UX/UI DESIGN · HEALTHCARE
Recovery Companion: An AI Guide That Turns Discharge Paperwork Into a Plan Patients Actually Follow
type
Self-initiated
ROLE
Product Designer
& Researcher
EXPERTISE
UX/UI Design
UX Research
AI-Assisted Design
Accessibility
YEAR
2026
The moment a patient is discharged is one of the highest-stakes handoffs in healthcare, and one of the most poorly designed. A patient leaves with a stapled packet of instructions and is expected to manage medications, watch for complications, and follow up correctly, often while still in pain, medicated, or frightened.
The evidence that this handoff fails is overwhelming:
Patients don't understand what they're sent home with. One study found 78% of patients had a comprehension deficit in at least one area of their care and discharge instructions, with the most dangerous misunderstandings involving medications. (Engel et al., Academic Emergency Medicine, 2012)
The instructions are written for the wrong reader. Discharge paperwork is routinely written at an 11th-grade reading level, while the average patient reads at a sixth-grade level. (Spandorfer et al., Annals of Emergency Medicine)
The system pays for it. About 20% of Medicare patients are readmitted within 30 days, costing roughly $26 billion a year, and about $17 billion of that is potentially preventable. (CMS / MedPAC)
The cause is as much a design problem as a clinical one. MedPAC attributes a large share of avoidable readmissions to inadequate communication of discharge instructions and poor follow-up planning.
That last point is the opening. The information already exists. The problem is that it reaches patients in a format no overwhelmed human can act on. This is a translation, hierarchy, and trust problem, which makes it a design problem.
The hard part of this project was restraint and safety. I was designing for a user at their most vulnerable: tired, anxious, possibly older or low-literacy, holding information they didn't fully absorb in the room.
That created three tensions to design against:
Cognitive load is the enemy. Every extra choice, screen, or word taxes someone who has very little left to give, so the product has to do as little as possible.
The stakes are real. The bar for clarity is absolute, because a mistake with medication timing or a warning sign can turn into a safety event.
AI has to know its limits. An AI companion in healthcare cannot diagnose, cannot guess, and absolutely cannot offer false reassurance. The central design challenge was building an assistant whose most important skill is knowing when to stop talking and hand the patient to a human.
Secondary research framed the problem. I started in the literature to map where comprehension breaks down, and a few patterns showed up again and again. Medication instructions are where the most dangerous misunderstandings happen. The gap between how instructions are written and how patients actually read sits at the root of it. And the people most at risk tend to be older adults, lower-income patients, and those who didn't finish high school. (Cited studies, plus Choudhry et al. on income and education correlates.)
Provisional personas:
Ruth, 68, recovering from a knee replacement. Low tech comfort, manages five medications, and leans on her adult daughter. She needs large text, a simple daily plan, and a way for her daughter to help from a distance.
Marcus, 34, discharged after an asthma flare-up. Literate and busy, but he left the ED stressed and only skimmed the packet. He wants a fast, plain-language summary of what matters in the next 48 hours, plus clear red flags.
Dana, 41, caring for an aging parent. She wasn't in the room at discharge, yet she's responsible for the follow-through. She needs shared access and notifications.
Primary research plan. To turn these personas into something evidence-based, the study includes short interviews with recently discharged patients, caregivers, and discharge nurses, who see firsthand where comprehension falls apart. I want to learn how people actually use the paper packet once they're home, where they get confused, and what they reach for when they're worried.
Rapid ideation with AI-assisted tools. With the problem framed, I used AI-powered design tools (Stitch, Visily, and Figma Make) to generate several structural directions before committing to high-fidelity work. I explored how much to surface on a home screen, how to sequence a daily plan, and how to present warning signs. The tools let me move through a lot of options fast, while I kept control of the calls that mattered: what was safe and clear enough for a vulnerable user. I ended up throwing out several AI-generated directions that looked polished but quietly added to the patient's cognitive load.
This is the section that defines the product, so I treated it as a first-class design problem rather than a disclaimer.
Grounded in the patient's own documents. The assistant answers only from the patient's own discharge document and clinician-approved content, never the open internet. When a patient asks "can I take ibuprofen with this?", it answers from their actual medication list, or says it doesn't know and routes them to their care team.
It admits what it doesn't know. Where the source material is silent, the design surfaces "I can't answer that safely. Here's how to reach your care team," rather than guessing. False reassurance is treated as the worst possible failure.
