shuorenhua (说人话) is a Chinese-first editing skill for Claude Code, Codex, Cursor and ChatGPT that removes filler, repetition and stiff phrasing from a draft while leaving its facts, conditions, commitments and tone intact.

What it is

shuorenhua is an agent skill: a set of editing rules loaded into an assistant rather than a standalone application, distributed through the skill and plugin mechanisms of Claude Code, Codex and Cursor and through a hosted ChatGPT GPT. It sits in the Chinese-language AI-writing toolchain, is written in Python, and is released under the MIT licence. The concrete thing it works against is the layer of boilerplate that survives AI drafting — the announcement and thanks that open a report, the signpost, the sentence that says the same thing twice — so that a draft can be sent without a manual line-editing pass.

What it keeps defines it as sharply as what it cuts. Numbers, dates and version numbers stay, as do conditions and commitments (「如果」「得到批准后」「可能」), quotations, terminology, code blocks and the writer's own judgement and tone, including evaluative remarks that read like promotion, which are retained by default. The default output is only the revised full text, with no score and no explanation attached, and a passage that is already fine is returned unchanged. The skill does not judge whether a text was written by AI, and it does not promise that output will pass AI detection.

Key capabilities

• Clears contentless greetings, previews and thanks, signposts such as 「值得注意的是」, roundabout constructions such as 「完成了对……的调整工作」, and statements that repeat the same point.
• Preserves numbers, dates, version numbers, conditions, commitments, hedges, quotes, terminology, code blocks and authorial voice, and returns the original when nothing needs fixing.
• Offers three change-scope modes: free adjustment / structural, keep structure / bounded (sentence-internal cleanup only, with whole-sentence deletions suggested rather than applied), and in-place / in-place (no deletion, merging or reordering).
• Defaults to preserving the original paragraph and sentence structure for Chinese public long-form writing of roughly 1,000 characters or more.
• Adds a plain-writing step (v2.6) that opens only on explicit requests such as 「别吹」「写平实点」「去掉营销味」, replacing degree-only intensifiers like 「极致」「完美」 and dropping contentless slogans while retaining checkable "first" claims, figures and terminology; it only lowers the tone, never raises it.
• Supports review-only requests (「只标问题，不改写」) that flag problems without rewriting, and handles English as well as Chinese.
• Ships with a 123-case test set — 65 cases needing cleanup, 58 that should not change, 7 readability cases — with published results in evals/results-v2.6.0.md and criteria in evals/v2.5/criteria.md.

Who uses it and how

• Developers working inside Claude Code, Codex or Cursor paste the draft and state the request in conversation; no command has to be memorised, and naming 「用说人话」 in the request ensures the skill is invoked.
• Teams standardise on one skill across repositories by adding --global, which installs to the user directory so every project can use it.
• ChatGPT users skip installation entirely and work through the hosted 说人话 GPT.
• Manual file copying and OpenClaw are supported for users who want to place the rules by hand or work from a pasted prompt.
• Writers of Chinese public long-form articles use it with structure preserved, and reviewers who want problems marked rather than text rewritten use the review-only mode.

Getting started

Install with npx skills add MrGeDiao/shuorenhua and select the tools to install into; Claude Code users can instead run /plugin marketplace add MrGeDiao/shuorenhua followed by /plugin install shuorenhua@shuorenhua, and ChatGPT users can open the hosted GPT with no installation.

How it compares

The projects named alongside it — stop-slop for English redundancy rules, humanizer for English AI-writing patterns, avoid-ai-writing for an AI-writing problem taxonomy, and speak-human-tw for Traditional Chinese editing rules — are largely English or Traditional Chinese rule sets, which leaves shuorenhua as the Simplified-Chinese-first option among them. Its distinguishing material is the layered change-scope controls and the published 123-case evaluation, rather than a broader rule catalogue.

When to use it — and when not to

Pick it when the draft is already written and the job is to cut padding without losing facts or voice, and accept that the defaults are conservative: promotional-sounding judgements are kept unless 「别吹」 is requested, and on longer public articles slang and stacked metaphors are cleaned up less thoroughly than the project would like (issue #16). Be aware that plain-writing mode has occasionally demoted legitimate praise in testing, and that a sentence consisting only of a preview may be deleted outright (case B-33). It is not a tool for detecting AI authorship, and anyone expecting a guarantee about AI detectors should look elsewhere.

npx skills add MrGeDiao/shuorenhua

/plugin marketplace add MrGeDiao/shuorenhua
/plugin install shuorenhua@shuorenhua

npx skills add MrGeDiao/shuorenhua

把这段去掉套话，保留信息和我的口气：……
只在句内改，不删句、不并句：……
别吹，写平实点：……
只标问题，不改写：……