What I Actually Fixed with AI This Week
This week, the stuff I did with AI wasn't exactly flashy.
Mostly I was going through real problems in a live product, the kind that hurt the user experience, slow down shipping, or turn into maintenance headaches later, and taking them apart one by one.
That's also my biggest takeaway from working with AI lately: it's not just a code-writing tool. It's more like a dev partner that helps me turn a messy problem into a clear plan of action.
A tiny copy change, a real-world workflow
MakanCloud has a POS scan-to-order flow for table service where customers order first and pay later.
The old confirmation message just said "Order placed." That's technically correct, but it doesn't match how things actually work in the restaurant. After ordering by QR code, the customer still has to go to the counter to pay. If the system only says "success," someone at the front desk or a server ends up having to explain the next step.
So I had AI help me change it to "Order placed. Please pay at the counter." and update the Chinese and English versions together.
It looks like a tiny change, but it closes a gap between the product and what actually happens on the restaurant floor.
You can't eyeball a permissions bug
The more serious issue: users whose subscriptions had lapsed could still log in and use everything as normal.
That's not a UI bug. It's about where the product's boundaries are and how access control works. So instead of just telling AI to fix the code, I first had it help me work out the next steps:
- Where the subscription status needs to be checked
- Which entry points might get around the restriction
- How to verify the fix once it's in
Once the problem was clearly broken down, I moved on to fixing, committing, tagging, and deploying.
This round convinced me of something: once a problem gets even a little complex, AI is most useful when it helps me build an actionable checklist first, not when it jumps straight to an answer.
Scan-to-order refresh is a mobile UX problem
Another issue came up in scan-to-order. When several phones scanned the same table's QR code, the shared order didn't refresh on its own. And on iPhones there was no obvious way to refresh the page by hand.
It's easy to write this off as a minor bug. But for a group ordering together, it breaks things: A orders a dish and B can't see it. The page doesn't update, and nobody knows what to do about it.
So this fix was about handling both sides: automatic refresh plus a manual fallback, not just patching the symptom.
For a new project, design the collaboration first
Outside of MakanCloud, I also used AI to design the product framework for a fortune-telling website/app.
I didn't rush AI into building full features this time. Instead, I had it produce framework docs for the rest of development: what to do at each step, how to test it, how to sign off on it, and how the next AI session picks up where the last one left off.
While I was at it, I also fixed a page display bug and filled in a long list of Chinese cities that were missing from the birthplace options.
The lesson for me: if you're going to build a project with AI over the long haul, the first thing to build isn't any particular page. It's the way you work together.
Wrapping up
Looking back on the week, my answer to "what have I been doing with AI lately?" is pretty down-to-earth: I fixed a prompt message, a permissions hole, and a refresh bug, and I set up a framework for a new project.
None of these are flashy features, but every one of them solves a real problem.
If you're using AI to build products or write code too, drop a comment: what do you most want AI to help with right now? Writing code, debugging, or figuring out what to do next?