Both will find the loophole if you leave one.
“Clean your room” gets you one sock moved from the floor to the chair. Technically, the room got cleaned. Nothing you said was false. But you meant something specific, and you did not say it, so you got the letter of the instruction and none of the spirit.
AI does the same thing, just faster and with more confidence. Give it a vague acceptance criteria and it will build something that matches the words on the page and misses the point you had in your head. It is not being difficult. It is doing exactly what a kid does: following the actual rule, not the rule you assumed was obvious.
I used to think this was an AI problem. It is not. It is a me problem that AI exposes faster than a human ever did. When I write a spec for an engineer, I get a Slack question three days later asking what I meant. When I write the same vague spec for an AI, I get a wrong answer in ten seconds. Same gap, different speed. The AI just removed the three days of buffer where I could pretend the spec was fine.
The fix is the same in both cases. Say the actual constraint instead of the thing you assumed was obvious. Not “clean your room,” but “everything off the floor, put away, not stacked on a chair.” Not “handle the edge case,” but “if the list is empty, show this exact message, not a blank screen.” The instruction gets longer and less elegant. It also stops leaving room for a technically-correct wrong answer.
I have started treating every prompt I write like I am talking to someone who will take me completely literally and has no interest in guessing what I really meant. Because that is exactly what I am talking to. A kid does not fill gaps with your intent, they fill gaps with the most convenient reading of your words. Neither does a model. The convenient reading is always the one that requires the least effort: move one sock, ignore the case you didn’t mention, ship the version that technically satisfies the ticket.
None of this means writing longer specs for the sake of length. It means finding the one sentence that was doing the work of ten unstated assumptions, and writing that sentence down. Most bad specs are not too short. They are missing the one line that was carrying all the meaning in your head and none of it on the page.
I still get this wrong constantly. I write a prompt, get back the AI equivalent of one sock on the chair, and only then notice what I forgot to say. But at least now I know where to look when the output is technically right and still wrong: not at the tool that followed my instructions, at the instructions.
Leave a Reply