Using AI Assistants Across Korean and English: A Practical Guide
Maker's note: 기준일 2026-10-03. This one is deliberately light and practical. It comes from a team whose project README is written in Korean, whose blog posts are written in English, and whose product family includes traditions that traveled out of Korea.
Working in two languages
Bilingual work is not a hypothetical for us. The AZET README is Korean; this blog is English; Seowoon is a Korean saju product and Tonemoa is personal color analysis, a practice familiar from Korean styling culture that has since traveled well beyond it. Most days, the same question arrives in both languages before lunch.
That makes us heavy users of AI assistants in Korean and English, and after enough of those days, a short list of what works and what bites has settled out. Here it is.
What works well
Reading in either language. Pointing an assistant at Korean text and asking for an English summary — or the reverse — is now a routine, reliable move for everyday material: announcements, product pages, email threads. Treat the output as a summary, not a substitute for the original when the wording matters.
Drafting in the target language directly. Asking for the draft in the language it will be read in usually beats writing it yourself and translating, because the draft never passes through your language's sentence shapes on the way.
Mixed conversations. Asking in one language and expecting the answer to follow the question's language works well enough that most bilingual users stop noticing it. When it doesn't, saying which language you want is enough.
Preferences that persist. Assistants that keep notes
can hold a language rule once, so you stop repeating it. AZET's own
example is
remember "Use concise Korean explanations" in "preferences.md"
— after that, relevant memory is searched before a question whenever
user context would help. One instruction, stored, reused.
A related boundary worth knowing as a user: AZET includes only explicitly requested relevant memory in task context — the whole personal memory store is never sent wholesale to an external classifier. In a bilingual workflow, where notes may mix languages and private details, that restraint matters more than it first appears.
What to watch
| Situation | What to do |
|---|---|
| Fixed command vocabularies | Keep them in one language. AZET's task verbs (open,
click, type, select) are fixed
keywords, not prose to translate |
| Output language | Say which language you want; do not leave it to inference, especially mid-conversation |
| Politeness register | Korean has honorific levels — specify the formality you need, or the draft will pick one for you |
| Facts | Verify in the language and source you will be quoting; a translated summary is still a summary |
| Locale-bound services | Logins, formats, and behavior differ by locale; the Korean and English faces of a service are not guaranteed to match |
| Names, dates, numbers | Check formats across locales before anything important ships |
Two of those rows deserve a sentence more.
The first: when a tool's grammar is fixed keywords, the temptation is to localize the keywords too. Resist it. Keeping the verbs in one language is what keeps parsing unambiguous across users; the local language belongs in the content — what to type into the field, what to remember for next time — not in the machinery.
The second: register. English drafts mostly vary between formal and casual. Korean varies along a steeper ladder, and picking the wrong rung turns a competent draft into an awkward one. If you draft in Korean with an assistant, name the register in the request — it is the single highest-leverage word in the prompt.
A small habit that pays off
Decide your language policy once and write it down. "Answers in English, drafts in Korean, commands in English" is a complete policy, the kind of thing you store in a persistent preference and stop thinking about. The alternative is re-deciding it in every conversation, which means occasionally getting an English answer to a Korean question at the worst moment.
Language handling in AI products keeps improving, but the person who states their language and register explicitly gets the good outcome today, without waiting for the model to guess better. Say it once, remember it, and most of the bilingual friction is gone.
If you want to see how this shows up in what we're building, the waitlist at https://azet.io is open.
FAQ
Should I translate AZET's task verbs into Korean?
No. The task verbs — open, click,
type, select — are fixed keywords, not prose.
Keep the machinery in one language and let Korean live in the content:
what to type into the field, what to remember for next time.
Do I need to specify politeness level in Korean drafts? Yes. Korean varies along a steeper formality ladder than English, and if you don't name the register the draft will pick one for you. Naming it is the single highest-leverage word in the prompt.
Does the assistant send my whole memory store to external services? No. AZET includes only explicitly requested relevant memory in task context — the personal memory store is never sent wholesale to an external classifier. In bilingual workflows, where notes mix languages and private details, that restraint matters more than it first appears.
Is a translated summary good enough for quoting facts? No. A translated summary is still a summary. Verify dates, numbers, and quotes in the language and source you will actually be citing.