Renwei-zh: Auditoria de Escrita Chinesa Baseada em Evidências
Zaoqu-Liu/renwei-zh
Audita chinês para remover tom formulaico sem alterar fatos
Ver no GitHubSkill Python que audita e revisa textos em chinês, identificando padrões formulaicos e tom de IA enquanto preserva fatos e voz do autor. Retorna evidência estruturada, clusters de padrões e proteção de dados factuais (números, datas, citações). Disponível como skill para Claude ou CLI reproduzível.
Instalação
npx convez-ai Zaoqu-Liu/renwei-zh1 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
- Revisar rascunhos em chinês que soam muito 'gerados por IA'
- Validar segurança factual antes de publicação (redes sociais, releases de produto)
- Otimizar clareza e credibilidade de conteúdo em chinês mantendo voz autoral
Veredito

Ferramenta sólida e diferenciada para quem escreve profissionalmente em chinês. Foca em qualidade textual com proteção factual, não em detecção genérica de IA—abordagem honesta e reproducível via CLI ou skill Claude.
README
A tradução para o português deste README ainda está na fila.
renwei-zh

renwei-zh is a Chinese writing audit and revision skill. It helps users turn formulaic, model-like Chinese drafts into clearer, more credible writing while preserving facts and the author's voice.
It is not an "AI detector bypass" tool. It is a writing-quality tool for people who care about natural Chinese, factual safety, and repeatable review.
What You Get
| User need | What renwei-zh does |
|---|---|
| "这段中文太 AI 了" | Finds formulaic wording, syntactic shells, PR tone, overclaims, and model-like fingerprints. |
| "不要把事实改错" | Marks numbers, dates, citations, named sources, statistics, and claims as preservation-sensitive. |
| "我想知道哪里有问题" | Returns evidence, clusters, metrics, and prioritized guidance instead of vague feedback. |
| "我要在 Claude/Codex 里用" | Ships a compact SKILL.md with a Load Router and Golden Template. |
| "我要命令行/CI 可复现" | Ships a Python CLI with stable JSON output and tests. |
5-Minute Start
git clone https://github.com/Zaoqu-Liu/renwei-zh.git
cd renwei-zh
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[zh]"
renwei-zh audit examples/11-pure-ai-rewrite/11-before.md --scenario business
No API key is required. The optional zh extra installs jieba for better Chinese tokenization. Without it, the CLI still works with a deterministic fallback tokenizer.
Use It As A Skill
Symlink the repository into your Claude skills directory:
mkdir -p ~/.claude/skills
ln -s "$(pwd)" ~/.claude/skills/renwei-zh
Then use:
/renwei-zh 请审计并改写这段中文产品发布稿。保留所有事实,不要新增数据。
The skill will follow this user-facing flow:
flowchart LR
A["用户中文草稿"] --> B["场景判断"]
B --> C["证据审计"]
C --> D["事实保护"]
D --> E["结构性改写"]
E --> F["复审与残留风险"]
F --> G["可发布终稿"]
Use It From The CLI
Human-readable report:
renwei-zh audit draft.md --scenario business
JSON report:
renwei-zh audit draft.md --scenario business --json > audit.json
Read from stdin:
cat draft.md | renwei-zh audit --scenario wechat
Exit codes:
| Code | Meaning |
|---|---|
0 | Score is at least 70. |
2 | Score is below 70; structural revision is likely needed. |
How To Read A Report
The CLI does not accuse a text of being AI-written. It reports editing evidence.
Score: 0 / 100 (rewrite)
Evidence
- Red hits: 21
- Yellow hits: 5
- Clusters: 2
- Fact-risk items: 1
Guidance
- Fact guard: preserve numbers, citations, dates, names, and sources...
- Break dense clusters first...
- Priority pattern families: perfect_ending, product_pr, preposition_disease
Interpretation:
| Field | How to use it |
|---|---|
Score | Editing severity. It is not authorship proof. |
Red hits | High-priority formulaic patterns to revise first. |
Yellow hits | Softer evidence; only revise when dense or contextually wrong. |
Clusters | Dense windows with multiple pattern families. Fix these before isolated hits. |
Fact-risk items | Facts that must not be invented, deleted, or silently altered. |
Guidance | The shortest path to a better rewrite. |
Scenarios
Use --scenario to reduce false positives and get better guidance.
| Scenario | Best for | Protected style |
|---|---|---|
business | 产品稿、周报、提案、融资材料 | metrics, source attribution, risk boundaries |
academic | 论文、摘要、基金、研究总结 | citations, p values, passive voice, nominalization |
official | 公文、机构材料 | formal register |
wechat | 公众号长文 | paragraph rhythm and author stance |
zhihu | 知乎回答、知识解释 | technical specificity and personal experience |
xhs | 小红书短内容 | price, product model, actual usage condition |
creative | 小说、诗歌、文案 | voice, ambiguity, character diction |
general | fallback | balanced defaults |
Example Walkthrough
Try the dense business sample:
renwei-zh audit examples/11-pure-ai-rewrite/11-before.md --scenario business
renwei-zh audit examples/11-pure-ai-rewrite/11-after.md --scenario business
What you should notice:
- the before sample has dense PR tone, overclaims, perfect-ending language, DeepSeek-like imagery, and clustered formulaic structure
- the after sample keeps the business topic but becomes more concrete and less ceremonial
- fact-risk items remain visible because numbers and sources should still be protected
Examples live in:
examples/
├── 01-xhs-skincare/
├── 03-wechat-essay/
├── 07-deepseek-imagery/
└── 11-pure-ai-rewrite/
Configuration
Most users only need --scenario.
Advanced users can edit the default rule catalog:
src/renwei_zh/data/patterns.v1.json
Repository rule:
- add new default rules to the catalog
- do not duplicate regexes in standalone scripts
- add tests or false-positive notes for high-severity rules
Install Troubleshooting
If renwei-zh is not found:
source .venv/bin/activate
python -m pip install -e ".[zh]"
python -m renwei_zh audit examples/11-pure-ai-rewrite/11-before.md
If package installation is blocked by your system Python, use a virtual environment:
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[zh]"
If you only need the skill and not the CLI, symlink the repo into ~/.claude/skills and use /renwei-zh directly.
Project Layout
.
├── SKILL.md # Claude/Codex skill entry point
├── src/renwei_zh/ # Python package and CLI
├── src/renwei_zh/data/ # Single source for default audit rules
├── references/ # Deep scenario and pattern knowledge
├── examples/ # Before/after samples
├── assets/ # README and project visuals
├── tests/ # Regression and contract tests
├── benchmarks/ # Corpus layout for future calibration
└── .github/ # CI, issues, PR template
Development
python -m pip install -e ".[dev,zh]"
python -m pytest
python -m ruff check .
python -m mypy src
python -m build
Responsible Use
Use this project to improve writing quality, not to fabricate authorship, fake citations, impersonate someone, or evade required AI-use disclosure.
Text-only signals have limits. Treat renwei-zh reports as editing evidence, never as proof that a person did or did not use AI.
License
MIT.