Direction 02 · 5-mockup set

Workspace / Canvas

The agent and the buyer co-edit a shared workspace, not a chat transcript. The agent renders interactive components from a fixed kit — never a wall of text — and they become pinned, manipulable artifacts you keep, rearrange and share.

Anchored at the tool-rendered end of generative UI (Vercel-AI-SDK pattern): the model picks and fills a pre-built component — AppraisalCard, CompareMatrix, FitChecklist, AreaPanel, SourceCard — it never authors the valuation number or its markup. That is the literal UI of □ KNOW (computed, auditable, cited) vs RATE (the agent's read, visually softer).
conservative
inventive
Conservative · shippable
↗ Open full screen

1 · Conversation spine

Chat thread, but every reply is a widget
Idea
A familiar thread where the agent never answers in prose — each reply materialises as a rich card (appraisal, comp table, fit checklist, honest-gap). A pinned-artifacts tray holds what you keep.
Interaction
Ask via chips or text → live step-trace → a tappable card. Star any card to pin it.
MOST SHIPPABLE
For: universally understood, low risk, small component kit, latency hidden by the reasoning trace. Against: still a linear scroll — weaker for side-by-side comparison than the board.
Middle · workspace
↗ Open full screen

2 · Command canvas

No chat — a stack of artifacts, driven by a command bar
Idea
A scrolling workspace of artifacts (compare matrix, appraisals, suburb profile). A command bar + chips materialise new ones at the top; the compare matrix is the centrepiece.
Interaction
“Compare these” / “Explain a suburb” build cards; each is pinnable & shareable. No thread to scroll back through.
THE BRIDGE
For: persistent scannable surface + a command grammar, still a fixed kit (Lynx-friendly). Against: vertical stack, not yet 2-D spatial arrangement.
Bold · spatial board
↗ Open — drag the cards

3 · Comparison board

A draggable 2-D canvas; the agent annotates in place
Idea
A spatial board — listings, the compare matrix (centrepiece), comp-provenance & a working node, all as draggable cards wired together. The agent pins notes in place (“Best buy”, “guide optimistic”).
Interaction
Arrange cards in space; a command bar (no chat) drives it. Drag is genuinely functional — try it.
HIGHEST CEILING
For: buying is spatial + comparative; couples lay out contenders, KNOW/RATE physically split. Against: re-opens some blank-canvas discoverability risk; hardest to build on Lynx.