WorkBuddy: Managing context matters more than managing prompts

I've been using WorkBuddy for about a month. On Tuesday night, our club needed to submit weekly course project reports. I threw over a dozen Feishu-exported PDFs, chat log screenshots, and an Excel task sheet into WorkBuddy. At first, it was fine—summaries came out, and tables merged correctly. But as the conversation got longer, it started mixing up the "Owner" and "Notes" columns. I initially thought the model was dumbing down, but after checking the official tips, I realized the problem was on my end: the context window was nearly full, too many skills were enabled, and I hadn't switched modes.

My setup is WorkBuddy Desktop, which might not suit everyone. First, organize your directories clearly: D:/ClubProjects/RawWeeklyReports/, D:/ClubProjects/ReportOutput/, and D:/ClubProjects/ReportBackup/. Keep only member-exported files in the raw folder; don't let WorkBuddy modify them there. After opening WorkBuddy, go to Workspace and find Skill Management on the left. You'll see a row of skill cards, each with a toggle switch on the right. Previously, I tried to save effort by enabling Document, Spreadsheet, PPT, and Data Analysis all at once. Later, I found that more skills aren't always better. Now I only keep File Organization, Table Conversion, and Document Summarization, turning off the rest. Note: just turn them off, don't delete them. If I need to write a PPT later, I can re-enable them; switching costs are low.

Don't ask it to do everything immediately. Click the mode entry next to the + button in the input box and switch to Ask Only. Ask mode is just for chatting, no execution, so it doesn't mess with files randomly. In Ask mode, clarify the goal: extract task status from raw files, merge into one CSV with fixed fields: Item, Owner, Deadline, Progress, Risk, Source File. If you're worried your prompt is messy, click Enhance Prompt. It refines colloquial requests into cleaner ones, expecting shorter prompts with clearer constraints. Once confirmed, switch back to Craft to execute. This step is crucial: define boundaries with Ask first, then run execution with Craft. It's much more stable than asking for full automation in one sentence.

During execution, I watch the Context Usage indicator at the bottom right of the dialog. One solid tip from the official docs: the model performs best when the usage circle is around half. Beyond half, responses slow down, and fields get missed easily. Don't rush to start a new task; instead, have it generate a task checkpoint detailing current goals, completed work, key constraints, modified files, unresolved issues, and next steps. Then copy this checkpoint to D:/ClubProjects/ReportOutput/TaskCheckpoint.md. Finally, send /compact to let WorkBuddy compress the previous conversation. Compression isn't lossless, so saving the checkpoint to disk before compressing gives me confidence to continue.

I also hit pitfalls in collaboration. Initially, I had everyone drop files into one cloud drive directory. WorkBuddy read them quickly, but someone accidentally deleted the original Excel file. Later, I split permissions into three layers: members only submit Feishu document exports; my local WorkBuddy has read-only access to RawWeeklyReports; aggregated results go to ReportOutput first, where I manually verify them before uploading to Feishu. Feishu sheets are set to comment-only (no edit) until owners confirm, then switched to editable. This way, WorkBuddy handles the grunt work, and humans handle acceptance.

Now I run maintenance weekly: check for redundant skills on Sunday, clear expired raw files on Monday, and write checkpoints at each stage for long tasks. Try not to switch models mid-task; if you must, use /compact before starting a new session. With this workflow, my time spent organizing weekly reports has dropped significantly, from about half an afternoon to just over twenty minutes. Next time I encounter field mix-ups, I'll check context usage and skill toggles first, rather than tweaking prompts.