name	renwei-zh
description	Audit and revise Chinese text to reduce formulaic AI-style writing while preserving facts and author voice. Use for 中文润色, 去机器味, 去 AI 味, 公众号/知乎/小红书/商业/学术稿改写, 模型化表达审计, 声线保留, fact-safe Chinese editing.
allowed-tools	["Read","Grep","Glob","Bash"]

renwei-zh

Use this skill when the user wants Chinese text reviewed, revised, or audited for formulaic model-like writing. The objective is better Chinese writing and voice preservation, not detector evasion.

Critical Rules

Preserve facts. Do not add numbers, dates, sources, institutions, citations, people, claims, or quotes unless the user supplied them.
If concrete evidence is missing, write [需补事实] or ask for the missing fact instead of inventing it.
Do not over-humanize. If the text already has a clear human voice, clean only formatting and obvious friction.
Scenario first. Academic, official, legal, medical, and business writing must stay appropriately formal.
Scores are editing signals, not proof of authorship.

Workflow

Classify the task: audit, rewrite, voice-match, or fact-safe rewrite.
Identify scenario: business, academic, official, wechat, zhihu, xhs, creative, or general.
Run the CLI audit when a text file is available:

PYTHONPATH="$CLAUDE_SKILL_DIR/src" python3 -m renwei_zh audit "$FILE" --scenario "$SCENARIO" --json

Read the highest-severity evidence first: fact-risk items, red hits, clusters, and model fingerprints.
Rewrite structurally: delete formulaic scaffolding, replace abstract verbs with concrete user-supplied actions, vary sentence rhythm, and keep domain-appropriate terminology.
Re-audit the revised text when practical.
Output the final text plus a short evidence report: what changed, facts preserved, residual risks.

Load Router

If the text is business, academic, official, platform-specific, or creative, read references/05-vertical-scenarios.md.
If the audit shows Claude, GPT, DeepSeek, Qwen, Gemini, or Agent-style fingerprints, read references/03-llm-fingerprints.md.
If many syntactic shells appear, read references/02-syntactic-shells.md.
If the user asks why something is considered AI-like, read references/01-pattern-library.md.
If the user asks for metrics or thresholds, read references/04-statistical-features.md.
If the user only wants command-line usage, read references/06-detection-tools.md.

Golden Template

When rewriting, follow this output contract:

【门检】
- 场景:
- 是否需要重写:
- 事实风险:

【改写后】
...

【证据报告】
- 删除/改写的高风险模式:
- 保留的事实:
- 未确认但需要用户补充的事实:
- 残留风险:

Boundaries

NOT this skill: English prose editing, plagiarism, impersonation, fake citations, SEO spam, detector bypass guarantees, or rewriting content to hide misconduct.

renwei-zh

Use this skill when the user wants Chinese text reviewed, revised, or audited for formulaic model-like writing. The objective is better Chinese writing and voice preservation, not detector evasion.

Critical Rules

Preserve facts. Do not add numbers, dates, sources, institutions, citations, people, claims, or quotes unless the user supplied them.

If concrete evidence is missing, write [需补事实] or ask for the missing fact instead of inventing it.

Do not over-humanize. If the text already has a clear human voice, clean only formatting and obvious friction.

Scenario first. Academic, official, legal, medical, and business writing must stay appropriately formal.

Scores are editing signals, not proof of authorship.

Workflow

Classify the task: audit, rewrite, voice-match, or fact-safe rewrite.

Identify scenario: business, academic, official, wechat, zhihu, xhs, creative, or general.

Run the CLI audit when a text file is available:

PYTHONPATH="$CLAUDE_SKILL_DIR/src" python3 -m renwei_zh audit "$FILE" --scenario "$SCENARIO" --json

Read the highest-severity evidence first: fact-risk items, red hits, clusters, and model fingerprints.

Rewrite structurally: delete formulaic scaffolding, replace abstract verbs with concrete user-supplied actions, vary sentence rhythm, and keep domain-appropriate terminology.

Re-audit the revised text when practical.

Output the final text plus a short evidence report: what changed, facts preserved, residual risks.

Load Router

If the text is business, academic, official, platform-specific, or creative, read references/05-vertical-scenarios.md.

If the audit shows Claude, GPT, DeepSeek, Qwen, Gemini, or Agent-style fingerprints, read references/03-llm-fingerprints.md.

If many syntactic shells appear, read references/02-syntactic-shells.md.

If the user asks why something is considered AI-like, read references/01-pattern-library.md.

If the user asks for metrics or thresholds, read references/04-statistical-features.md.

If the user only wants command-line usage, read references/06-detection-tools.md.

Golden Template

When rewriting, follow this output contract:

【门检】 - 场景: - 是否需要重写: - 事实风险: 【改写后】 ... 【证据报告】 - 删除/改写的高风险模式: - 保留的事实: - 未确认但需要用户补充的事实: - 残留风险:

Boundaries

NOT this skill: English prose editing, plagiarism, impersonation, fake citations, SEO spam, detector bypass guarantees, or rewriting content to hide misconduct.