Structure · Aesthetary Original
Localize with explicit uncertainty
You'll be able to translate a site with AI and show exactly which lines still need a fluent speaker's check.
In this skill
Machine translation is fine. Unmarked lines are the risk.
Translate slot by slot
- Start from the approved manuscript
- Keep each slot's name
- Glossary: what stays as it is
- Mark each line's status as you go
A translation you can't read is a draft until someone fluent signs it off.
1 more scene is in the video and the transcript.
Transcript
AI will translate a whole site in seconds, and the result will look fluent in a language you can't read. That's the danger: from the outside, a good translation and a wrong one look exactly the same.
So use it, and expect some lines to be wrong. A wrong price, a formal 'you' where the brand is casual, a customer's quote quietly rewritten: none of it looks wrong to someone who doesn't speak the language. Your job is to translate, then show exactly which lines you're unsure about, so the person who can read them knows where to look.
Translate from the finished manuscript, never from the built page, and keep the slot names so every translated line sits next to its source. The bakery launches in English: the owner settled on one language on the ninth of October. Say she adds Albanian later, for neighbors who speak it at home. Give the AI the content model, the manuscript and a glossary of what stays as it is, like the bakery's name and the loaf names. Every line gets a status as it's produced: ok, check or hold. The customer glossary doesn't carry over, because it came from English reviews. What Albanian speakers call the sourdough is a question for a fluent reader, so those lines start at check.
Some lines need marking more than others. Numbers first. Prices, hours and phone numbers should come through exactly, though a time or number format may need to change for the reader: here, six p.m. became eighteen hundred. Many languages make you choose between a formal and an informal 'you', and the AI will pick one without telling you. Translating a customer's review changes their words, so show the original beside it, or ask. Anything legal or about safety is held for a fluent check, every time: on a bakery's site, that's the allergens. And translations rarely come out the same length, so check each line still fits its slot.
This is the sheet that goes with the translation: every slot, a status, and a reason someone can act on. Check means I've found something I can't settle myself, like a button that translates back as 'order forward' instead of 'order ahead'. Hold means it doesn't publish until someone fluent has signed it off. A reviewer who gets this sheet goes straight to the lines that matter, instead of reading the whole site and hoping to notice.
You can verify more than you'd think without speaking the language. Check every number against the source, and check the glossary terms survived. Translate it back with a different translation service and compare the meaning. That catches drift, though not whether a line sounds natural. And make sure the page declares its language in the code, lang equals s q for Albanian, so screen readers pronounce it. W3C's guidance on declaring language in HTML shows how. What you can't verify is whether it reads like a person wrote it, whether the tone is right, and whether the legal lines say what they should.
So someone fluent checks it before it's published for real: the client or a native speaker, planned for from the start. You don't need to speak the language to do this work, but you do need that person. A review build can carry the marked draft, so they see it in place. A published page can't. Before launch the sheet has no check or hold left on it. If a line never gets signed off, it comes out, or that page stays in the original language until it is.
A translation you can't read is a draft until someone fluent signs it off. Your job is to show them exactly where to look.