Have AI wait until I write three lines, then give the answer
Predict how it will move at the boundary, then have it get something wrong on purpose — you point it out.
One more scroll and the answer sits there in full. Your finger rests on the wheel. The screen is waiting. This second you can still guess. After you scroll, that chance is gone.
You are checking this month’s electricity bill for the family: total use is 210 kWh, and the first tier caps at 200 kWh. The AI puts the question above the answer.
Once you pass the first tier, does all 210 kWh switch to the second-tier rate, or only the extra 10 kWh?
You vaguely lean toward the second, but you don’t write it down. You feel one confirmation away, so you flick the wheel. The answer says: only the extra 10 kWh is billed at the second tier; the first 200 stay at the first. You read it for 7 seconds and jot, “Oh right. That’s what I thought.”
At dinner someone holds up the same bill and asks why it jumped the moment you passed 200 kWh. You say at once: “Cross the tier, and all 210 kWh go to the second rate.” The line items sit right there. You just said the opposite of the answer you read.
You go back for your first thought. It isn’t there. The on-screen answer already handed you a direction, the wording, and the result. Recall follows those ready-made cues. “That’s what I thought” and “that’s what I thought after I saw it” get mixed.
If you had paused one second and written the result, the rule, and when it fails, you would see which line the error sits on. The three lines are evidence from the you before the answer walked in. That evidence may be wrong. Wrong is useful — it tells you what to patch next.
Once you have seen the answer, you can never know how much you would have thought of on your own. Every sentence after that may already have borrowed its words. The chance to check what you currently understand exists only before the answer appears.
Write it first, and the page keeps a version the answer has not edited. When it is revealed, you can check line by line: did you guess the result, name the rule, and find its range? The line that misses is the line you patch next.
The AI can write a boundary problem for you, and it can hide the answer. The key is to put the order in the request, and make it wait until you have handed in three lines.
- Show only the problem first. Hide the answer, the hints, and the keywords.
- Wait until you have written three lines. In between, only confirm receipt. Do not correct you along the way.
- Then compare line by line. Point out where the prediction, the rule, and the failure condition each miss.
What is each of the three lines actually checking, and what extra layer do you get by writing them before the answer? Let’s run that electricity-bill problem.
Line one writes the result. It keeps the prediction you made with no hint, and turns “I kind of get it” into a sentence you can mark right or wrong.
Line two writes the rule. You can luck into the result. Naming which rule is doing the work shows whether you actually ran the calculation.
Line three writes when it fails. Every rule has a range. Change the billing cycle, a shared meter, or a local rule, and the old cutoffs may move. Line three checks whether you understand a rule, or only memorized this problem’s answer.
Send the next two sentences in order. The first locks the answer. The second does the comparison. Swap the material name and the concept for what you are studying.
- Set the problem and hide the answer
Write a problem with a boundary condition about 【concept】. Show only the problem first. No answer, hint, or keyword. I will write three lines first: what I think will happen; which rule I am using; under what condition that rule fails. Until I have sent the three lines, only confirm receipt.
- After I write, compare line by line
Now reveal the answer, and compare it line by line with what I just wrote. Point out where the result, the rule, and the failure condition match, and where a step is missing. Quote my words. Do not rewrite them into a fully correct version for me.
“Until I have sent the three lines, only confirm receipt” is the line that matters. Drop it, and the AI will happily hint while you write line one. The original judgment gets edited again.
The three lines check whether you can form a judgment on your own. Flip it: have the AI hide an error inside a smooth explanation, and see if you can point to the sentence that breaks the rule.
- Hide the error in the mechanism
Deliberately tell a wrong version of 【concept】. Hide the error in the mechanism — not an obvious typo. After you finish, do not say where it is wrong. Wait until I point it out, then tell me if I am right.
Both self-checks are practical moves. Predicting keeps the version from before the answer. Pointing out the error checks whether you can see a mistake on the screen. For walking through your thinking after you already got it wrong, see have AI find the step I got wrong. To keep drilling changed conditions, see solve the variants.