Draft a multilingual message while keeping ambiguous terms visible
"Check the charge before the release" is clear to the person who wrote it. Translated without context, a billing reminder can arrive as advice about a battery and a shipment, in perfectly fluent French.
What you'll make
a draft of your message in another language, a short term map showing how each tricky word was translated, and a list of questions for anything still unclear.
What you'll need
the message, who it’s for and what they should do, a few lines of glossary for product names and technical terms, and a chat your organisation lets you use for the content: ChatGPT, Claude, Microsoft 365 Copilot Chat or Gemini. Paste the glossary in with the prompt. For external or contractual messages, plan for a fluent colleague or qualified reviewer to check the final text. Allow about 15 minutes.
Example result.
| Source term | Meaning confirmed | French |
|---|---|---|
| charge | fee on the invoice | frais |
| release | software release | mise en production |
| 14 November | unchanged | 14 novembre |
Merci de vérifier les frais avant la mise en production du 14 novembre.
1. Give the context and glossary with the message
Say who will read the message and what they should do. Add a few glossary lines for product names and any word with a specific meaning in your work. Then ask for the draft, with ambiguous words flagged rather than guessed:
Translate this message into [language] for [reader and purpose].
Use this glossary: [glossary]. Keep names, figures, dates and conditions
exactly as they are. Before translating, list any word that could mean
more than one thing here. If the context doesn't settle it, ask me a
short question or leave a labelled placeholder instead of choosing.
Give me the draft, a term map (source term, meaning used, translation)
and any questions. Don't make a request stronger or softer than the
original.[language] is the target language. [reader and purpose] is who it’s for and why, such as “our French billing partner, who needs to act before go-live”. [glossary] is a few lines like “release = software release to production”.
Why this works: a model translating without context picks the likeliest meaning and writes it fluently. Asking for the list first makes the choices visible before they’re buried in good grammar.
What you should see: a short list of ambiguous words, either a question or a placeholder for each one the context doesn’t settle, and a term map.
2. Check meaning and commitment, not just fluency
Read the term map against what you meant. Then compare the draft with the source for names, numbers, dates, conditions and the strength of every request. Does “may” become “will”? Does “please review” become “please approve”? Does a software release become a shipment?
If you can’t read the target language well, ask for an explanation of the choices:
For each term in the map, explain in plain English why you chose that
translation and what other meanings it could have had. List anywhere
the draft is more or less certain, or asks for more or less, than the
original. Don't back-translate the draft as proof that it's correct.Why this works: commitment words carry the risk in a work message. A back-translation by the same model tends to undo its own choices, so it can make a wrong draft look right.
What you should see: a short explanation for each term, and either “no change in commitment” or a specific sentence to fix.
3. Answer the questions, then redraft
Answer each question in the same chat, or fix the source message if the ambiguity was yours. Ask for a new draft and term map. If the source itself is unclear, asking for a smoother translation only hides the problem.
Why this works: one confirmed meaning per term turns the placeholders into words the reviewer can check against the glossary.
What you should see: a draft with no placeholders and a term map where every row has a confirmed meaning.
Check the result
- Every row in the term map matches what you meant.
- Names, figures, dates and conditions are identical to the source.
- No request has become stronger, weaker or a different action.
- Anything still unresolved is a visible question, not a fluent guess.
The usual failure: the draft quietly picks one meaning of an ambiguous word and reads perfectly. Spot it in the term map: a row whose “meaning used” you never supplied is a guess.
The draft translated [term] as [chosen meaning], but I meant
[confirmed meaning]. Correct that term and every sentence it affects,
keeping names, figures, dates and conditions unchanged. If anything
else is still unclear, leave it as a question. Return the new draft
and the updated term map.Worked example: “check the charge before the release”
Fictional example. A product team writes to its billing partner in Lyon:
Source message: Please check the charge before the release on 14 November.
Glossary: none supplied
Reader: billing partner's operations team, French-speakingThe first draft, with no glossary, came back fluent and confident:
Weak: “Veuillez vérifier la charge de la batterie avant l’expédition du 14 novembre.” (Please check the battery charge before the 14 November shipment.)
With the step 1 prompt, the model flagged both words instead: “charge” could be a fee or a battery level, and “release” could be a software release, a publication or a dispatch. It returned “Merci de vérifier [CHARGE] avant [RELEASE] du 14 novembre” and two questions. Answering “fee on the invoice” and “software release” gave the draft in the example result.
- The failure was invisible in French. The weak draft is grammatical and specific; only the term map would have shown that “battery” and “shipment” came from nowhere.
- The date never moved. 14 November stayed 14 novembre in every version, which is the check that the facts survived.
- Two lines of glossary fix next time. “charge = fee on the invoice” and “release = software release to production” stop the same guess on the next message.
Task kit for this job: Download (ZIP) · Reusable instructions
Go deeper
Keep a reusable glossary for repeat terms
If you write to the same partner often, keep the glossary as a short file and paste it in each time, with each term’s meaning, not just a word pair. For high-volume work, a dedicated translator’s glossary feature keeps terms consistent: DeepL offers glossaries on its paid plans. The searchable procedure guide shows how to turn one into a searchable page without changing its definitions.
Translate diagram labels without changing the logic
A flowchart or decision tree can keep its branches while its labels change language. Keep the conditions and numbers fixed and add a bilingual key. The bilingual diagram key guide covers it.
Keep the reviewed version tied to its approval
When a reviewer checks the final text, note which version they checked. If the message changes afterwards, even by one term, it needs another look; a later AI rewrite doesn’t inherit the earlier sign-off.