WorkBuddy consistently misorganizes AI game asset tables; anyone else experienced this?

Worked overtime until 3 AM last night. I saw an article saying AI can generate maps, characters, and UI assets. My first reaction was a headache. Our design docs already have dozens of asset tables, and the export formats differ across platforms—some use asset_type, while others just provide a string of filenames. I wanted to use WorkBuddy for a daily task to automatically categorize newly generated images into the corresponding tables, split them into three tabs by Scene, Character, and VFX, and add descriptions for each asset. Sounds very SSR (Super Simple Right?), but it immediately forced different platform fields into the same column. UI paths ended up in the character table, and scene notes popped up in the VFX table.

I tried setting up folders and standardizing filenames first, then asked it to "identify type by filename prefix." The first time, it recognized ui_panel_01 as a character. The second time, it threw an unreadable error, something about column names being unparseable. Later, I added conditional logic, telling it to write only if fields were complete, but then it started dropping the remarks column. The final written result only had titles and paths. The most absurd part: I asked it to put failed items in a separate table, and it split the failure reasons into a dozen empty-value fields like "missing type, missing desc, missing source." That table was messier than the original.

During this month of using WorkBuddy, I've also built daily report bots, so I know about fallback mechanisms like retries, checkpoints, and run logs. What gave me a headache this time was handling ambiguous materials; its judgment logic was too jumpy. Maybe my prompts were too scattered, or maybe I didn't set the target table as the sole source in collaboration permissions. Has anyone used WorkBuddy for game asset ingestion? For issues like misaligned fields and drifting generated content, should I give it a stricter table structure first, or just break it into steps where it only classifies without completing fields? Seeking some ideas—I really don't want to manually wipe up the mess in spreadsheets anymore.