Zenn guide: five checks catch AI code that runs but is silently wrong

The starting point is a specific kind of unease: the code ran, no error appeared, but the writer cannot tell whether the result is right. It separates mistakes into two kinds — ones that stop the program, which are low-risk because they force attention, and ones that keep running with a quietly different result, which it calls the high-risk kind and the one AI-written code tends to produce. Its example is a discount function where 100 and 1000 were swapped, returning 990 yen instead of 900 with nothing to flag it.

The five checks are framed as process, not expertise. Deciding the expected number before asking, probing the exact boundary and its neighbours, checking what happens with zero or one item, watching whether digit counts reveal text being treated as numbers, and comparing counts in against counts out are all things a reader can do without reading the code. These confirm only the cases tried, not overall correctness, and some decisions — such as whether zero records should show a dash, zero, or an error — belong to the person specifying the work, not the AI.

Its closing argument is about how to ask. Requests for an opinion, it says, get a reply either way, so the stronger form is to ask for actual run output, since results cannot be invented the way opinions can.