files, PR descriptions, emails, essays, product copy, and technical documentation
- Rewriting toward a provided writing sample
It does not:
- Invent data, citations, sources, examples, or claims
- Turn every text into casual prose
- Remove important details just to sound more human
- Randomly replace stable terms, API names, product names, or domain language
Install in Claude Code
Claude Code loads personal skills from ~/.claude/skills/<skill-name>/SKILL.md. After installation, invoke this skill with /text-naturalizer, or let Claude trigger it automatically from the skill description.
Install from GitHub:
mkdir -p ~/.claude/skills
git clone https://github.com/ripplek/text-naturalizer.git ~/.claude/skills/text-naturalizer
If you maintain the repository locally, use a symlink so local edits take effect immediately:
mkdir -p ~/.claude/skills
ln -s /Users/dodo/projects/text-naturalizer ~/.claude/skills/text-naturalizer
Claude Code watches existing skill directories for changes. If ~/.claude/skills did not exist before, restart Claude Code after creating it.
Install in Codex
mkdir -p ~/.codex/skills
ln -s /Users/dodo/projects/text-naturalizer ~/.codex/skills/text-naturalizer
Invoke it in Codex with:
Use $text-naturalizer to make this Chinese or English draft sound more natural and less AI-generated.
Usage examples
Claude Code:
/text-naturalizer Make this README section sound more natural while preserving the technical details:
...
Codex:
Use $text-naturalizer to polish this PR description. Keep it factual and concise.
With a writing sample:
Use $text-naturalizer to rewrite the draft below in my style.
Writing sample:
...
Draft:
...
How it works
SKILL.md keeps the main workflow and editing constraints small. Language-specific rules live in references/:
references/chinese-patterns.md: Chinese and mixed-language issues such as translationese, inflated significance, marketing jargon, mechanical transitions, and slogan-like endings.references/english-patterns.md: English issues such as significance inflation, promotional language, superficial-ingclauses, rule-of-three phrasing, and chatbot artifacts.
When the skill runs, it first identifies the language and register, then loads the matching reference file. Technical documentation stays restrained. Personal writing can keep stronger rhythm, uncertainty, and point of view.
Maintenance
Validate the skill after editing:
python3 /Users/dodo/.codex/skills/.system/skill-creator/scripts/quick_validate.py /Users/dodo/projects/text-naturalizer
If local Python does not have PyYAML, run validation in a temporary virtual environment:
python3 -m venv /tmp/text-naturalizer-validate
/tmp/text-naturalizer-validate/bin/python -m pip install PyYAML
/tmp/text-naturalizer-validate/bin/python /Users/dodo/.codex/skills/.system/skill-creator/scripts/quick_validate.py /Users/dodo/projects/text-naturalizer
Sources
This skill reorganizes and rewrites ideas from two MIT-licensed projects:
blader/humanizerop7418/Humanizer-zh
See LICENSE for details.