The 30-second version
Lodestone is a concept for a personal finance assistant. The interesting design problem is the response, not the prompt. Each answer is a small generative card built from the user’s own data, with sources cited and confidence shown.
Naming
The original guide
Lodestone — a naturally magnetized piece of mineral. Before satellites, before maps, sailors carried lodestones to find true north. The name says: this tool is a guide, not an oracle. It points; it doesn’t pretend to know.
Problem
Chat is the wrong shape for financial questions.
A plain LLM chat over a person's money is the worst-case interface. Questions need a chart, an allocation breakdown, a side-by-side. Long paragraphs of explanation hide the answer. Citations get dropped. Confidence gets faked.
I wanted to design what a serious version looks like. Conversational where conversation works. Generative UI where a card is what the question actually needs. Trust UI as a first-class element, not a footer.
Approach
Treat every answer as a small designed artifact. The card is the answer. The prose is just the read-aloud version.
Product
A live conversation
Step through the demo. Each prompt builds the next answer in place with its own card primitive, source pills, and confidence sliver.
Lodestone Workbench · click a prompt to advance the thread
Pattern library
Three card primitives. Composed per question.
Allocation
Compare
Tick mark = category median
Cashflow
Trust UI
Three signals that earn the user's trust
Most assistants pretend to be confident at everything. Lodestone shows what it actually knows.
Source pill
Every claim points to a specific dataset. No "based on general knowledge."
Confidence bar
A 4px sliver. Green above 85%. Amber below. Not a number, not a percentage.
Don’t know
Cells that grey out and label themselves when data isn’t there. May vs June in the cashflow card.
What I'd do next
The third turn
Right now the demo handles single questions well. The next design problem is the third turn in a conversation, where the user is iterating on a card. “Re-run that with only the equity sleeve.” “Compare this to last year.” The card needs to know how to update in place without losing the user's reference point.
That's the boundary where generative UI gets actually useful. Not a fresh card per question, but a card the user can mold.