+ Covers every category and supplies a usable final line.
- Quotation translation remains optional instead of decided.
Keeping quotations and proper nouns in the original is often correct, and number conventions get forgotten.
| Category | Using AI › Context management |
|---|---|
| Tags | ReviewingRewritingAnalyzing |
The output keeps mixing languages. **Diagnose and fix it.** 1. **Find where it switches**, and which of these it is: - **My input was mixed**, so the output followed - **A technical term has no natural equivalent**, and the model switched rather than glossing - **Quoted material** is in the other language. ***This one may be correct*** — quotes usually should not be translated - **The instruction is in one language and the content in another** - It drifted back to the language the conversation started in 2. **Decide the rule for each category and state it:** - **Technical terms** — translate, keep, or keep with a gloss on first use. **Pick one** - **Proper nouns, product names, code** — ***normally kept.*** **Say so explicitly or they get translated** - **Quotations** — kept, with a translation alongside or not - **Numbers, units, dates** — ***which convention.*** *This is the one people forget and it is the most visible* 3. **Write the instruction line** that enforces this, to paste at the end of the prompt. ***At the end, not the start*** — a language rule given early is the first thing to drift ⚠️ **Where keeping the original is correct, say so.** ***Forcing everything into one language is a different error, not a fix.***
Languages mix for five different reasons, and in some of them keeping the original is right — forcing everything into one language is a different error. This sets a rule per category and puts the directive last.
ChatGPT is the most complete and concise. Gemini is specific but has an omission and speculation; [C] is absent.
+ Covers every category and supplies a usable final line.
- Quotation translation remains optional instead of decided.
+ Pinpoints the switch and provides a concrete revision.
- The final rule omits quotations and overstates the cause.
| Criterion | ChatGPT | Gemini | Leader |
|---|---|---|---|
| Instruction following | 9 | 7 | ChatGPT +29% |
| Accuracy | 9 | 7 | ChatGPT +29% |
| Specificity | 9 | 9 | Tie |
| Structure | 9 | 9 | Tie |
| Right length | 9 | 7 | ChatGPT +29% |
Scored 1–10 by gpt-5.6-sol with model names hidden (2026-09-25). This is an AI review, not a measurement.
We gave three models the same input and copied their answers unedited. Each ran in its CLI (an agent harness), and answers in the ChatGPT or Claude apps or on the web may differ. Outputs are in Korean.
Target language: 한국어 The output: 저희 서비스의 latency는 평균 120ms이며, 이는 caching layer 도입 이후 개선된 수치입니다. Overall, the system performance has improved significantly since last quarter. 다음 분기에는 throughput을 늘리는 방향으로 scale-out을 검토 중입니다.
원하는 언어는 한국어입니다. 따라서 기준 언어는 한국어로 잡아야 합니다.
진단:
latency, caching layer, throughput, scale-outOverall, the system performance has improved significantly since last quarter.적용할 규칙:
120 ms처럼 쓰고, 날짜는 한국식 2026년 9월 24일을 사용합니다.수정 예시:
저희 서비스의 지연 시간(latency)은 평균 120 ms이며, 이는 캐싱 계층(caching layer) 도입 이후 개선된 수치입니다. 지난 분기 이후 시스템 성능이 크게 향상되었습니다. 다음 분기에는 처리량(throughput)을 늘리는 방향으로 수평 확장(scale-out)을 검토 중입니다.
프롬프트 맨 끝에 붙일 instruction line:
최종 답변은 한국어로만 작성하세요. 기술 용어는 한국어로 번역하고 첫 사용 때만 영어를 괄호로 병기하세요. 고유명사·제품명·코드는 원문을 유지하고, 인용문도 원문을 유지하되 필요할 때만 한국어 번역을 함께 제공하세요. 숫자와 단위는
120 ms형식을, 날짜는 한국식YYYY년 M월 D일형식을 사용하세요.**
| Chat Summary and Export Instructions | |
| Context Migration | |
| Continue and Recap Assistant | |
| Plan how to split long material for a model | |
| Find instructions that contradict each other |