I built a companion. She only wanted the calendar.

Every essay before this one on this site was written about a product called pal.fun — a companion AI you could see all the way into. I still think it was beautifully built. I also think it was, at its core, a bet the market quietly declined to take. This essay is about the day I admitted that, and what I did next.

If you've read the judgment and the hands, you already know the rule I keep coming back to: my real job isn't building, it's saying no — and the biggest, hardest no is deleting something I was proud of. I've killed a marketplace that way. This time I killed the whole product.

The beautiful bet

pal.fun ran on an a-priori thesis, and on paper it was a good one: people don't want another tool, they want a someone. So the product was a companion. You named it. It had a naming ceremony, a soul written in prose, an identity you could read, a rolling memory of your relationship. The whole glass box existed to make that someone legible and trustworthy. Every essay here was, in some way, in service of that bet.

I believed it. I still find it compelling. But "compelling to me" and "compelling to the person actually using it" are two different measurements, and I'd spent weeks optimizing the first one.

The one thing that landed

Here's the fact that reorganized everything. I put the product in front of my own household — I wrote about that in one pal, the whole household — and I watched what my wife actually reached for. It wasn't the companion. She didn't want to name a buddy, or read its soul, or keep up a relationship with it. What she did, and loved, and did again, was snap a photo of a kid's birthday invitation and watch every date on it land on the calendar by itself, with a reminder attached.

That was the whole signal. One validated fact in a few weeks of building: the photo-to-calendar magic was real, and almost nothing else was. The Someone I'd designed the entire product around was the part she walked past on her way to the utility.

The economics said the same thing, louder. pal.fun runs on your own API key — you pay per token. Paying real per-message money to chat with a companion was never going to pencil out for a normal person, because the thing they compare it to is the free chatbot already on their phone. But paying a few cents to have a season of game dates read off a photo and filed for you? That's a chore you hate, done. That's worth money. A feature has to earn its round-trip — and the companion conversation never did, while the data entry earned it every single time.

Target user beats a-priori thesis

I'd learned this exact lesson before, one size down. I once built an iCal clone off a theory about what users wanted, and the actual users quietly told me the theory was wrong. The rule I wrote down then was blunt: target user beats a-priori thesis. When the person in front of you and the idea in your head disagree, the person wins. Every time.

pal.fun was that same mistake, one level up — not a wrong feature inside the product, but a wrong bet about the whole product. The a-priori thesis was "people want a someone." The target user, standing in my kitchen, wanted her calendar filled in for her. I'd been out-arguing my own user with a theory for weeks.

So I stopped.

What I kept, what I cut

The pivot was brutal in the way the good ones are: it deleted far more than it added. Out went the entire companion layer — the ceremony, the soul, the vibe, the identity files, the rolling relationship memory, the whole apparatus for being a Someone you have to maintain. Out went the thesis every earlier essay was written to hold up.

What stayed was the part that worked: the calendar, the notes, the photo-to-events magic, the shared-household layer, the reminders — and the agent underneath, kept on as the invisible engine that does the entry and the remembering, not a friend you keep up with. Same boring chassis I've written about more than once. A fraction of the surface. An order of magnitude less cost per interaction, because a clean "add the dentist Tuesday at 3" is one cheap hop, not a relationship turn.

The new product is Dustav — the name means "do stuff," and it happens to read like a real one. (It was almost Endar, for cal-endar, the validated wedge sitting right there in the name — a trademark snag sent me back to the drawing board, which turned out to be a gift.)

The reframe that saves it from being "just a calendar"

Here's the objection I had to answer for myself before I could commit: every phone ships with a calendar. Why would anyone want yours?

Because the product was never the grid. The grid is the free part. The product is the thing families actually hate doing — the data entry and the remembering — done for them by an agent that reads the photo, files the dates, keeps the notes, and surfaces the reminder in its own voice at the right moment. Nobody is short on calendars. Everybody is short on someone to do the calendar. That's the whole company.

Which is the sharpest irony of the pivot, and I'll say it plainly, because I try to say the awkward things here: pal.fun's deepest problem was that I couldn't pitch it — its value only showed up after you'd lived with it for a while. Dustav I can pitch in one sentence. Snap a photo of the schedule; every date's on your calendar. Same builder, same chassis, same transparency — but a thing you can hand someone in a sentence, instead of a thing you have to convert them into over an afternoon.

Why pal.fun's essays are still up

I could have deleted everything above this line. The whole archive argues for a product I no longer ship. But taking it down would break the one rule this project has never broken: you build in the open, including the parts you got wrong. The pal.fun essays aren't wrong — most of what's in them is still true, and the engine they describe is the engine Dustav runs on today. They're just the first chapter. This is the hinge. Everything after it is about the thing that survived.

I built a companion. She only wanted the calendar. It took a few weeks of building and one honest afternoon to tell the difference — and the product got smaller, cheaper, and far more itself the moment I did.