Let AI Find the Step I Got Wrong
After a miss, do not look at the answer first. Tell it what you were thinking, and have it point to the step where the reasoning bent. When you finish, you can locate the one wrong turn and patch only that.
“Route A is faster.” Saturday afternoon at the table, a practice set: 12 km. Route A does 6 km at 30, then 6 km at 60. Route B is 40 the whole way. Which is faster.
On scratch paper you wrote (30＋60)÷2＝45. 45 is bigger than 40, so A. The numbers were neat. Every step computed. You filled it in.
The system marked it wrong. The explainer said: A takes 12 minutes then 6, 18 total. B is also 18. You nodded and wrote down “split distances need split times”, then went on.
Four days later another problem swapped the speeds to 40 and 80, the other route 55 the whole way. You averaged 60 again and picked the split route. Wrong again. Last time’s sentence was still in your notes. It did not stop you.
Because you saw the right method and never kept the reasoning you actually walked. The path you took was this:The two stretches are equal, so each speed owns half; average 30 and 60 and the whole trip averages 45.The miss starts at “each speed owns half.” The slow stretch ate more time. The two speeds did not split time evenly. The answer covered the wrong result. The road that led there is still sitting there.
Here AI can do three things:
- Take your original answer and each step of thought first, and do not let the right answer rewrite memory first.
- Check the reasoning step by step, keep what was right, and mark only where it first bent.
- Write new questions only around that step, and see if you can correct it after the numbers and the scene change.
Which step took you to 45? Both paths saw the right answer — why do the results differ a few days later? The demo starts from that question.
The explainer tells you both routes take 18 minutes. That can fix this item. It does not record why you got 45. Change the numbers and you will average the two speeds again, because that old road was never pointed at.
Say the thought first, and AI has something to check step by step. Steps 1–2 are right, so it keeps them. Step 3 treats equal distance as equal time, so it stops there. What you need to fix is the first step that left the right direction. Step 4 only kept computing along step 3. Reteaching it alone does nothing.
The block below can be copied as-is. Put the problem, your answer, and the scratch work in the brackets. Clumsy notes are fine. Leave them as they were, or you will not see that reasoning path.
I will say what I was thinking. You tell me which step first bent. Problem: [paste the problem here] My answer: [the answer I picked then] This is how I thought, step by step: 1. [step 1] 2. [step 2] 3. [step 3] 4. [step 4] Do not publish the right answer yet. Check my thinking step by step. Keep the steps that were right. Name the first step that left the path, and explain why I went wrong there. Then give me two sibling questions on that step only. After I answer, tell me the right answer.
If you cannot recall the whole path, paste the scratch work, the formulas you wrote, and any reason you still remember — and mark clearly which stretch is already gone.
This ask hands over three things “what is the right answer” never does: your choice, the steps you walked, and the reason you used. AI can then tell which steps you already have, which one needs explaining, and which one condition the next question should change. Extra practice only trains the first bent step. What you already computed right does not need to be learned again from zero.
There is also the generation effect. Slamecka and Graf’s 1978 experiments saw this: try to generate an answer first, then see the full answer, and later memory is usually better. The boundary is concrete. You have to leave your own attempt. Waiting blank for the answer gives nothing to compare.
Where sibling questions come from: have it write a new question. The cost of grabbing the conclusion: Let AI give the reasoning back to me.