I'm building Unchore, a personal AI that offers to automate the things you keep repeating. This week we added something simpler: say what app you want, and it builds it and sends you a link.
A real example from our screen recordings. The prompt, spoken into a phone:
"Build me a photo booth app. Use my camera, add soft film filters, and put four shots in a photo strip I can save."
A few minutes later there's a link to a working app.
What broke on the way
1. Long builds got cut off. Anything slightly complex took 4–5 minutes, and our worker killed the request at 280 seconds. We raised the limit to 10 minutes and retry once if the connection drops mid-build.
2. Deploys killed builds in progress. When our worker restarted for a new version, a build that was halfway done died with terminated. Same fix: one retry on a dropped connection.
3. Apps overflowed the phone screen. One generated timer put hours, minutes and seconds in one row that refused to shrink, so the page was 1,306px wide. The builder now has a hard rule: 360px phone width, no horizontal scroll, min-width: 0 on side-by-side fields.
How it's paid for
It runs on your own Claude or ChatGPT account, so we don't resell AI usage.
Happy to answer questions about the build loop.
Top comments (1)
The restart case seems worth separating from a failed build. If the app was deployed just before the connection dropped, does the retry resume the same build ID or start a second one? I'd test a disconnect after deploy but before the result link is returned: the user should get the existing link without another generation run. For the phone-width rule, an actual 360px screenshot check would catch regressions that a prompt instruction can still miss.