Draft the client reply in their language, and give me an English mirror so I can verify it says
what I meant.
1. Detect the client's language from their messages in the thread — the language they actually
write in, not their country or name. If the thread mixes languages, match their most recent
substantive message; if genuinely ambiguous, ask me which language.
2. Draft the reply NATIVELY in that language — composed in it, not word-for-word translated
from an English draft — applying full house standards: answer first, no filler, greeting
only on a thread's first message, defensive writing, and locally natural register (match
their formal/informal address — vous/usted/Sie level).
3. Keep technical anchors unambiguous: product names, error messages, and ticket numbers stay
as-is; error text the client must recognize on screen is quoted in its original language.
4. Provide the English back-translation directly below the draft, labeled "English
back-translation (for your review)" — a faithful mirror of what the foreign-language draft
actually says, including anything that shifted in translation.
5. Show me both together (draft on top, back-translation below for removal before send) for
review. I confirm before anything is sent.
The back-translation is mandatory whenever I don't read the target language — never ask me to
approve text I can't verify, and remind me to remove it before sending. The two versions must
match: never add nuance, promises, or softening in one that isn't in the other; if a concept
doesn't translate cleanly, flag it ("<term> has no direct equivalent — I used X, which means
Y"). All standard reply guardrails apply across languages: no invented details, no jargon,
defensive writing intact, reopen line on close-outs (translated naturally). If the language
uses honorifics the thread doesn't reveal, default to formal.
