A little room: the postpartum app I'm not shipping
When I started this blog, I said I’d also write about the things that didn’t work. This whole post is about one of them: an app concept I’m putting on pause.
What I need this time around
One of the things I know I need for this pregnancy is more postpartum support. I struggled last time as a first-time mom. Everything felt overwhelming. I’m someone who wants to control every outcome and solve every problem (like most PMs), so I put a lot of pressure on myself and let things get to me that I should have let go.
I also forgot to take care of me. I filled up with mom guilt for every moment I wasn’t spending with my baby, and I turned down help I should have accepted because I didn’t want to get used to support that wouldn’t always be there. There was no room left for me.
Interviewing myself
So I spent a lot of time with ChatGPT talking through what I wanted from a postpartum app. It interviewed me like I was the user (because I am) to figure out what I actually needed.
It quickly went in a common direction I didn’t need: an app that tracks how often your baby feeds, how many wet diapers they have, and so on. In hindsight, that was the first clue to how this would end. Left to itself, the model reached for the most common pattern it knew. Those apps already exist, and honestly, this time around I want less of that data. I don’t need to know how many ounces my baby drinks a day and how that trends. I know that isn’t good for my mental health.
My mental health. That’s when I realized what I wanted most was an app that helped me make a little room for myself: something to help me hold on to my identity and well-being outside of being a mom. To state the obvious: being a mother is wonderful, but it’s so important to also remember that the mother is a regular human with non-maternal needs too.
What mattered to me
- Gentle and flexible. I wanted to set goals for things I’d get to eventually, not immediately, without creating yet another daily to-do list to feel overwhelmed by.
- Skipping a day isn’t failing. Not doing something today doesn’t mean I am not interested. It just means other things got in the way, or my energy wasn’t right for it.
- Adapting to what I can handle. One of my goals is to get back into running, and that will come in steps. A 15-minute walk is much better than no walk, and so is a 5-minute walk. Going from 15 minutes to 5 isn’t a regression in my recovery.
- More than getting back in shape. I also want to protect my relationships, keep up my hobbies, and so on. So it needed to coach me across different areas, but focus on only one or two a day so I’m not overwhelmed.
- A coach, not a tracker. Coaching became the theme we built around. What I wanted was a postpartum coach that could adapt and rebuild my plans as my needs changed. Pregnant me can’t really predict what postpartum me will need, and I fully expect postpartum me to keep shifting over time. I also knew I’d have very little time to iterate on the app after the baby comes, so the idea of a coach that adapts itself for the user was exactly what I wanted.
I still believe in this list. These principles came from living through postpartum once, and that’s the part of this I’m qualified to design. The problem was everything underneath it.
What I learned
The coach didn’t know what good looks like. I don’t mean training data, since I wasn’t training a model. I mean expert knowledge: when it’s actually safe to start running again, what pelvic floor recovery involves, what the early signs of postpartum depression look like, and how to nudge someone without adding pressure. Without that, the coach fell back on the general knowledge of the underlying model plus the rules I gave it. My design made it feel different from ChatGPT, but the advice underneath was the same.
I can’t do user research on a future version of myself. I was the user, but the user I was designing for doesn’t exist yet. My first postpartum helped, and it’s where the list above came from. But one experience tells me what was difficult. It doesn’t tell me what would have helped, or whether it’ll be the same this time.
You can’t write evals for something you can’t judge. I tried to work around all of this by building an eval set: example situations and what a good coach should say. That’s where it fell apart. To grade the coach’s answers, I needed to know what a good answer was, and I’m not an expert in postpartum care or in how to coach someone through recovery. This is an area that requires real expertise, ideally from someone with a breadth of experience. You can’t build an app like this with knowledge of just one user.
Why I’m pausing it
Publishing this app would be irresponsible. Imagine the coach cheerfully telling someone three weeks postpartum that it’s a great day to get back to running. Or telling someone they just need to listen to some relaxing music when what they need is to talk to their doctor. This app is indirectly dealing with mental health, at one of the most vulnerable times in a person’s life, and bad guidance, or even unsupervised guidance, can do real harm. When someone is vulnerable, you need a human involved to intervene.
So I’m not publishing this app, not even as a beta. Can’t I strip out the coaching and ship a gentle goal tracker? I thought about it. But a to-do list with kinder copy isn’t what I needed, and the adaptive coaching was the whole point.
I’d love for an app like this to exist, but I’m not the right person to build it, at least not today. What would change that is a real expert in the loop: a postpartum doula, a pelvic floor physical therapist, or a perinatal psychologist who could define what good looks like and grade the coach against it.
In the meantime, not publishing isn’t the same as throwing it away. Using it myself puts no one else at risk, so I may keep it on my own phone to learn where its edges are and what I’d do differently, for this project or others.
Building was easy - shipping is not
This is a good reminder that AI can build a lot of things. I think the scaffolding for this app is solid. What’s missing is the knowledge and judgment underneath it, which has to come from a human expert, not a generic LLM.
I keep learning this. When I built Eat with GDM, the code was easy, and what I couldn’t build was the deep integration Dexcom already had. This time the code was easy again, and what I couldn’t build was the expertise. In both cases, the hard part was never the app.