FIELD NOTE / INVESTMENT APP
The app looked done.It wasn't.
A fresh review found four practical problems the first checklist had missed. The interesting part wasn't the score. It was what each problem could mean in everyday use.
A passing checklist wasn't the same as finished.
This was a planning app used by a small business. Before I called the update done, every existing automated check was passing. Those checks were useful, but they didn't cover every situation.
A separate AI reviewer started fresh, tried a few less obvious things, and found four practical problems.
One passing test answers one question. It doesn't mean you asked every important question.
This is where knowing the work mattered.
Before the review, the app ran into an empty status in older business data. It couldn't know whether that blank meant something or whether the data was broken.
I knew the answer because I knew the work: here, an empty status meant Unplanned. That wasn't programming knowledge. It was everyday business knowledge. Once that was clear, the app could handle this one blank value without accepting other values nobody understood.
The four things we missed.
The reviewer used safe copies, never the live planner. All four problems were in the earlier version and were fixed later.
01 / A failed save could still make the app look ready.
If the app couldn't save its first planner, it still showed starter information as if everything were fine. Showing an error wasn't enough; the app also had to stop pretending it was ready.
02 / A file arriving at exactly the wrong moment could be replaced.
If a real planner appeared while the app was preparing its first file, the app could overwrite it. Weird timing, serious result.
03 / An invisible character could turn “1000” into “1.”
A hidden character inside a number made the app accept only the digits before it. The input looked almost normal, but the answer was wrong.
04 / Some unusual dates were judged incorrectly.
One impossible date got through, while another perfectly valid date was rejected. These were odd examples, not proof that everyday schedules had already been harmed.
One round of fixes wasn't enough.
The first changes addressed all four problems. Then the fresh reviewer looked again and found one smaller rule that was still too loose.
We tightened that rule and checked the whole path again. The important part wasn't the final number. The app got fixed, questioned again, and improved one more time.
Then we tried it with the real planner.
After making a verified backup, I opened the real planner, checked a few familiar things, closed the app, and opened it again. The planner file was exactly the same before and after.
I kept this deliberately boring. I didn't try to break the live planner. I only wanted to answer one question: could the updated app open the existing information without changing it?
So, can a non-programmer do this?
Yes—at least further than I would have guessed. AI can turn your knowledge into working software. But it doesn't know what your business means, which mistakes would hurt, or what “correct” looks like unless you spell those things out.
You don't need a programming background to bring the most important knowledge. You do need patience, careful checks, and the habit of asking, “What could this still be getting wrong?”
I knew the work. AI helped build and review the tool. The result got better because both kinds of knowledge were in the room.
Want the nerdy details?
The original work records show 31 distinct tests passing before the problems were found and 39 passing after the final fixes.
The fresh review was performed by a separate AI reviewer working from the same version of the app. It was not an outside company or human audit.
The real-data check was limited and read-only. It showed that the existing planner stayed unchanged during that check; it did not prove that every future action would be safe.
This is one app and one round of maintenance. It does not prove that this approach finds every problem or that building software is always better or cheaper than buying it. Identifying details and private business data have been left out.