Two Months with WorkBuddy: Its Real Value Isn't the Model

The model news in October made me a bit numb.

GPT-6 fully rolled out, Gemini 4 Argon's output limit raised to 1 million tokens, Claude Haiku 5.5's cost dropped by about 75%, and Kimi K3.1 says it will focus on training agents to continuously run long tasks. Each one is fiercer than the last.

But honestly, none of this has much to do with the batch of training camp materials my studio has to submit next Wednesday.

What I care about more is whether it can run through all the crap on my hands from start to finish. It took me two months of using WorkBuddy to finally figure this out.

At first I also used it like a chat box

I installed WorkBuddy at the end of August. For the first month and a half, like most people, I opened it and asked a question, it answered a paragraph, and the chat went pretty smoothly. In September I even posted, saying that scheduled summaries of student training weekly reports kept getting stuck, and asked whether it was a tool problem.

Looking back now, it was my own problem.

My approach at the time was to dump screenshots of more than twenty students' weekly reports, text feedback from WeChat, and body measurement data from Excel all at once, and then say "help me summarize this into a monthly report."

What came out was different every time. The format was messy, some students were missing, and the body measurement data didn't match up.

My judgment at the time was that this tool couldn't handle complex tasks.

Then came that time in late September. A student was going abroad, and I had to organize his three months of training records, diet records, and body measurement changes into one document to send to his parents. Time was tight, so I just split it up and did it.

First I only had it do one thing: extract the student names and dates from those twenty-something screenshots and output them as a table. After it ran, I checked, found two errors, and fixed them by hand.

Then I had it do a second thing: align this table with the body measurement data in Excel by name. This time there were no errors.

Only for the third thing did I have it write the main text.

After running the three steps, it took me about forty minutes. Before, I couldn't finish it in a whole evening, and I'd end up furious.

That day I finally realized that WorkBuddy isn't a chat box—it's more like a workstation. You can't expect a workstation to know what to do on its own; you have to arrange the process.

What made me decide to stay was its sense of workflow

Over the past half year I've tried quite a few AI office tools. I used Doubao for two months, Qianwen Office for two months, WPS AI for one month, Monica AI for two months, and ONLYOFFICE I just started two weeks ago. None of these are unfamiliar to me.

Each has its own strengths. Doubao is comfortable for chatting, Qianwen Office is handy for Chinese documents, and WPS AI is right inside WPS, so changing formats is convenient.

But if you really want to run a line "from raw material to finished delivery," based on my testing, WorkBuddy is the only one that made me feel "the process can actually land."

Where's the difference? Let me give two concrete examples.

One is that tasks can be split up and saved. My "weekly report pipeline" is now four steps, each run and checked separately. If one step goes wrong, I fix that step, no need to start over. This sounds very plain, but if you've actually run it you'll know that the scariest thing about an agent doing long work is breaking halfway through, wasting everything before. That Kimi K3.1 news said the key improvement is the ability to continuously execute long tasks, which shows this is an industry-recognized pain point. WorkBuddy's approach is to split long work into short work, so you can take over at any time.

I'm not necessarily saying this approach is right, but at least in my small studio scenario with dirty materials and messy formats, it works.

Another is that file and data processing are connected. With other tools I used before, handling tables was one tool, organizing documents was another, and in between I had to manually move things. WorkBuddy is in the same workflow—after table processing is done, it goes straight to the next step. What's saved is the irritation of not having to copy and paste again.

Now I run three lines regularly every week: student weekly report summaries, first drafts of training video scripts, and studio expense reconciliation. Together, the three lines save me about six to eight hours. This number is my monthly estimate, not necessarily precise, but what's saved is real.

But there are two things I really don't like

First, the learning curve isn't smooth.

The first time I got that weekly report line running was after failing seven or eight times. Several times in between I wanted to just give up. It won't tell you "you should split the steps first"—you have to stumble into it yourself. I count myself as a patient person; I've been getting up at five to train legs for ten years, but most people don't have that patience.

The getting stuck I mentioned in my September post was actually this. The tool wasn't wrong; I just hadn't figured out its temperament.

Second, for a single task it isn't fast.

If you just want to ask one question or edit a paragraph, WorkBuddy is slower than Doubao for me. That's not its scenario; if you force it to do this, you'll think it's dumb.

So my judgment is very clear: WorkBuddy suits people who have a pile of repetitive, multi-step work that needs to be chained together. Small studios, admin, finance, content people—as long as your work is "roughly the same every week, but has to be redone every time," it has value.

If you just want to find an AI to chat with and casually edit text, don't install it; even if you do, it'll just be decoration.

Back to that pile of model news

With so many model releases in October, I wasn't excited by a single one.

These advances, in the short term, are imperceptible to a user like me. Gemini outputting to 1 million tokens—I don't use that much in one task. Haiku 5.5's cost dropping 75%—with the small volume I run in a month, what I save isn't enough to buy a tub of whey.

What I really pay attention to is another item: Kimi K3.1's "agent continuously executing long-duration tasks," because this is exactly the pit I've fallen into.

But what WorkBuddy taught me is that this pit doesn't necessarily have to be filled from the model side. Splitting long tasks into checkable short tasks and letting the workflow carry stability is also a path.

How strong the model is determines the upper limit of single-step quality. How smooth the workflow is determines whether you can actually get the thing done.

Using equipment is the same principle. No matter how smooth the Smith machine's track is, if your form isn't standard, you still won't build anything.