WorkBuddy messes up field mapping with mixed inputs; anyone else seen this?

Bottom line first: Today I saw a post on UISDC saying Agents have low success rates in real work scenarios. I was initially stubborn about it. Then WorkBuddy turned my weekly report into a collage. It can definitely do the work, but as soon as it does, it mixes up delivery sheets, PRDs, and color value files into a mess—like a bug pushing cat litter onto my color swatches.

I've used WorkBuddy for document archiving this past month, which went pretty smoothly. Today I wanted to slack off, so I threw in a PRD Markdown file, a schedule table from Feishu (Lark), and a design-tokens.json file, asking it to generate a weekly report based on a fixed template. The goal was simple: summarize this week's pages, blockers, and pending color values, and output a table ready to paste directly into the group chat. I clicked "Organize by Template" and selected the Markdown template. Result? It treated the color and font-size fields in the tokens as requirement text, merged the table rows incorrectly, and finally gave me an error: "Field mapping not recognized." I thought it might be a Chinese path issue, so I renamed all files to English and ran it again. Still wrong.

Strangely enough, having it read the PRD alone works fine, and organizing tables alone is also okay. But once you mix them, it starts getting creative with its imagination. Last week I wrote that it shouldn't compete with 3-second posters on aesthetics; now I feel it at least shouldn't compete with 3-second layouts on formatting. The time saved by automation shouldn't all be spent fixing its ugly formatting.

Has anyone else encountered this? Is it because my field names in the template are too colloquial, or does it just half-understand mixed inputs of nested headers and JSON? Maybe I messed up some setting somewhere. Asking for help here, otherwise next week's report will turn into overnight latte stains again.