Helping physicians gain context when time is precious

The pre-visit summary a physician reads in the seconds before walking into the room was working as a concept and failing as a page. Clinicians were being asked to read where they needed to scan.

The users see 40+ patients a day with no dedicated prep time between them. Each section delivered several sentences of continuous prose where a clinician needed to scan: we had structured the data and then re-flattened it into paragraphs. 

Two standing decisions shaped what I could propose: surface data without editorializing, “like the most humble med student”; and one experience, three data profiles, same UX pattern, different data density.

Where the idea started

The first pass at the vision: what a physician would actually see and touch in the seconds before walking into the room, moving at the speed the visit moves.

Direction one, after the workshop

I ran a workshop with clinical stakeholders and the feedback was clear: what a physician is actually scanning for in the minute before a visit starts. This direction leads with the card — conditions, recent hospitalization, specialists, key values, care gaps — and opens each assessment as a sheet over it, with an evidence rail behind every claim.

It runs across three patient types: a multi-hospitalization case, an oncology case, and one with nothing to report but noise to filter.

And the second direction

The same clinical content, ordered differently. Putting the two side by side turned an argument about layout into a concrete choice about what a physician should hit first, which is the question the workshop had actually surfaced.

Built on real patient data

I shared both directions with stakeholders and, working with the engineering and data teams on the feedback, put this version together against representative patient data.

Testing length, and what to leave out

What we heard

Work overview

I laid five real patient briefs side by side and found the system emitting the same template for all of them — for a patient with three uncontrolled conditions, one fresh out of a hospitalization, one whose only news was eight new medications from five prescribers, and one with nothing to report. That reframed the question: not what should the layout be, but how many kinds of patient does this need to handle.

Clinical software has no mature convention for this, so I went looking outside it. Briefing a busy expert on what changed since they last looked is a general problem: sales teams solved it for pre-call prep, operations teams for incident response. I analyzed thirteen products across three reference classes. The headline finding was not the one I expected; nobody in the mature class ships narrative AI summaries. Every serious product uses templated structure, type-aware sections, and traceable evidence. That retired the two-sentence “AI read of the story” I had been sketching.

I built five archetypes against real anonymized briefs — high-acuity, recent event, medication reconciliation, sparse/stable, single-domain deep — same data model, different presentational shape, with a routing rule, and took the whole exploration to product, clinical, and design leadership.

What the brief had to obey

Four experience principles carried through every archetype, and the structure spec fixed what each brief had to say and in what order. 

1. No wall of text

   Headers and bullets only — the brief has to stay scannable in thirty seconds.

2. The three-tap rule

   Reading a chip, assessing in the drawer, and saving to move on takes exactly three taps.

3. Source transparency

   Every chip and card carries a source link. The working assumption is trust but verify.

4. Mobile-first

   The drawer takes no more than 90% of screen height, so the agenda underneath stays in view.

• Section

  What it carries

• Byline

  One sentence: age, sex, and time since the last visit — or new patient if there wasn't one.

• Required assessments

  The value-based care requirements that apply to today's visit date, named explicitly.

• Clinical activity

  Medications by status — active, recently changed by outside providers, possible gaps in filling. Labs with outside results first, then chronological, values shown with their deltas.

• Preventive care

  Immunizations and screenings, each with a date or an explicit no record found.

Five patient shapes, one data model

Each archetype was built against a real anonymized brief. Same sections and same underlying data — what changes is which one leads, how much room it gets, and what collapses to a single line.

Predictability has real value at volume. A clinician reading forty Briefs a day builds muscle memory for where things live. Varying structure by patient trades that away. I tested the assumption that only showing what changed would reduce the length of the copy in the brief, what I learned was having predictable sections actually helped scannability. We used feedback on what was important to improve the model on what to display.

The Brief and the Agenda are architecturally distinct. The Brief is catch-up context with a wide net; the Agenda is focused action. Several of my shapes had started to read like action plans — useful work, but belonging to a different surface. The product lead's read was that shape-adaptive thinking might have a real home in Assessments, which opened a separate track rather than closing the idea.

Narrative synthesis was already settled. My own competitive research had found that nobody in the mature reference class ships narrative AI summaries.

The shipped brief, phone to desktop