Agent-to-UI

Messy intent in.
Useful interface out.

I design the rules, states, trust signals, feedback loops, and component systems that let agents create useful interfaces safely.

Interactive prototype

From ambiguity to action

Change the intent, inspect what the agent inferred, then correct its memory.

02 Extracted signals
Time context20 minutes
96% · Stated
Goal typeGet organised
82% · Inferred
Friction toleranceKeep it simple
91% · Stated
Input modalityPhone
94% · Stated
03 Generated micro-UI
Ranked by fit

Useful is a behaviour, not a personality.

01 / Trigger

Show UI when comparison matters

Use structured UI when the user needs to compare, rank, edit or confirm. Use text when a direct answer will do.

02 / Ranking

Make the ordering legible

Rank by stated constraints first, inferred preferences second. Explain why the top option fits.

03 / Correction

Correction belongs in the flow

Every inference can be changed where it appears, without asking the user to start again.

04 / Memory

Confidence needs provenance

Show what was stated, inferred or confirmed. Store corrections with source, confidence and a way back.

Explore more work
Laptop showing the Generative UI playground with Open Brain artifact

Generative UI, A2UI and a knowledge base

Six weeks building a generative UI for a personal knowledge base typed components, agent composition, write-back actions then discovering Google had formalised the same pattern space as A2UI.

Read the story