Multilingual prompts without muddy mix
Goal: Lock instruction vs draft languages; keep proper nouns intact—no Chinese–English salad in one sentence.
Best for: Teams that research in one language and ship in another; HK/bilingual workers switching mid-day.
Models will happily mix languages mid-sentence if you do not declare the rules. Product names, error codes, and UI strings often must stay in the original form. Say that out loud in the prompt.
Why beginners skip this
Mashing ZH/EN in one sentence feels tidy—the model follows the muddle and both readers tire. Pick a primary language, then mark exceptions.
Step 1 — Declare languages like a contract
Declare the output language at the top like a contract; one language for the whole thread.
Input language: [en / zh / mixed notes]
Output language: [en / zh] — one language only
Do not mix languages inside one sentence.
Keep unchanged (original form): [product names, SKUs, error codes]
If a term has no natural translation, keep English and add a short gloss in parentheses once.
Audience locale: [US / HK / mainland / …]
Step 2 — Chinese notes → English sales bullets
Example
Filled prompt
Input language: Chinese meeting notes
Output language: English bullets for US sales
Keep unchanged: “iswarp”, “Claude Projects”, error code E-204
Do not mix languages in one sentence.
Audience locale: US
Task: Turn notes into 6 bullets: problem → what we ship → proof point → CTA.
Notes:
- 新手不知道 ChatGPT/Claude/Grok 怎么选
- 我们按场景写指南,不堆参数表
- 免费阅读,不用注册
What good output looks like
- Clean English only in the bullets
- Product names untouched
- No “我们 ship 一个 seamless 体验” mashups
- CTA matches what you can actually offer
Step 3 — Two-pass when you need both languages
Do not ask for “both languages in one block” first—sequence it: full language A, then translate the whole block to language B.
Pass A: Output English only (rules above).
Pass B: After I approve, localize to [zh-HK / zh-CN] for [audience]. Keep product names. Prefer natural phrasing over literal calques.
If you need both in one doc, request a labeled section structure: ## EN then ## ZH—never interleaved sentences.
Example
Filled prompt
Live task this week (not a toy): apply this guide's Goal — Control input and output language clearly when you work across English and Chinese—so drafts stay pu
Fill with your real artifact names/files. Constraints: no invented facts; label Unknowns; stop before send/delete/pay.
Return the working prompt + one pass/fail check I can mark in a minute.
What good output looks like
- Goal/Format/Boundaries all present for “Control input and output language clearly when you”
- One example beats stacked IMPORTANT words
- Version/date kept when template changes
- Pass/fail: zero clarifying turns on live task
Do not bookmark-only
Finish Practice now, then tick this guide on your path. Titles unread are not skills.
Reality check
Open this guide beside a live task. If you cannot name the task, pick one before reading further.
Close the loop
Write one line: next time I mash ZH/EN in one sentence with no exception list, I refuse send.
Example
Filled prompt
EN bullets → ZH-HK email. Freeze names/numbers.
Pass1 meaning · Pass2 HK register. No mixed EN leftovers in body except product names.
What good output looks like
- One example beats stacked IMPORTANT words
- Version/date kept when template changes
- Pass/fail: zero clarifying turns on live task
- Goal/Format/Boundaries all present for “Control input and output language clearly when you”
Red flag
Output language unlocked; half-translated salad shipping—lock input/output languages.
Save this card
Card: When switching languages, lock input/output language so drafts ship—not a half-translated salad. Keep this beside the tool; read before you start.
Before you start
Keep the last failing output visible. You will repair the prompt for “Control input and output language clearly when” against that failure—not a fresh toy ask. Guide: prompt-multilingual-without-muddy-mix. Keep the real job visible the whole time; a blank box means stop and fetch the artifact first.
How to use this week
This week set a primary language on a real bilingual draft, list exception terms aside; rewrite every mashed sentence.
Close the week by quoting aloud the refuse line you actually used once. If you cannot, rerun Practice now and date the pass/fail (Asia/Shanghai).
Pitfalls
| Pitfall | Fix |
|---|---|
| “Translate and improve” with no output language | Name input + output every time |
| Letting UI strings get “helpfully” translated | List keep-unchanged tokens |
| One-shot “make it bilingual” | Two passes; approve EN before ZH |
| Mainland tone for HK readers (or reverse) | State locale; see bilingual HK workflow |
Example
Filled prompt
Temptation: skip the habit for prompt-multilingual-without-muddy-mix to save five minutes.
Job still: Control input and output language clearly
Smallest correct pass: Done visible · checkable artifact · Unknowns [?].
Boundaries: draft only; no silent send; no invented metrics.
Return: artifact + refuse-line said aloud.
What good output looks like
- Names the shortcut refused (specific to prompt-multilingual-without-muddy-mix)
- Still delivers checkable output for “Control input and output language clearly”
- Unknowns labeled; draft-only
- Pass/fail: refuse-line dated (prompt-multilingual-without-muddy-mix #4)
Practice now
Glossary three: three locked bilingual terms on Save (Asia/Shanghai).