Escalation is a designed path. Red-flag symptoms (for example, the specific warning signs on the patient's own discharge sheet) trigger an unmissable, one-tap route to a nurse line or emergency guidance. The AI's job at that moment is to get out of the way.
It never diagnoses. The assistant translates, reminds, and organizes. It does not interpret symptoms or change the plan.
The thesis is simple: in healthcare, an AI assistant's credibility comes from its boundaries. Designing those boundaries clearly, so a frightened person knows exactly what the tool will and won't do, is the actual work.
1. Translate the packet into plain language. The companion ingests the discharge instructions and presents a sixth-grade-level version alongside the option to see the original. Medication entries show what, when, and why in plain terms. Tradeoff: It adds an AI translation layer that has to stay clinician-verifiable, but it directly attacks the literacy mismatch behind most misunderstanding.
2. Collapse everything into one "Today" view. Instead of a document to parse, the home screen answers one question: what do I need to do right now? It shows today's medications, today's tasks, and the next appointment, with nothing else competing for attention. Tradeoff: Less information is visible at once, but the cognitive load drops sharply for someone who has little to spare.
3. Make warning signs impossible to miss, and acting on them effortless. The patient's specific red flags live in a persistent, high-contrast location with one-tap escalation. Tradeoff: It gives prominent space to something used rarely, but the cost of missing it is a safety event, so the space is earned.
4. Build a caregiver mode. Shared, permissioned access lets a daughter or partner see the plan, get reminders, and help, which reflects how recovery actually works at home. Tradeoff: It adds account and privacy complexity, but the research shows caregivers are often where follow-through actually happens.
5. Accessibility as a baseline. Large type, high contrast, voice input and read-aloud, and graceful offline behavior, all designed for older adults and low-vision users from the first frame. Tradeoff: It constrains the visual design space, but for these users accessibility is the whole point.
Step 1: Frame from evidence. Mapped where discharge comprehension breaks down using published research, and identified medications and warning signs as the highest-risk areas.
Step 2: Validate with people. Interviews with recently discharged patients, caregivers, and discharge nurses to turn provisional personas into evidence and pressure-test assumptions.
Step 3: AI-assisted ideation. Generated and compared multiple structural directions with Stitch, Visily, and Figma Make, keeping what reduced load and discarding what merely looked finished.
Step 4: Prototype in Figma. Built a clickable prototype of the core flows: Today view, plain-language instruction reader with "ask about this," medication schedule, warning-signs and escalation, and caregiver view.
Step 5: Usability testing plan. Test with representative users, mainly older adults and caregivers, using teach-back ("show me what you'd do tonight") to measure genuine comprehension rather than task completion alone.
The status quo (before): a stapled, multi-page packet written above most patients' reading level, handed over at a stressful moment and rarely revisited. It's the format the research shows people can't act on.
The companion (after): the same clinical information, restructured around the patient's actual day.
Today view: the one screen that answers "what do I do now," showing today's meds, tasks, and next appointment.
Plain-language reader: the discharge instructions rewritten at a readable level, with "ask about this" on any line, answered only from approved sources.
Medication schedule: what to take, when, and why, with reminders and a simple taken/not-taken state.
Warning signs and escalation: the patient's own red flags, persistent and high-contrast, with one-tap routing to help.
Caregiver view: shared visibility and reminders for the person actually helping at home.
Because this is a concept, I'm not claiming results I haven't earned. Here's what I'd instrument and the hypotheses I'd test:
North-star: 30-day readmission rate, the real-world stake the whole product is built to move.
Comprehension: teach-back accuracy at home versus paper instructions. The hypothesis is that plain language plus a daily plan raises genuine understanding.
Medication adherence: doses taken on schedule.
Warning-sign recognition and escalation quality: whether the design helps patients act on real red flags while avoiding false alarms. This is the safety metric that matters most for an AI assistant.
Caregiver engagement: activation and ongoing use by the support person.
Good healthcare design adds clarity before it adds intelligence, and it respects its own limits. This concept took a documented, expensive, human problem (patients can't act on what they're told at discharge) and treated it as what it is: a translation, hierarchy, and trust problem.
It's also a deliberate demonstration of how I work in an AI-era product environment: research-led, fluent with AI tools but in command of them, and most careful exactly where the stakes are highest. The same approach that made a real-estate dashboard clearer is the one that makes a discharge plan safe. Find the highest-leverage gap, close it precisely, and design for the human on the other side.