Using WorkBuddy on Customer PDFs: Lock Down Your Output Fields First

I came across that Kimi Docs article introducing the document Agent, saying you describe the task, generate, preview and download, and it can output a ten-thousand-word long document or hundreds of payslips in one go. My first reaction after reading was thinking of that batch of client PDFs on hand. In translation, being a bit slow doesn't matter, but something that comes out fast and can't be checked sentence by sentence is the real trouble.

Let me give the conclusion first. When using WorkBuddy to process documents, lock down the output fields first, then talk about automation. If you can't lock down the fields, the more diligent it is, the more tired you get.

Environment first: I'm a freelance translator, clients are mostly small companies, a batch of files is usually a dozen to a few dozen PDFs, mixed Chinese and English. I've used WorkBuddy for about a month, early on mainly to break down tasks, and only started running the full document process in the last three weeks. The steps below may not apply to everyone.

A job I took last Wednesday works as an example. The client gave a zip file with 18 PDF inquiry sheets, plus two Excel quote tables, to be organized into a Chinese comparison list, then a formal email reply. My old approach was to manually open the PDFs, copy, paste into Excel, then translate — a whole evening. This time I went through WorkBuddy the whole way.

Open WorkBuddy, click workspace in the left sidebar, create a new one, name it client name plus date. Don't skip this step. I learned the hard way — all clients' files piled in the default space, and when it retrieved, it brought client A's quote into client B's email. I wrote about workspace isolation before, and it's still my judgment now.

After the workspace is created, drag the 18 PDFs into the materials area. It auto-parses, and after parsing lists the file inventory on the right. Two scanned files had recognition errors, needed manual box-selection and re-run, this step is similar to iFlytek OCR's manual correction.

Then the key part. Click new task, choose document processing, and the interface gives an output field table, default is all-on, anything can be written in. I deleted all the unused fields, keeping only four: file number, original summary, Chinese translation name, notes. The notes column I left empty, forbidding it from filling in itself.

Further down there's a switch, something like "only use uploaded materials," turn it on. Next to it there's "allow online supplementation," turn it off. If you don't turn these off, it'll add drama for you. I once had it write a Chinese email, and it took the liberty of adding "looking forward to long-term cooperation," the client never said it, I never said it either. Since then I treat these switches as hard requirements.

The terminology base I only brought in two weeks ago. In translation, inconsistent terminology is more trouble than mistranslating. Export the client's fixed product names and company names as CSV, attach to the workspace's terminology base, check it when running tasks. It'll prioritize the terminology base value in the translation name column, and only translate itself if it can't find one. I suggest trying three to five files first, confirming the terms hit, then going full-scale.

Set the permissions while you're at it. If several people share an account, the workspace can set read-only and editable roles. On my end it's just me, but I set client materials as read-only mount, to avoid accidentally changing the original files.

The 18 files took about twenty-some minutes to run. In between it lays out each step in the execution log — which file was read, what was filled in each column, all clickable to view. This step is what I value most. That Kimi article talks about result-orientation, describe and download directly, I don't deny it's fast, but for work like translation and proofreading, if the intermediate process is invisible you can't sign off on it.

For export format I chose CSV, then converted to Excel myself. Exporting Excel directly is fine too, but on my end I need to run through formulas, CSV is cleaner.

Three pitfalls to avoid, all ones I hit myself. Only have it do one thing at a time, don't cram organizing the list and writing the reply into the same task, cram it in and it gets muddled. Delete fields where you can, delete until only what you want is left, with no room to play it can't add drama. Before running a batch, run one first, check the intermediate log, and if it's right then go full-scale.

From my use, only a tool that lets me check sentence by sentence and roll back anytime is one I dare take client work with. Being fast at generating — that's not hard this year.