Create / update / verify Roblox in-game UI (ScreenGui / Frame / TextLabel / …). Always start with `design_brief`. For ambiguous input, it inspects Studio context and returns candidate brief patches; Do not dump enum lists at the user. An empty StarterGui does not prove runtime UI absence: for an existing-UI review, follow `runtime_review_required`, collect a structured Play Client PlayerGui snapshot, and pass its saved `runtimeEvidenceRef`. Pass relevant repository evidence in `projectContext`; explicit screen job, visual thesis, forbidden defaults, tokens, and approved assets determine the direction without a built-in subject preset. `design_brief` returns a primary-view-first Design Contract and an advisory Layout Plan with neutral regions, AnchorPoint strategy, and responsive behavior; it is not a hard tree preset or style quota. A calling agent may declare provider-neutral image generation capability through `agentCapabilities`; missing or unavailable capability must use the explicit existing, user-provided, Creator Store, or assetless fallback. Keep the returned `brief_id`, `design_contract`, and `layout_plan` through creation, update, preview, and check. The asset recommendation gate considers user references, project assets, existing UI, and user-approved `manage_assets.search` intent; do not use a recommendation until the user accepts it. After each meaningful create or update, preview the representative primary view and submit a calling-agent critic verdict (`recompose`, `refine`, or `approved`) with `visualReview`. Read `primary_visual_verdict`, `functional_safe`, `responsive_requirement`, `responsive_ready`, `required_dimensions`, and `readiness` separately from mutation success. Responsive evidence blocks readiness only when the Design Contract marks it required. Read `weppy://ui-studio/guide` for the authoring workflow, `weppy://ui-studio/tokens` for optional visual examples, and `weppy://ui-studio/functional-rules` for machine-readable advisory evidence rules. Design Check permits `priority_high` only for observed functional visual failures such as meaningful overlap, unreadable content, invalid geometry, or primary-view occlusion; style-only findings remain medium or low.