Two Months With WorkBuddy: I Deleted the "Full-Process Automation" Idea

Let me give the conclusion first. The most valuable thing about WorkBuddy is that every little task it does can be run on its own, verified on its own, and redone on its own. That's way more useful than how much it can finish in one go. I used it for two months, and in the end I tuned it from an imagined end-to-end butler into a worker responsible for just a single step on the assembly line—and overall efficiency actually more than doubled.

This shift was forced on me by the tool; I didn't figure it out myself at first.

The other day I scrolled through the recap of Tencent TDS's batch of tech events—Workshop Agent talking about generating cross-platform Kuikly App projects directly from natural language, Lingji SmartRun talking about AI executing tasks on real devices, and the whole testing-agent-driven quality engineering thing. After watching, my DNA definitely twitched a bit—that narrative is end-to-end takeover: one sentence in, a runnable project out. My first reaction was envy; my second was that this is a completely different species from what I do every day.

On my side I've got three WeChat official accounts plus one video account. I write during the day and chase anime at night, and the most time-consuming part of each day is the shuffling before writing—the writing itself isn't that draining. Drafts, screenshot assets, titles, publish times, publish channels—all have to be manually copied into one master table. Copy one number wrong and you have to go dig through chat logs until you question your existence. There's no skill in this, but it's rigid: skip it one day and the next day is chaos.

When I first started using WorkBuddy two months ago, my idea was very simple: state all the requirements at once and let it do everything in one shot. Create a big folder, split subdirectories by date, and have it extract all the draft titles, first paragraphs, images, and publish times, and merge them into one master table.

Ran it, looked at the result—all the dates were scrambled.

I even wrote a post on October 16th specifically to complain about this. It doesn't read by folder name at all; it goes into the body text to find dates. If the body says "a certain event," it dutifully records that date. The folders I created by date might as well not exist to it.

After that I changed my approach: only let it do one thing at a time.

Step one, only extract file names and titles, nothing else. Step two, only grab summaries from the first paragraph of the body, and I scan through them myself afterward. Step three, merge the results of the first two steps with publish times into a table. I keep the three steps disconnected—each can be rerun independently, and a mistake doesn't drag down the whole thing.

The effect was better than I expected. Before, manually merging the asset tables for four accounts took half an afternoon; after splitting into three steps, by my own measurement it's about twelve minutes, and I can go make a cup of coffee in between. Categorizing topic assets used to take forty-some minutes, now it's about ten. I don't calculate things like "saves 90% of time"—it's just genuinely one less afternoon of staying up late.

And there was more than one pitfall. Once, to save time, I mounted several folders and ran them at the same time; the third task was halfway through when the client popped up a box saying something like Local Agent Error, too many requests, and it was stuck for almost an hour. I was rushing a draft at the time and my mindset just collapsed. Later I learned my lesson: one at a time, queue up if you have to—slower, sure, but no mishaps. This made me completely give up any fantasy of concurrent scheduling with it.

So my judgment on this kind of tool is now very clear: its ceiling, in the end, is your ability to break down the work—model capability is secondary. The finer you slice the work, the more stable it is; the more you want it to do it in one step, the more it starts improvising, and the cost of its improvisation is you redoing it all over again.

Comparing with Tencent's stuff makes it clearer. TDS's testing agents and cross-platform generation—their fault tolerance is propped up by automated testing and regression as a safety net; if it runs wrong, there's a pipeline to catch it. On my side there's no testing—I am the testing. So my workflow has to be designed so that every step can be visually accepted by me. That's the constraint of my scenario; WorkBuddy itself doesn't have any real shortcomings. Forcing an engineering team's end-to-end thinking onto content operations is just asking for trouble.

I also don't love some of its design choices. For example, by default it tries to guess what you want to do, and it guesses quite eagerly—sometimes I just want it to faithfully read out a column of numbers, and it goes ahead and adds formatting and subheadings, and I have to manually delete them. This kind of thoughtfulness might be friendly to beginners, but for someone like me who's already broken down the process clearly, it's interference. Now I write my instructions extremely rigidly—if I can avoid giving it room to improvise, I do.

So who is it for? For people with a pile of repetitive, structured chores that can be broken into small chunks—like me, an operator shuffling assets every day, or people who need to make daily/weekly reports, merge spreadsheets, or organize folders. Who is it not for? Not for people who want to generate a whole plan in one sentence and deliver it directly—you'll be disappointed, and you'll most likely blame the tool. I've seen too many people jump straight to wanting full-process automation, run it for two days, find it doesn't work, and say this thing is no good—when really they just didn't break it down themselves.

After two months, my usage has basically settled. It's now just one workstation on my assembly line—hardly a butler.

I'm responsible for breaking down the work and accepting it; it's responsible for running the repetitive parts. Once this division of labor is set, the collaboration flows.

The term "full-process automation"—let's leave that for the launch event slides.