Deai zh gpt style

De-AI Chinese GPT Style

Use this skill to make Chinese text sound less like GPT-5.5/GPT-5.4 output while preserving meaning. The goal is not to make the text sloppy; it is to remove model-like scaffolding, replace formulaic phrases with natural equivalents, and clean spacing in mixed Chinese text.

Core Workflow

1. Preserve the user's meaning, facts, numbers, names, and required format.
2. Remove or rewrite obvious GPT-style scaffolding.
3. Replace formulaic phrases with natural equivalents only where the sentence still reads smoothly.
4. Remove spaces between Chinese and English, and between Chinese and numbers.
5. If editing chat records, rewrite by speaker and channel instead of polishing isolated sentences.
6. Read the result aloud once. If it still sounds like a checklist, consultant memo, scripted dialogue, or model answer, simplify again.

Work-Order Naturalness Mode

Use this mode for task prompts, long queries, review notes, platform fields, project briefs, or any text that should sound like a real person assigning work. The goal is to turn instruction-manual language into a believable workplace handoff while keeping all facts, constraints, file names, numbers, dates, output formats, and scoring requirements unchanged.

Principle: keep requirements exact, but make the surrounding explanation sound like work. Background, purpose, and deliverable use can be direct and ordinary; sheet names, formulas, thresholds, exclusions, and pass/fail checks should stay precise.

Prefer:

• Concrete actions: 做清洗、关联和汇总, 把10月至次年3月拉出来复盘, 给运营总监快速看结论.
• Short active phrasing: 只纳入..., 用...补齐..., 最后交1个xlsx, 评分里要放进..., 写清..., 不要....
• Business purpose in plain words: 这个中文说明是给运营总监快速看结论的, 用来安排Q2排班, 下周会上要看.
• Slightly uneven sentence rhythm. Not every sentence needs the same 请...需...并... pattern.

Avoid overusing:

• Instruction-manual words: 面向, 基于, 纳入范围为, 交付, 口径, 维度, 形成, 支撑, 核心, 确保, 需, 应, 不得.
• Abstract noun piles: 数据清洗关联汇总工作, 结论快速阅读用途, 运营管理决策支撑.
• Repeating hard modal verbs in every sentence. Keep 必须/不得/至少 for real hard constraints; use 要/只/最后/写清/不要 elsewhere.

Punctuation pass:

• Remove unnecessary comma pauses when the sentence reads naturally without them. AI-style Chinese often inserts commas after every short phrase, which makes a work note sound segmented and mechanical.
• Good candidates: 单看某一条FRED序列，只能知道... -> 单看某一条FRED序列只能知道...; 合在一起，才能看到... -> 合在一起才能看到...; 没有它，存款/资产...算不出来 -> 没有它存款/资产...算不出来; 少了其中一张，就很难判断... -> 少了其中一张就很难判断...; 只有银行信贷总量的话，看不出... -> 只有银行信贷总量的话看不出....
• Keep commas for long parallel lists, inserted explanations, contrast turns, or places where removing punctuation would make the sentence hard to parse.

Examples:

该中文说明面向运营总监阅读，用于快速了解主要结论 -> 这个中文说明是给运营总监快速看结论的
纳入范围为PULocationID或DOLocationID属于机场zone的行程 -> 只纳入PULocationID或DOLocationID属于机场zone的行程
交付1个xlsx工作簿，并附800字以内中文说明 -> 最后交1个xlsx工作簿，再附一段800字以内的中文说明
评分公式至少纳入异常率、收入变化率、现金占比变化和订单量变化 -> 评分里要放进异常率、收入变化率、现金占比变化和订单量变化
基于附件无法确认原因的异常 -> 附件本身看不出原因的异常
用于安排Q2车辆排班、司机沟通和异常收费复核 -> 用来安排Q2车辆排班、司机沟通和异常收费复核

Chat Naturalness Mode

Use this mode for group chats, DMs, customer-service logs, work coordination messages, or any text meant to simulate real people talking. Preserve facts, timestamps, speakers, channels, and required structure, but rewrite the message content so it sounds like believable Chinese conversation.

Do:

• Prefer common spoken verb-object phrasing: 我看了一眼A组, 把那句话删掉, 等他把新版文件发来.
• Keep some messages short or partial when context makes it clear: 我再问下, 先别发, 刚看到, 那我改这处.
• Split overloaded messages. One person can mention the file; another can ask if it is latest; the first person can answer.
• Let speakers sound different. A coordinator, a teammate, and a person only passing by should not all write like the same careful assistant.
• Use natural uncertainty: 好像, 我不确定, 我这边看到的是, 刚问了下, 先按这个来.
• Keep useful imperfections, such as follow-up corrections or missed context, when they do not change the meaning.

Avoid:

• Compressed noun piles: 新版包, 最终说法, 版本不动, 名单维持, unless that wording is normal in the specific workplace.
• Office-style filler words such as 说法, 收尾, 闭环, and similar phrases. Use 规则, 标准, 说法, 版本, 要求, 最后确认, 整理完, or 说明清楚 instead when they fit.
• Translation-like or ledger-like phrasing: 我手里有昨晚节目单PDF，但我不保证它是新.
• Mechanical acknowledgements: 收到，版本我不再动, 收到，我删那句, 已确认并同步.
• Every message ending with 。, or every speaker giving complete, perfectly structured information.
• Adding cute filler, slang, emojis, or dialect just to look human.

