After Reading That WorkBuddy Field Guide, I Turned "One Chat Box" Into "One Workbench"

Sisters, last night while the little one was asleep I came across a hands-on guide about WorkBuddy, and one line from the author stuck with me. The gist was that a lot of people, the first time they open WorkBuddy, treat it like just another AI chat box, when really it's more like their own workshop. I sat there stunned for a bit after reading it, because that was me. I've used WorkBuddy for three months, and the first two I basically used it as a chat box.

I wrote a post before about having WorkBuddy organize my Taobao side-hustle bookmarks, and it went haywire the moment it ran. At the time I thought the table wasn't built right, and I even went and redid the headers. Looking back now, the problem was I never set up a workbench at all — every time I started from zero and explained what I wanted all over again, so of course it guessed differently each time. Redoing the table however many times wouldn't have helped.

There's a line in that guide that really hit me:

A regular chat product is like going to a convenience store — you push the door open, say what you want, grab it and leave, and next time the guy at the counter doesn't know you again.

A workbench is different. A workbench is a fixed space for a fixed task — you lay out your folders, your usual tools, your processing flow on it, and next time you start work you don't have to set it all up again.

I redid it following that, and let me tell you exactly how I did it. Open WorkBuddy, there's a + New Workbench entry on the home page, click in and name it first. Don't skip this step — I named mine "Product Picking Bench," because later you'll rely on that name to find its context memory again. If the name is vague, even you won't be able to tell which is which. After it's created, the first thing is to bind a folder — click Add Directory and select the folder on my computer called "Side-Hustle Screenshots." Once bound, it knows where my materials are and doesn't have to ask every time. Second thing is to attach Skills — a Skill is a pre-installed little tool. I attached two, document parsing and spreadsheet processing, and that's enough.

Step three is choosing a model, and this is the step most likely to go wrong. There's an article in my materials that puts it bluntly: the tool itself is just a shell, and what determines output quality is the model being called behind it. I started with the default config, and it kept mixing up prices in the screenshots — 3.8 written as 38. Later I switched to another model and reran it, and it was much better. I won't recommend which one — my setup is an old Win laptop with a two-year-old who comes over to mash the keyboard at any moment, so it may not apply to everyone.

I use permissions only lightly, just me on my own. Workbenches can be set to shared, and supposedly in a team you can split who can view and who can edit — I haven't tried it. I only set one thing: made the output directory read-only for myself, to stop it from overwriting the previous version while it's running.

Two more pitfalls I hit. One was greed — I stuffed product picking and copywriting into the same workbench, and it ended up applying the copy's formatting to the table, producing a pile of garbage. Later I learned my lesson: one workbench handles only one thing. The other was file naming — at first I named screenshots casually as 1, 2, 3, and after it ran I couldn't match anything up. Now I stick to the format "platform_date_keyword," same principle as that AI short-drama post I wrote before — only files you can manage count as assets.

For daily upkeep I added two rules: every Sunday night I clear the temp files in the workbench, and after each run I save the results under a fixed name instead of leaving them in the default directory.

I still haven't fully achieved one-workbench-one-task — I keep piling things into one as I go. I'll fix it slowly.