WorkBuddy document translation: lock the glossary first, then the fields, everything else later

Didn't take any orders today. Came across Weglot's piece comparing the best AI translation tools of 2026, which talks about adding multilingual support to websites — first running a layer of AI translation, then fine-tuning with Visual Editor or a glossary, the model learns brand guidelines, glossaries, and historical edit records, the more content the more consistent.

The AI translation model learns your brand guidelines, glossary, and historical edit records, so as content scale grows, translations always maintain brand consistency.

My first reaction after reading wasn't websites. It was the batch of offline documents on hand.

That website setup works because there's a CMS underneath holding it up, the glossary is live, change once and it takes effect site-wide. Offline documents don't have this. A client sends a manual, the glossary is on page two of the attachment, changing one term means back-and-forth emails. I started moving this work onto WorkBuddy two months ago, and now it's running smoothly. Let me explain how I set it up.

My environment is a one-person shop, clients are mostly European language pairs, the setup below may not apply to everyone.

If it's your first time, open WorkBuddy, click "New Task" in the left taskbar, choose the document processing category. Before uploading, don't rush to upload — first go to the settings in the top right and lock down the output fields. This step is the easiest to skip, and the most critical. There's a field called "Allow model to write content not provided," default is on. First time I didn't notice, it added a line at the end of the translation: "Looking forward to long-term cooperation." The client never said it, I never said it either. Turn it off. On the same page there's "Output format," choose plain text or docx, don't choose its own Markdown, the heading hierarchy gets messed up on export.

After locking the fields, handle terminology. I only started using the glossary properly three weeks ago, before that I relied on memory. The method is to create a glossary file in WorkBuddy, CSV format, two columns, source on the left, translation on the right. CSV is a comma-separated table, Excel save-as produces it. After creating it, go back to task settings, click "Glossary," choose "Prioritize over model judgment." That line isn't long, but it determines whether it looks up the table or guesses. The product names and regulation numbers in client manuals all depend on this one line.

I'm even more aggressive with permissions. In my WorkBuddy there are two folders, one called To Process, one called Verified. The task only gets read-write permission on "To Process," "Verified" is read-only. What it translates lands in To Process, I go through it, and manually move the fine ones over. This move is dumb, but useful. It can't touch already-finalized things, and I won't one morning find a submitted draft has been changed.

Then pitfalls. The biggest pitfall is saying too much at once.

Early on I always wanted to explain formatting, terminology, tone, and word count all clearly in one task, and it did a bit of each, and got each wrong. Later I learned — one thing at a time. First just translate, no formatting. After translating I review, if terminology is wrong I fix the glossary and rerun, if tone is off I send a separate revision instruction. Slower, but the checking cost is much lower. In this line slower doesn't matter; what's really troublesome is fast output you can't check sentence by sentence.

The effect isn't stunning, it's steady. An eight-page product manual, I used to compare source and translation back and forth, one evening. Now with terminology and fields locked, I can get through it in about forty minutes, focusing only on sentence structure and the places it's unsure about. The rest of the time is mine.

By the way, client-returned revisions also go through this setup. The logic of rejecting item by item in revision mode, I wrote about it at the end of September, after moving into WorkBuddy it's the same process, just the rejection target changed from the client to it.

Looking ahead, I guess the glossary will move from the client's hands into the workflow. Right now it's still the client sending attachments and me importing manually; next year it'll probably be the tool pulling from the client's library itself, and writing changes back after translating. Weglot is already doing this, just for websites. For offline documents, whoever first connects the glossary into the workflow wins half the battle.