After Seeing WorkBuddy Enterprise's Skill Management, I Built a Simplified Version for My One-Person Shop

I can't get into the backend of the WorkBuddy Enterprise Skill management, but I went through the docs twice. The two-layer policy framework in there, I broke out a simplified version and brought it back to my own shop. Over this stretch, I've cut rework by about three or four times.

Let's start with the phenomenon. Tencent Cloud's WorkBuddy Enterprise docs are written in great detail. Every entry in the Skill list is tagged with category, type, version number, status, and visibility scope. Admins can upload enterprise Skills and configure distribution policies, whitelists and blacklists, with priority being blacklist > whitelist > default allow. There's also a column called member custom Skills with four options: allow all, whitelist, blacklist, forbid all, which controls whether employees can upload their own Skills.

My first reaction was that this has nothing to do with me. I'm a one-man show—customer service, design, operations, packing all done by myself. Where would members and departments come from?

Later I figured it out. What a solo shop really needs to manage is boundaries. With no one managing you, you just change things on a whim, and after changing you forget, and only next month when a batch ships do you find the table was wrong. The essence of the enterprise version is: first define who can use it, which version they can use, and whose account a change falls under. These three things I also have to face alone.

First, let's clarify what a Skill is. In WorkBuddy, a Skill is something that adds specific tool capabilities to the AI. It packages a process into a set of executable steps, and the AI, under your authorization, works according to these steps—for example, reading Excel files in a folder, merging by field, exporting to a specified location. Unlike telling it something off the cuff in the chat box, a Skill is fixed.

My scenario is monthly reconciliation. Shopify backend order exports, Xianyu chat log exports, and main images from Gemini Nano Banana and Flux Art all need to be merged into one table. Before, I'd re-explain the requirements each time, and it would understand slightly differently each time—just explaining took over ten minutes.

How I actually packaged it. Open the WorkBuddy client, find Skill in the left sidebar—some versions call it skills—click in and you'll see a list sorted by category, with a + Upload Skill in the top right. Click it, fill in basic info: name, description, then drag your local Skill package in, .zip supported. Below that is Permission Config and Visibility Scope. I'm just one person, so I set visibility to myself. After confirming, click Save and Enable. I named it myself: Monthly Reconciliation.

The key is the description field. I always write it in three parts: what goes in, what it does, which folder it outputs to. For example, "Input xlsx from the To Sort folder, deduplicate and merge by order number, output to Reconciliation_Done." Once it's hardcoded, next time I call it I don't need to explain—those ten-plus minutes saved are pure profit.

Three policies I copied over. First, for the material categorization Skill, I set it to read-only. It can only drop results into To Sort, it's not allowed to touch my original images. That's the whitelist idea, except the list only has me. Second, any Skill whose process has been changed must have a date appended to its name, Monthly Reconciliation_0930. Old ones aren't deleted, kept for comparison. That's copying the version number idea. Third, cross-folder deletion actions are always manual, no automation.

One pitfall to mention. WorkBuddy is genuinely good with plain text and tables, but when it comes to categorizing by image content, it falls short. Once I had it sort main images by color, and it ended up sorting by the red blue prefixes in filenames. A few images that had their colors changed but not renamed were all misplaced. Later I split it into two steps: first have it read image attributes into a table, I manually scan through, then have it categorize by the table. That's stable. One thing at a time—I wrote a post about this conclusion in mid-September, and it hasn't changed.

Daily maintenance is just ten minutes every Friday afternoon, fishing out the outputs that went off-track this week to take a look—was it the description that wasn't clear, or the Skill itself that has a problem. Most of the time it's the description. Changing the description is much cheaper than changing the process. Also, don't enable auto-run. All my Skills are manually triggered. I'd rather click one extra time.

The enterprise version's SkillHub and Connector, supposedly covering tens of thousands of skills, I don't need for now. I'll deal with that when I actually hire a second person. My environment is the personal version plus a computer I use myself. It may not apply to everyone, but the thinking should be universal. A solo shop has to draw a line for itself. Once the line is clear, you dare to keep using the tool.