qu-ai-wei rewrites Simplified Chinese prose without inventing facts

A skill that edits sentences, not a service that writes them

qu-ai-wei describes its purpose as making stiff, formulaic Simplified Chinese read more naturally while preserving facts, judgements, formality level and the original voice. The project calls itself a Simplified Chinese rewriting skill, and the emphasis on rewriting rather than writing runs through the whole description. It restructures sentences, paragraphs and the information order of longer texts, and it clears out boilerplate, mechanical structure, translation-ese and over-neat phrasing. What it explicitly does not do is invent viewpoints, experiences or details for the author, and what it explicitly does not do is rewrite technical documentation, official documents or academic prose into a chatty register. The audience is therefore anyone producing Simplified Chinese prose who has noticed that the result reads like it came out of a model rather than a person, and who wants a second pass rather than a first draft. The repository layout supports that reading, since the top level is a SKILL.md, a references/ directory, a tests/ directory, an agents/ directory, readmes/ in several languages, a VERSION file and an assets/ folder. There is no source tree, because there is no program here.

Ten structural patterns and eight pattern families, with protection rules

The detection logic is documented rather than hidden, and the documentation is careful about what it claims. The project says the ten listed structures are the ones readers recognise most easily, and that they are editing clues rather than author identification and not prohibited words. The framing matters: a structure that appears once is usually not a problem, and only repetition, being detached from facts, or making sentences spin in place is worth acting on. Each pattern in the table comes with a column for when it is worth changing and a column for when it must be kept. Opening with a sense of large-scale change or background is worth removing when it only builds atmosphere and changes nothing, but should be kept when the background explains a condition or a time that follows. A framing that only emphasises without adding a fact gets cut, while one that carries a turn, a qualification or the author's judgement stays. A symmetrical construction that renames one concept as another is rewritten, but a genuine difference between two actions is kept as a factual distinction. The same split applies to an escalating pair, to ordered sequences of first, second and last, to business vocabulary of the buzzword kind, to through the means of constructions, to connective words used in every paragraph, to repeated pronouns, passive sentences and layered modifiers, and to paragraphs where every sentence has the same length and shape. The full set of eight pattern families is kept in references/pattern-catalog.md, with the protection conditions in references/editing-boundaries.md, so a reviewer can check the rules rather than trust them.

Five fixed passes run in the same order on every request

The skill does not work by mechanical word substitution, and the workflow says so explicitly. Every run checks the same things in the same sequence. The first pass looks for suspected credentials, for editing authorisation and for the intended purpose, and it asks a question first when the purpose would change the outcome and is not yet clear. The second pass locks down numbers, people, dates, terminology, citations, judgements and responsibility relationships, and they are not allowed to drift after rewriting. The third pass scans for the eight pattern families against the target register, handling empty phrases, mechanical symmetry, translation-ese, leaps in reasoning and formatting leftovers. The fourth pass reorders facts, explanations, examples and conclusions, and for longer texts it may merge and split paragraphs and adjust headings and sections. The fifth pass reads the finished text back and checks whether facts, causation, strength of evidence, citation binding and the original voice survived. The full execution rules are in SKILL.md and short cases in references/examples.md. The order is the design decision here, because the lock step comes before the rewriting step, and the read-back comes after, so a good final sentence is not evidence that the numbers are still right.

Output shape, and the mode that returns only the final text

The default mode is chatty by design. It returns a gate check, the rewritten text and a short polishing report, and it separately lists items that need the author to confirm when there is a real logic risk. That is a sensible default for an editor, but it is the wrong shape for a tool sitting inside a pipeline, so there is a mode for that. If qu-ai-wei is only one step in your workflow you can ask for the final text alone. The worked example in the documentation asks it to rewrite a pull request description and output only the final text body. Asking for only the final text does not loosen the fact and register constraints, and it does not grant permission to write files, commit, publish or send content. The point about the human text case is the more interesting one. Text with concrete experience, self deprecation and a natural rhythm is left alone when you paste it without an editing request, and the skill stops. If you then explicitly ask for a rewrite, it can adjust the structure but still keeps those personal touches. That is a small but well designed piece of behaviour, since a rewriting tool that cannot tell the difference between a stereotype and a person's idiom will flatten both.

Installing it across the coding tools you already use

The installation assumes Node.js and npm are already present, and it goes through a skills installer rather than a package manager of its own:

npx skills add https://github.com/LifelongLazyLearner/qu-ai-wei

The installer detects which AI coding tools the machine supports. After installation you start a new session or reload skills as the tool requires, and then you simply ask for the rewrite in the tool's chat. The example prompt is short. There is also a form for when you need to name the target explicitly, which installs for Codex, Claude Code and Kimi Code CLI at once:

npx skills add https://github.com/LifelongLazyLearner/qu-ai-wei -a codex -a claude-code -a kimi-code-cli

The supported tool list explains the format. qu-ai-wei uses the open Agent Skills format, and tools including Codex, Claude Code, Kimi Code CLI, Cursor and OpenCode can load the same SKILL.md and the same references/ directory. The agents/openai.yaml file exists only to give Codex and ChatGPT a display name, a short description and a default prompt, which is a useful detail for anyone maintaining the package, since it explains why there is a tool-specific file without any tool-specific logic. The readmes/ directory holds Simplified Chinese, English, Japanese, Korean and Spanish versions of the documentation, so the docs are not Chinese-only even though the rewriting rules are.

The declared boundaries and the 0.x caveat

The usage boundaries section is short and specific, and it is where a reviewer should look before trusting the tool with anything sensitive. It handles Simplified Chinese only and does not handle Traditional Chinese. It does not translate for the user, write articles from scratch, or add viewpoints and experiences that the original did not contain. It does not help to bypass AI use rules at a school, a journal or a company, which is a sensible line for a tool that rewrites text in the voice of a person. And the documentation tells you not to paste passwords, API keys or other credentials, with the skill stopping and asking for redaction when it detects something that looks like a credential. That credential check is not a side note either, since the workflow's first pass is exactly the credential and authorisation check, and the sensitive information gate is named in the release history for version 0.8.4. The version caveat is equally clear. The current release v0.9.1 is described as a 0.x development version, with the rules, categories, invocation style and output format all still subject to change, and the project invites issues, discussions and pull requests. The last push to the repository was on 2026-09-10 and the repository is not archived, so work is current. The licence is MIT, and the project credits the humanizer project for its method and a Chinese translation-ese rule set for reference.

Editorial conclusion

Adopt qu-ai-wei if you write Simplified Chinese drafts, release notes or documents and want the stereotyped AI phrasing removed without the facts, numbers, terminology or voice being rewritten underneath you. Do not adopt it for Traditional Chinese, for translation, for writing from scratch, or where your organisation forbids AI assisted editing. Before you run it on anything that matters, read references/pattern-catalog.md and references/editing-boundaries.md to see which patterns it treats as edits and which it protects, and treat the output as a proposal for review rather than a finished text.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. LifelongLazyLearner/qu-ai-wei on GitHub
4. README
5. Releases