Why does WorkBuddy output dirty data from design briefs?

Just saw a post saying that in Penpot's MCP AI prompt engineering, the Design Agent shouldn't just listen to adjectives; it needs a structured design brief. I got itching hands after reading that, so I threw my local pile of backend screenshots, a few design-tokens.json files, and a PRD snippet into WorkBuddy. I wanted it to parse the brief first, breaking down Context, Goal, Inputs, and Constraints into verifiable fields, and then generate an acceptance checklist on the side.

I've been using WorkBuddy for a month, usually just letting it organize tables, but this time I wanted to push it further. I wrote out Context, Goal, Inputs, and Constraints following that brief structure, but it still parsed dates from filenames into fields and turned color values into plain text. I clicked on structured extraction and asked it to output acceptance items according to the instructions, but all I got back was a bunch of crammed text blocks—ugly like day-old latte.

I'm not sure if WorkBuddy's Skill can't directly ingest screenshots, or if I need to set up fixed fields first. Maybe I misunderstood, treating tokens in the design draft as constraints when it treated them as attachments. Has anyone tried this approach of feeding a clean brief first and then having it validate fields? Does it get messed up halfway through?