I Used WorkBuddy to Compare Google's VLA and Reddit's New Feature: On Dumping Batch Work into Folders
Community Discussion · Humanoid Robots

I Used WorkBuddy to Compare Google's VLA and Reddit's New Feature: On Dumping Batch Work into Folders

Galaxy BrothersGalaxy Brothers3d ago2026/11/16 208 views

Late at night I saw two pieces of news. One was that Google DeepMind released Gemini Robotics 2.5, saying that in an evaluation of about 200 tasks, pastry-picking success went from 5% to 100%, and untying ribbon bows went from 0% to 70%. The other was that Reddit is testing AI-generated summaries of product reviews, able to analyze the pros and cons of a product, and it reportedly raised user concerns.

My first reaction was that these two things are essentially the same action: throw in scattered material and have it produce a readable result. The only difference is Google throws in robot arm motions, Reddit throws in comment sections. And I throw in planning documents.

So the next day I just did a controlled experiment with WorkBuddy—using it to organize a settings table on one hand, and on the other thinking about how it differs from the logic in those two news items. Let me put the conclusion first: its biggest value is that it can work in bulk, can set rules, and can be checked afterward. These three things matter far more than how smart it is.

How I set it up specifically

My environment is Win 11, WorkBuddy desktop, I won't report the version number, anyway it auto-updated once last week. May not apply to everyone, especially if your team needs multi-person collaboration.

When creating a new task, don't go straight into the chat box and type. First click "Workspace" on the left, choose "New Workspace," and give it a folder. This step is crucial—it's like drawing a plot of land for the AI, and it only works within that plot. I created PlanningDoc_SettingsArchive, then dragged in over a hundred docx files of settings, a few xlsx files, and a few handwritten screenshots.

Next click "Output Contract," which is the column I value most in WorkBuddy. In plain terms, it's writing clearly in advance what format, what fields, what naming rules you want. I filled in: output one xlsx, columns are ID, Name, Category, First Appearance Chapter, Related Entries, Notes. Leave the Notes column empty, tell it not to improvise.

This step is where Reddit's feature can't compare. AI summarizes reviews, but you can't control what the summary looks like. But if you write the contract rigidly, it can't escape that frame.

Then comes the place most prone to failure: permissions. There's a "Collaborate" button in the top right of the workspace; click in and you can set three roles: can view, can edit, can manage. I set myself as can manage, and the lead designer as can view. Why set this? Because I've been burned.

The first time I ran it I didn't set permissions, and just threw the workspace link into the group chat. A colleague then casually changed a field name, changing "Category" to "Classification," and the next round the whole table came out empty. I stared at the screen for about ten minutes, thinking the model had glitched.

Later I understood: WorkBuddy recognizes field names, and if you change the name it won't match the definition in the contract. This is actually the same principle as Google's VLA model—the success rate in the evaluation reached 100% because the task was defined rigidly. Loosen the definition and the success rate drops immediately. So my habit now is: the first thing after creating a workspace is to lock permissions, leaving only myself and admins able to change the structure.

Three things I do regularly for maintenance

First, after each run, check the "Run Log," just scroll down on the task detail page. It lists which files were read successfully and which were skipped. Once it skipped all three screenshots, and the log said "image resolution too low"; I retook them and it passed. If you don't check the log, you have no idea it slacked off.

Second, clear the workspace cache once a week, in Settings under "Storage Management," click "Clean Temporary Files." If you don't clear it, it gets laggy after many runs—mine started noticeably slowing around the tenth-something run.

Third, I set a rule for myself: only let it do one thing at a time. Before, I wanted to produce the settings table and the chapter outline all at once, and it mixed up the fields from both sides, producing a mess. Later I split it into two workspaces, each running separately, and it was actually faster.

I can sort of understand the concern about Reddit's AI review summaries. What users fear is AI making judgments for them without telling them the basis. My solution on the WorkBuddy side is to add a "Source File" column in the contract—every entry must indicate which docx and which paragraph it came from. That way even if it wants to make things up, it can't, and when you check back you know whether it's right.

No matter how strong Google's model is, someone still has to define the task for it. No matter how controversial Reddit's feature is, the controversy is over who sets the rules. On my side, I just write the rules rigidly, go through the source files after each run, and keep it if it matches. The more rigid the rules, the less likely it is to go wrong.

1 replies

?
Ctrl + Enter to reply
Yanshi
Yanshi2d ago

The "change the category to classification and everything goes empty" trap is way too real—the moment you rename a field the contract no longer matches. I've been burned by this too.