Good recommendations arrive when you can't act on them, then disappear into screenshots, notes and texts. I designed, built and shipped Rexx solo, a native iOS app that keeps each one with the person who gave it.
Founder, designer & developer · Solo · On the App Store since February 2026
Impact: about 300 installs, about 90 active users, $0 spent on marketing.
Everyone has a show for you
Here’s how it works
Tap Tell Rexx, say it the way you’d tell a friend, and the Claude API fills in the title, who said it and a note.
The model proposes, the person commits
Nothing saves until the person says so. Every result lands on a Ready to save card, with edit, retry and cancel one tap away. That one design choice does more for reliability than any model change. The model can get things wrong, so the person always has the last word.
What was hard
Keeping the model from recommending. A chat-trained model wants to be helpful, and in an app built on recommendations from people, helpful means inventing suggestions. My prompt opens with a restriction, not a capability.
A correct answer can still break the screen. Tell Rexx about a shampoo and the model suggests a Beauty category the person doesn’t have. At first, the app just refused. Now Rexx offers to create it, and nothing is added until the person taps Save.
Reflection
The interesting problems weren't about making the model smarter. They were about deciding what it's allowed to do, what it has to ask, and where the person's hand stays on the wheel. Every change that improved Rexx strengthened the person's position, not the model's.