op7418/Humanizer-zh

Humanizer‑zh – 中文文本润色 Skill

What it is

• A Claude skill that lets the Claude‑based “Claude Code” agent rewrite Chinese prose, comments, or documentation.
• It focuses on removing empty‑talk, repetition, and template‑style phrasing while preserving facts, certainty levels, and the author’s voice.
• The skill does not claim to detect authorship or guarantee that the output will pass AI‑detector tools.

How to install

Method	Command / Steps	Notes
1️⃣ One‑click (recommended)	npx skills add https://github.com/op7418/Humanizer-zh.git	Installs the skill into the correct ~/.claude/skills/ folder automatically.
2️⃣ Manual clone	git clone https://github.com/op7418/Humanizer-zh.git ~/.claude/skills/humanizer-zh	Directly clones the repo into Claude’s skill directory.
3️⃣ Manual copy	Download ZIP → copy Humanizer-zh folder to ~/.claude/skills/ (macOS/Linux) or %USERPROFILE%\.claude\skills\ (Windows)	Useful when you need to edit the files locally before installing.

After installing, restart Claude Code (or reload skills) and type /humanizer-zh in a conversation to invoke the skill.

How to use

• Inline text: paste the passage you want polished and prepend the request with humanizer-zh.

  请用 humanizer-zh 润色下面的文章，保留技术评论的语气：
  [原文]
  

• File editing: ask the skill to rewrite a Markdown file, e.g.

  请用 humanizer-zh 润色 article.md 中的正文。
  

• Suggestion‑only mode: add a note like “只审阅，不修改”， and the skill will return edit suggestions instead of a rewritten file.
• You can also provide a short author‑sample so the skill can mimic the author’s style, but it will never copy the sample’s data or opinions.

Core editing principles

1. Information first – never fabricate facts, numbers, sources, or personal experience.
2. Maintain certainty – keep words like 可能 (maybe) or 尚未确认 (not confirmed) unchanged.
3. Respect voice & tone – preserve the original level of formality (academic, business, casual) and the author’s personality.
4. Selective rewriting – only modify phrasing that falls under the 31 defined “check points”; sentences that are already clear are left untouched.
5. File‑aware handling – code blocks, paths, YAML, links, headings, and anchors are kept intact unless an explicit structural change is requested.

What the skill actually does

• It is not a standalone detector; it is a set of editing rules that an Agent (Claude) follows when asked.
• The rules are organized into six categories (A‑F) covering:
  • filler‑style introductions, formulaic rhythm, inflated language, formatting quirks, leftover draft chatter, and Chinese‑specific constructions.
• Each check point includes a condition, before‑after example, and boundaries for what must be retained.
• The repository ships with a test suite (tests/README.md) containing 18 short‑text cases, a Markdown file example, and a structural‑check script to verify that the skill respects the rules.

Limitations

• The skill only works inside Claude Code; it cannot be called from other environments without the Claude skill infrastructure.
• It does not guarantee that the output will evade AI‑writing detectors.
• It does not attempt to infer missing data; if the original text lacks evidence, the skill will ask the author for it rather than invent it.
• The quality of the rewrite depends on the underlying Claude model; the skill merely guides the model with its rule set.

License

• Distributed under the MIT License (see LICENSE).
• The rule set builds on the original Humanizer project (v3.0.0) and incorporates contributions from several upstream repositories (e.g., blader/humanizer, hardikpandya/stop-slop).

---

Humanizer‑zh is a practical, rule‑driven Claude skill for polishing Chinese text while keeping the author's intent and factual content intact. It is installed via a single npx command or manual cloning, invoked with /humanizer-zh, and respects a detailed 31‑point editing checklist.