Output lang tag: first line states output language (Asia/Shanghai).
Lang lock: one system language; mark output language explicitly (Asia/Shanghai).
Multilingual drill: one language for instructions, one for output—label both (Asia/Shanghai).
Thin-spot drill: rewrite the Save card with real nouns from your work (repo, doc, inbox). Generic placeholders fail the drill.
Tonight set a primary language on a real bilingual draft, list exception terms aside; rewrite every mashed sentence.
Buffer line: name the artifact path you used (Asia/Shanghai).
One sentence takeaway
Primary language first; exception terms listed aside; rewrite mashed sentences.
Next
Translate then localize · Bilingual HK workflow · GFB pattern
多语言提示词:别写成中英夹杂沙拉
目标: 说明层与草稿层语言写死,专名原样保留;避免一句里中英泥浆。
适合: 英文说明和中文原文搅在一起、输出读起来很糊的双语工作者。
混搭不是通顺,是读者负担。品牌名、错误码、工单号列进「保持原文」;禁止模型为了通顺悄悄换语言。
为什么新手跳过
一句里中英混写像省事;模型跟混,读者两边都累。先定主语言再标注例外。
步骤 1——像签合同一样声明语言
在提示词顶部写死输出语言,像合同一样;整封线程只用一种。
输入语言:[en / zh / 中英混杂笔记]
输出语言:[en / zh],只用一种
不要在同一句里混用两种语言。
保持原文不改:[产品名、SKU、错误码]
如果某词没有自然译法,保留英文,并在括号里第一次加极短注释。
读者地区:[美国 / 香港 / 内地 / …]
步骤 2——中文笔记 → 英文销售要点
实例
填好的提示词
输入语言:中文会议笔记
输出语言:给美国销售的英文要点
保持原文:iswarp、Claude Projects、错误码 E-204
不要在同一句里中英混杂。
读者地区:美国
任务:把笔记收成 6 条:问题 → 我们交付什么 → 证据点 → 行动号召。
笔记:
- 新手不知道 ChatGPT/Claude/Grok 怎么选
- 我们按场景写指南,不堆参数表
- 免费阅读,不用注册
像样的输出大概长这样
- 要点全是干净英文,一句里不中英夹杂
- iswarp、Claude Projects、E-204 等专名原样保留
- 没有「我们 ship 一个 seamless 体验」式混搭
- 六条结构清楚:问题 → 交付 → 证据 → CTA
- CTA 是你真能兑现的,不画大饼
步骤 3——需要两种语言时用两遍
不要第一次就要求「一块里双语」;串起来:先出完整语言 A,再整段译成语言 B。
第 A 遍:只输出英文(规则同上)。
第 B 遍:我确认后再本地化成 [zh-HK / zh-CN],面向 [读者]。产品名保留。优先自然说法,不要生硬直译。
如果同一文档要两种语言,用分区:## EN 再 ## ZH,不要句句交错。
实例
填好的提示词
我把英文说明和中文原文混在一起,输出很糊。请重写提示词:说明层用英文、草稿层用港式中文,顶部写死输出语言,禁止一句里中英泥浆。
边界:不编造;未知标 [?];发送前停下。
完成检查:一分钟内能判过/不过。
像样的输出大概长这样
- 说明层用英文、草稿层用港式中文,边界写死
- 禁止模型「为了通顺」悄悄换语言
- 混搭问题被点名,并给出可粘贴的重写提示
- 读起来像双语团队的工作约定
- 不靠「你自己注意一点」这种空话收尾
- 未知标出;不编造负责人、日期或指标
不要只收藏
先做完「现在就练一次」,再在学习路径上打勾。只收藏不练,标题变不成技能。
现实核对
把这篇指南放在真实任务旁边。说不出任务,就先挑一件再往下读。
闭环
写下一行:下次一句里中英混写不标例外,我拒绝发出。
红旗
一句里中英混写,模型跟混,读者两边都累;先定主语言再标注例外。
保存这张卡
记下:今天改管用的那条提示词(含边界)+ 版本或日期。
开始前
调试时把上一次失败输出留在眼前。
本周怎么用
本周对一段真双语稿定主语言,例外词单独列表;混句一律拆开重写。
常见坑
| 坑 | 改法 |
|---|---|
| 一句里中英夹杂 | 说明层与草稿层语言写死 |
| 专名被「翻译」掉 | 列表保留原文(品牌、错误码) |
| 靠「你注意一点」 | 禁止悄悄换语言写进边界 |
现在就练一次
今晚就练:对一段真双语稿定主语言,例外词单独列表;混句一律拆开重写。
一句话带走
先定主语言;例外词单独列表;混句拆开重写。
下一步
← Back to home · Learn · Tools · Compare · Start