ROY SHILOH

Designing the core clinical workflow for an AI scribe and EMR used by mental-health professionals.

First designer, then Head of Design.

Documentation is the easy story. The clinical workflow around it is the hard one.

How clinicians begin and manage a recording.

Problem

Sessions start unpredictably. A recording that fails silently, or a UI that pulls attention away from the patient, is worse than no scribe at all.

Options

A modal “start session” flow; a persistent recorder dock; auto-start on calendar match. Each pulls attention away from the conversation in a different way.

Decision

A persistent, low-attention session dock that stays put across the app. State is always visible; controls are one action away; nothing blocks the clinician’s primary view.

Result

Clinicians can move between patient context, prior notes, and the assistant while a session is running without ever losing the recording or wondering whether it’s live.

Persistent session state — the recording control never leaves the frame, so clinicians can move between documentation surfaces without losing the active patient or the live capture.

Fictional data

How AI-generated documentation is reviewed and completed safely.

Problem

A generated note that looks finished invites the clinician to sign without reading. A note that looks unfinished slows every session. In mental health, either is unacceptable.

Options

Auto-approve with an edit affordance; force line-by-line confirmation; a diff-like editor between draft and clinician edit. Each trades speed against safety differently.

Decision

The AI produces a structured draft; the clinician sees it as an editable working document, not a finished artifact. Nothing is signed until the clinician acts. AI suggestions in every downstream surface (diagnosis, plan, assistant) stay optional and reversible.

Result

Clinicians move faster than typing, but the note that gets signed is unambiguously their note. That framing is what makes the product usable in real practice.

Generated note as a working document — the clinician is the author, the AI is a first draft. No AI output ships downstream without an explicit clinician action.

Fictional data

How information moves between notes, diagnoses, plans, and the assistant.

Problem

A note, a diagnosis, a treatment plan, and an assistant are four separate surfaces. If each has its own idea of the patient, the clinician spends the session reconciling them.

Options

Shared sidebar; global patient header; each surface as a tab of a single “patient workspace.” Each shapes the mental model of the product differently.

Decision

One patient workspace. Notes, diagnosis, treatment plan, and the assistant are surfaces inside it — they read from the same context and write back to the same record. Navigating between them never loses the active patient.

Result

The clinician stops re-selecting a patient on every screen. The assistant answers with the patient it’s already looking at. New surfaces slot into the same shell instead of forking it.

Diagnosis surface — same patient context as the note and the plan; the clinician doesn't reselect.

Fictional data

Treatment plan — reads from and writes back to the same patient record.

Fictional data

Assistant — answers about the patient the clinician is already looking at.

Fictional data

Sessions index — the shell every workspace is entered from.

Fictional data

A shadcn-based system that ships product and marketing from the same tokens.

I built the Berries system on a customized shadcn foundation. It was the right choice for a small team shipping fast: a familiar base for engineering, full control over primitives, and semantic tokens that carry across product surfaces and marketing without a second design layer.

The system covers color, type, spacing, and the component primitives that make up the clinical workspace. Marketing pages and PDFs consume the same tokens, so the brand and the product never drift.

iPad-first, because that's where the sessions happen.

The mobile app extends the workspace into the room the session happens in. It's the same design system, adapted for touch and for the specific moment where a clinician needs the fewest possible controls.

Shipped to production, adopted by active practices.

Safety is a design property, not a disclosure line.