Contact

Pillar 03 — Surfaces

The human side of AI.

AI is only as useful as the surface a person touches. We build small, well-designed tools that smooth the rough edges of using AI — the input affordances, the recovery paths, the moments where confidence and clarity actually matter.

What we work on Talk to the studio

Most friction with AI is design friction, not capability friction.

The model can do the thing. The interface does not make the thing easy to ask for, easy to verify, or easy to recover from when it goes wrong. The experience layer is where the next wave of AI adoption gets won or lost — and where good design has outsized leverage.

The Nielsen Norman Group's research on AI UX patterns shows the same thing from the user-research side: people's complaints about AI are mostly complaints about the surface, not the substance.

What we work on

Five places good design changes the math.

01

UX patterns for AI.

Reusable design patterns for the specific shapes of AI interaction. Streaming responses, partial confidence, graceful failure.

02

Prompt and palette interfaces.

Fast, keyboard-driven surfaces for invoking AI. Less chat box, more command palette.

03

Voice and multimodal flows.

Patterns for AI that goes beyond chat. Voice, vision, and combinations that feel like one product, not three.

04

Accessibility for AI tools.

Making AI interfaces actually usable for everyone. Screen-reader-friendly streaming, keyboard-only flows, captioned outputs.

05

Human-in-the-loop patterns.

When to bring a human in, and how to keep the experience graceful when you do. Approval flows that do not feel like paperwork.

Featured incubations

Surfaces in progress.

Prompt Palette — a library of prompts, by role A library UI: a left-rail list of job roles and personas; on the right, the selected role's prompt set with a small panel showing how the prompts adapt to current context. Hairline strokes; single Spectrum-purple accent on the selected role. FIG · 03 PROMPT LIBRARY Prompts that know your role. PROMPT · PALETTE · acme co ROLES · 12 Product manager 14 Engineer 22 Designer 11 Sales · AE 18 Customer support 9 Legal · counsel 7 Recruiter 6 Analyst · finance 8 Marketing 10 Founder · CEO 5 + Add a role ENGINEER · 22 PROMPTS Curated by the team, refined by use. ⌕ search prompts in this role ⌘ K Turn this ticket into a clear PR description. Reads the linked JIRA + diff. Drafts a PR title, summary, test plan. USED 42× Sketch a migration plan from this schema diff. Includes rollback steps. Asks before destructive changes. USED 31× Review this PR for our team's style guide. Knows the repo's conventions. Flags drift, not nits. USED 28× Explain this stack trace to a junior teammate. USED 19× Draft RFC from this notebook of ideas. USED 12× CONTEXT · ACTIVE Prompts in this role adapt to: › repo: acme/api › branch: feat/billing-v2 › on-call: Priya › last PR: 4 hours ago REFINED OVER 4 WEEKS accepted edits feed back into the role's defaults. A LIBRARY PER ROLE · PROMPTS THAT GROW WITH THE WORK
Experience

Prompt Palette

Prompt generators based on specific job roles and personas that evolve with context.

In incubation
Continuous Context — context beyond the window Editorial diagram on a light Canvas surface. A small CONTEXT WINDOW sits above a much wider PERSISTENT MEMORY band that extends across eight months of sessions. Two arrows connect them — one extracting facts down into memory, one retrieving relevant facts back up into the context. Single Aurora-teal accent on the active retrieval flow. FIG · 01 CONTEXT BEYOND THE WINDOW What the model holds. What it can reach for. CONTEXT WINDOW · NOW 12K / 200K TOKENS › goal · ship Q3 launch brief by Friday › user · Priya — product manager — declarative tone › project · acme/api — branch feat/billing-v2 › retrieved · pricing rationale from May review ✓ › working · §3 risks per team EXTRACT RETRIEVE PERSISTENT MEMORY · OFF-CONTEXT 4.2M TOKENS · 8 MONTHS launch v0 plan team norms priya — preferences tradeoffs log pricing rationale · may last review notes api conventions stakeholder map MAR APR MAY JUN JUL AUG SEP OCT · NOW A SMALL WINDOW · A LONG MEMORY · NOTHING LOST BETWEEN SESSIONS
Experience

Continuous Context

Extend context beyond context windows and optimize the context to specific outcomes and goals.

In incubation
Handoff — task moving between agents and a human reviewer An architecture sketch with three nodes (Agent A, Agent B, Human Reviewer) in a triangle. A task object moves along thin arrows; the handoff to the human is highlighted with a 2px Spectrum-purple arrow. FIG · 02 HAND-OFF AGENT · A Researcher finds sources AGENT · B Drafter writes prose HUMAN · REVIEWER You decides what ships TASK PASSES CONTEXT HANDOFF A · INTER-AGENT EXCHANGE B · APPROVAL TO A HUMAN
Experience

Handoff Loop

Smooth agent-to-human handoff with structured context summaries.

In incubation
Approval card — pending AI action A single, calm approval-flow card on a Canvas background. Pending review eyebrow, a one-sentence summary, and three buttons: Approve (Aurora teal), Edit (Ink), Decline (ghost). FIG · 03 APPROVAL CARD PENDING · REVIEW drafter wants to send a draft brief to the team channel. Channel #q3-launch · 4 recipients · 412 words · 0 attachments. + Brief v1 — 4 sections, one page. + Headline metric: time-to-useful = 1.84 s (–31%). + Open question: pricing footnote, §2. › Awaiting your call before send. Approve · ↵ Edit Decline A — APPROVE · E — EDIT · ESC — DISMISS ONE DECISION · CLEARLY STATED · WITH AN UNDO BEHIND IT
Experience

Approval Cards

Human-in-the-loop patterns that respect the human's time.

In incubation
Voice surface — phone showing a live transcript Editorial illustration of a hand holding a phone vertically. The phone shows a clean voice-input interface with a real-time transcript and a streaming summary. Hairline strokes, restrained palette. FIG · 02 VOICE SURFACE 9:41 ●●●● LISTENING Live transcript "Okay — I want a one-page brief on Q3 launch risks, grouped by team. Skip the summary slide. Keep it to four sections, max."▏ 12.4 s · 38 words · 1 speaker DRAFTING › 4-section brief: risks per team. › drop summary slide. › cap length to one page. › tone: declarative · no hedging. HOLD TO TALK · LIFT TO PAUSE · DOUBLE-TAP TO REWIND
Experience

Voice Surface

Voice-first AI flows that read well as text too.

In incubation

Designing in this space?

We trade notes with designers and researchers building thoughtful AI surfaces. Drop us a line.

Contact the studio