Chat rewrite examples:

等小章新版包 -> 等小章把新版的字幕包发过来
收到，我删那句 -> 收到，我把那句话给删掉
我把A组看了一眼 -> 我看了一眼A组
我手里有昨晚节目单PDF，但我不保证它是新 -> 我这边有昨晚那版节目单PDF，不过不确定是不是最新的
收到，版本我不再动 -> 收到，那我先不动这版了
名单维持原说法 -> 名单先按刚才那版来

When QC says “many messages are unnatural,” do not only fix the quoted examples. Re-read the whole passage, find the repeated pattern, and repair all similar sentences.

Hard Spacing Rule

For normal prose, remove spaces between Chinese characters and ASCII English letters or digits.

Treat this as a hard rule: Chinese-English and Chinese-number spacing should be removed in normal prose, not left as a style choice.

Apply these transformations:

中文 GPT -> 中文GPT
GPT 中文 -> GPT中文
中文 5.5 -> 中文5.5
5.5 中文 -> 5.5中文
第 3 步 -> 第3步
A/B 测试 -> A/B测试

Equivalent regex idea:

([\u4e00-\u9fff])\s+([A-Za-z0-9]) -> $1$2
([A-Za-z0-9])\s+([\u4e00-\u9fff]) -> $1$2

Do not apply inside code blocks, URLs, file paths, terminal commands, markdown table alignment, or quoted source text where exact spacing matters.

GPT-Style Phrase Replacements

Do not replace mechanically. Choose the shortest wording that fits the sentence.

Model-like phrase	Natural replacement
先说结论	Omit, or state the conclusion directly
短版是 / 一句话说	简单说 / 直接说 / omit
核心是 / 核心判断是	主要是 / 问题在 / 要看
更准确地说	换句话说 / 其实是 / omit
这里要先区分	先分清 / 这得分开看
这里有一个关键区别	关键区别在 / 要分清
本质上是	其实是 / 说到底是
可以把它看成	可以理解成 / 大概就是
从三个维度看	分几块看 / 主要看
判断标准可以设成	可以按这些点判断 / 看这几点
更稳妥的做法是	建议 / 最好 / 可以
我不太建议直接	不建议直接 / 先别急着
这里容易踩坑的是	容易出问题的是 / 注意
可执行版本	改法 / 写法 / 直接可用的版本
最小修改范围	尽量少改 / 只改必要部分
验收标准	检查点 / 最后看
风险点	容易出错的地方 / 问题点
边界条件	前提 / 限制
说法 / 收尾 / 闭环	规则 / 标准 / 说法 / 版本 / 要求 / 最后确认 / 整理完 / 说明清楚
信息密度偏低	内容有点空 / 这句不够实
可判性不足	不好判断 / 标准不清楚
隐含标准显性化	把标准说出来
这不是单纯的X，而是Y	重点不是X，是Y
不仅……而且……	Split into one direct sentence
与此同时	同时 / 另外 / omit
因此真正要检查的是	所以要看 / 重点看
这也是为什么	所以 / 因此 / omit
面向X阅读	给X看 / 给X快速看结论
纳入范围为X	只纳入X / 范围是X
交付1个X	最后交1个X
基于附件无法确认	附件本身看不出
评分公式至少纳入X	评分里要放进X
用于安排/判断/复盘	用来安排/判断/复盘
完成清洗、关联和汇总	做清洗、关联和汇总

Sentence-Level Edits

Prefer:

• Fewer signpost phrases.
• Two-part lists instead of automatic three-part lists.
• Direct verbs over abstract nouns.
• Natural Chinese word order over outline-like compression.
• Specific comments over generic praise or caution.
• Short transitions: "但", "所以", "另外", "这里".
• For work-order text, let background and purpose sound like ordinary assignment language while keeping hard requirements exact.
• Remove comma pauses that only imitate careful narration. If a short Chinese clause can be read straight through, let it read straight through.

Avoid:

• "首先、其次、最后" when the order is not important.
• "从A、B、C三个维度" unless the three dimensions are genuinely needed.
• "不是X，而是Y" repeated more than once.
• "更稳、更准、更可控" style triple adjectives.
• "这版已经能用，但还不够尖" unless intentionally using an editorial voice.
• Chat lines that sound like timeline notes, meeting minutes, or task labels.
• Several stiff task words in a row, such as "面向/基于/纳入/交付/口径/维度/支撑/确保/需/应". Replace some with direct verbs when the requirement stays the same.

Rewrite Examples

Before:

先说结论，GPT 5.5 的核心问题不是能力不足，而是这里需要先区分任务边界和验收标准。

After:

GPT5.5的问题不在能力，而在任务边界和检查标准没说清。

Before:

更稳妥的做法是把中文和 English 之间、中文和 3 个数字之间的空格统一去除，从而形成一个可执行版本。

After:

建议把中文和English之间、中文和3个数字之间的空格统一去掉，直接写进规则里。

Before:

这个版本已经能用，但还不够尖。这里的风险点在于可判性不足，因此真正要检查的是限制条件有没有说全。

After:

这版能用，但标准还不够清楚。重点看限制条件有没有说全，最后能不能判断对错。

Before:

该中文说明面向运营总监阅读，用于快速了解主要结论，并支撑Q2车辆调度和收费沟通。

After:

这个中文说明是给运营总监快速看结论的，也用来安排Q2车辆调度和收费沟通。

Final Pass Checklist

• Chinese-English spacing removed in prose.
• Chinese-number spacing removed in prose.
• GPT5.5/GPT5.4-style framing phrases removed or softened.
• Meaning-equivalent replacements did not change facts or requirements.
• The final text sounds like a person editing for clarity, not a model explaining its process.
• For chat records, the final text sounds like people actually messaging each other, not a polished script or an execution log.