WorkBuddy: Don't Connect to Groups Yet; Sandbox It in a Folder First

I've used WorkBuddy for about a month, and I actually advise against letting newbies hand over entire departments immediately. Today I saw a Tencent Cloud WorkBuddy desktop agent deployment practice article claiming zero-deployment, compatibility with OpenClaw skill systems, and integration with WeCom, QQ, Feishu, and DingTalk. My first reaction wasn't to connect everything, but to lock it in a folder. The easiest way AI office tools break isn't failing to work, but opening permissions so wide you lose track of what they touched.

The so-called OpenClaw skill system is simply understood as importable capability packs, like plugins, telling WorkBuddy what to do when encountering certain instructions.

My real scenario is client reconciliation. Every few days, a zip file arrives containing PDFs, Excels, and photos. Previous workflow: download, unzip, rename files, open Excel to merge, copy fields, run cleaning, export CSV. CSV is comma-separated table format, convenient for later processing with Pandas or SQL. Took about forty minutes each time, often late at night.

So I started with local isolation. On Windows, create D:/workbuddy_inbox; on Mac, ~/workbuddy_inbox. Open WorkBuddy installer, drag to Applications on macOS, run .exe on Windows. First launch requests file access and accessibility permissions; former lets it read specified folders, latter lets it operate UI. I only allowed access to this inbox, not the whole drive.

Then go to top-right Personal Center, click Claw Settings, select Enterprise WeChat, scan QR to bind test bot. Create a test group in Enterprise WeChat with only yourself, pull bot in. Don't go straight to work groups. I learned the hard way; a colleague shouted "process this sheet" in the group, and it messed up the wrong directory.

Drag zip into inbox. Return to WorkBuddy main interface, input: Scan workbuddy_inbox for new files, identify PDFs, Excels, images, extract date, supplier, amount, tax, file path, output UTF-8 encoded, comma-separated CSV to output folder. First time it gives a plan; I confirmed then executed. About ten minutes later, result.csv and log.txt appeared in output.

Later I fixed this as a skill. WorkBuddy supports importing .skill files or creating new ones in skill page. I created a skill called Reconciliation File Receipt, trigger word receive reconciliation statement. Execution actions written in three sentences: move to inbox, call extraction rules just defined, send result path to WeCom bot. Skill definition is essentially JSON, mapping triggers to actions.

What truly saves time isn't how smart the model is, but that it runs the same set of actions every time.

Daily ops are also lightweight. Weekly I check log.txt for failed files. Failure reasons mostly three types: skewed photos, inconsistent field names, scanned PDFs. I tried switching default model in config.yaml from hunyuan to deepseek. YAML config files are key-value pairs. My results were slightly more stable but slower, so I only switch for month-end reconciliation.

On permission collaboration, colleagues can't modify skills, only drop files in shared directory. WorkBuddy reads only inbox, writes only output. WeCom bot stays in test group. Only after three consecutive error-free weeks did I consider connecting to formal groups.

It also misidentified Chinese commas as separators, causing column shifts. Later I hardcoded in instructions: Field names fixed, separator English comma, amounts keep two decimals. Invoice remarks often missing, so I still manually checked lastly.

Last Wednesday night batch, uploaded then went showering, came out to review for ten minutes, saved about twenty. Don't underestimate twenty minutes; for 996 workers, that's life.

WorkBuddy's core value isn't thinking for you, but pulling you out of non-value-adding actions like downloading, renaming, merging, and cleaning.