Improve Chinese AI-assisted writing by fixing material and structure first

Human Writing Skill helps draft and revise Chinese prose. Its premise is that generic AI writing usually comes from thin source material and paragraphs that add no new fact, example, or consequence—not merely from a list of banned words.

What you give it and what it can return

Provide the source material and citeable references, intended audience and publishing context, facts and wording that cannot change, the voice to retain, and experiences or cases that must not be fabricated. Ask it to identify material gaps before drafting or revising. It can separate factual from fictional writing, remove repeated explanation and empty conclusions, adjust Chinese word order, sentence length, and pacing, then use check_prose.py for defined hard rules.

The full Skill includes entry files plus references for long-form, factual, fictional, formatting, and revision work. A lite version can be pasted into chat interfaces that cannot install the full Skill.

A practical revision sequence

  1. Decide whether the material supports the article, needs research, needs a question to the author, or should be shorter.
  2. Review each paragraph for a new piece of information or consequence.
  3. Revise structure before wording and rhythm.
  4. Run check_prose.py where useful, then let a human judge factual accuracy and voice.

Version 1.1.0 shifts attention from fixed forbidden strings toward rhetorical moves, such as inventing and overturning a reader misconception. Its checks include variations of reversal patterns, AI-style parallelism, sentence-length variation, and connective density. A script can flag a rule; it cannot decide whether prose is good.

Boundaries

It cannot create real interviews, data, or lived experience. It does not guarantee that text passes an AI detector, and it should not be used to conceal plagiarism or fabricated facts. Overapplying any "human" rule can become another template.

Source and license

The upstream says the repository contains original rules and tools, not third-party articles, training data, or model weights.