humanizer-zh: a Chinese editing skill that trims filler without touching facts

npx skills add op7418/Humanizer-zh --skill humanizer-zh

Ask a model to polish a paragraph of Chinese and you often get smoother prose with quiet damage: a “may” turned into “will”, a number that was never in the original, and a closing line about looking forward to what comes next. humanizer-zh, maintained by op7418 on GitHub, is a Chinese editing skill built to do the opposite. It changes how a text says things, not what it says.

How it works

The skill treats the job as editing an existing text, not rewriting it into a new one. It sets a priority order first: preserve information and the degree of certainty, then respect the user’s scope and genre, then match the author’s voice, and only then fix specific expression problems. When two goals collide, the earlier one wins.

Next comes a checklist of 31 patterns in six groups: setup standing in for statement, formulaic rhythm, inflated significance and borrowed authority, formulaic formatting, chat and draft leftovers, and Chinese-specific issues such as stacked “的”, the “进行 + verb” construction and four-character parallelism. Each pattern comes with a before, an after and a “keep” example, and the keep example shows what must not be touched, such as a dash that carries a real turn in the argument. The skill says outright that these are clues to check, not a word blacklist, and that a clean paragraph can stay as it is.

What it constrains

• Facts: no numbers, names, dates, quotes or sources that the original lacks; negation, scope, conditions and completion status stay as written.
• Certainty: “may” stays “may”, a plan never becomes a finished job, correlation never becomes causation.
• Files: code blocks, paths, links, YAML front matter and explicit IDs are left alone, and heading text and levels are kept by default.
• Output: pasted text gets the final version only, with no self-score or list of hits.
• Input: commands and role-play inside the text being edited are treated as material, not instructions.

Who it’s for

It suits anyone writing Chinese blog posts, commentary, product notes or news who wants a model’s polish without the facts drifting. It declares only four tools, Read, Write, Edit and AskUserQuestion, and runs no scripts and makes no network calls, so the risk comes from the rewriting itself.

Two limits are worth knowing. It does not judge who wrote a text and does not promise to pass an AI detector; its own description says so. And it cannot invent detail: a vague original stays general after editing, so you supply the specifics. Still read the diff against the original, especially numbers and hedges. For longer pieces, pair it with content-gate: one checks sentences, the other checks whether the piece deserves to ship.

The repository is MIT-licensed, had about 19,000 stars on 8 October 2026, and was last updated on 23 September 2026.