A Surprise Hospital Discharge, and the AI That Kept Me Cool

Early Friday afternoon, my aunt's sister texted me from the hospital. A social worker had just stopped by the room and said my aunt was being discharged. That day. The doctor who'd done rounds that morning hadn't said a word about it.

It was a cool but critically alarming moment. What do you mean she's being discharged today? How quickly? We have lots of outstanding questions. What does she think of that? What do we do?

That's when I went to the goals, what needed to be done here. Make sure her discharge was medically safe. Give it the best chance of being psychologically safe too, because a sudden change with almost no notice is upsetting to everyone, her most of all. Keep as much control as we could over the pace and the terms of her exit, since a hospital has a financial interest in the bed, and that isn't the same as her interest. And do all of it in the most harmonious way possible, keeping everyone calm and constructive, starting with me and my family. Knowing there were other factors in play that weren't 100% in my aunt's best interest has made us fiercely vigilant.

Minutes later they asked me to come in, and I had to go right away. I was home sick, so I put on a mask and got in the car. There was no time to think through a plan. I had to figure it all out, and make the plan up, on the drive.

Thankfully, I had a voice-capable assistant I'd been working with on my aunt's care for weeks. This was go-time. Live television, big stakes. Everything that had compounded to that point was about to find out whether it could keep up and hold up. It did wonderfully.

Two weeks ago I wrote about using AI to get through this emergency - the bedside, the flights, the family updates, her affairs. That afternoon was the pinnacle of all of it.

The Drive In, and Everything the Project Already Knew

I did the drive in voice mode, which means I talked and it talked back, like a phone call. I was frantic and I wasn't typing anything.

What made it work is what it already knew. A project, in Claude, is a workspace that keeps its own files and memory from one conversation to the next, and I'd been running this one since she went in. Every medical update had gone through it. It had helped me write every status text to family and friends, and helped me work out every question I took to her doctors. It had been with us for her whole stay, from her first weeks in the hospital through the search for a rehab facility and the review of her legal paperwork. That morning alone, before any of this, it had read in her sister's latest updates, refreshed the status document, the one file where everything about her care and placement is written down, and cleaned up a family text for me. So when I said the hospital was moving to discharge her that day, it didn't need a word of background. It knew what was still open on her care, what we'd agreed the conditions for leaving were, and which questions had never gotten an answer.

It also knew her rights. A week earlier, in that same project, we'd already looked up what happens when a hospital moves to discharge a patient on Medicare and the family isn't ready: the patient can say she doesn't consent, and there's a Medicare review organization you can call that starts a fast appeal, a formal second look at whether she's ready to leave. That afternoon, the call was simply the next thing to do.

So on a short drive, out loud, we built three question lists: one for the medical team, one for the social worker, one for the rehab facility. Claude pulled the questions from everything the project already held about her case. By the time I parked, I knew what was still open and who to ask about each piece.

Every Open Question, Asked of the Right Person

Her sister and my aunt had already said, out loud, that she didn't consent to leaving that day and would appeal, and they'd made the call to start one. That gave us the time to do the rest properly.

The medical team came in first and I went down their list. Every open clinical question got asked and every one got an answer we were comfortable with: how her medications carry over, how her therapy gets paced against her other treatments, what a couple of things we'd been watching turned out to be, what the facility would and wouldn't be handling. After they left, the social worker came in and we went through that list the same way. The hospital revised the insurance authorization, the approval a facility needs from the insurer before it can take a patient, and the insurer accepted it. The discharge was pushed back, which also let her finish a treatment before the move.

With everything answered and every concern addressed, we withdrew the appeal. Until that afternoon, our talks with her case workers had been tense, sometimes contentious. This one stayed calm from start to finish and ended on good terms. A perfect checklist, checked perfectly.

On the drive home, in voice mode again, I read every answer back in so the project had them and the loop was closed.

The Facility Call, Recorded

Later that afternoon, back home, I called the admissions contact at the rehab facility with the third list, to find out what my aunt was walking into. I recorded the call on my Plaud, a small voice recorder that produces a transcript, so I could pay attention to the person on the phone instead of scribbling.

Then I went through the transcript with Claude. It wrote all of it into the status document: a private room, visiting hours, where to park, how meals work, who checks on her and how often, what the recreation program looks like, and who the ongoing point of contact is. The one thing the contact didn't know, how many days the insurer had approved up front, went in as a question to ask at check-in.

I asked for a version I could text her sister, and sent it.

The Project Memory Behind That Afternoon

All of that was possible because the project remembers. I set it up on September 22, the week things got serious. Since then I've opened it about a dozen times, and each time it starts by reading what every earlier conversation learned. Its memory file holds more than thirty dated decisions and session notes: what the discharge conditions were, which facilities were screened and why the rest were dropped, what the care team said on which day, what her coverage does and doesn't pay for. There's a living status document, a care-needs profile, a ranked facility shortlist that started from 136 facilities, a legal-documents review, and a log of every bill paid on her behalf.

That's what was loaded when I got in the car. I didn't have to explain anything on the way in. I just drove and talked.

I've said before that I don't trust AI for medical advice, and I didn't ask it for any that afternoon. What I asked it for was to keep track of her case so I could keep my attention on the road, and then on the people in front of me. It told me what was still open and reminded me what we'd already decided, and a frantic afternoon turned into a short list of questions with good answers. She'll move to rehab in the next few days, with her blessing and ours.