How to make ChatGPT and Gemini actually useful for JEE doubts
A student opens a Mains-level Integration problem. They try King’s Rule. The substitution does not simplify. They paste the problem into ChatGPT and ask for help. The reply is technically correct, but it lands a level above where they are stuck — it assumes a step they had not yet considered. They ask a follow-up to clarify. The next reply references a substitution they have not seen, so they ask another follow-up. Twenty-five minutes pass. They never finish the problem. They feel less confident than when they started.
The AI is not the problem here. ChatGPT, Gemini, and Claude have become genuinely capable on JEE-level mathematics. The problem is the interaction — what the student told the AI, and what the AI then had to assume. Without context about you — the concept you are working on, the technique you already tried, the prerequisite gaps you carry into the problem — even a smart model has to guess at the right level. A guess plus three follow-ups is how 25 minutes disappear.
Why follow-ups make it worse
Modern LLMs preserve the original problem across follow-ups much better than they used to — large context windows mean the question is still in working memory. The trap is something different.
When you say “I don’t understand,” the AI’s default response is to make the explanation more thorough, not differently framed. It adds depth where you needed simplification. It introduces new vocabulary when you wanted the same vocabulary repeated more carefully. It pulls in adjacent concepts to be helpful. Each escalation adds new techniques, new notation, new dependencies on side-concepts. The student now has more vocabulary to be confused by, not less.
The deeper miss is that the AI does not know you. It knows the problem perfectly. It does not know what you already understand, what notation feels natural to you, which side-techniques you have not learned yet, or what the actual gap is between your reasoning and the step you stopped at. Without that, every explanation is calibrated for an average student — not for you.
The student rarely notices this happening. From their side it feels like “this should be working — let me try one more follow-up.” The trap compounds. Three follow-ups in, they are no longer debugging the problem. They are debugging the AI.
That is the rabbit hole. It is not a problem with the model. Modern LLMs are remarkably capable, and getting better fast. It is a problem with the missing student context — something no model can fill on its own.
The three-input rule
Most JEE students give AI one input: the problem. That single input leaves the model to guess everything else — your level, your attempt, what you already understand.
The three inputs that turn a generic AI response into a useful answer:
Input 1 — The concept being tested. Not “I cannot solve this integral” but “I am trying to apply King’s Rule but my substitution is not simplifying.” Name what the problem is testing before you paste it. If you cannot name the concept, that is itself a more useful diagnosis than anything AI will give you — it means the gap is concept-recognition, not concept-execution. Go to a textbook, not a chatbot.
Input 2 — The technique you have already tried. AI without this just suggests Plan A — which is exactly the approach you tried that did not work. Telling it “I tried the substitution u = π − x and got stuck at the denominator (sin x + cos x)² simplification” changes the AI’s response from a re-derivation to an actual diagnosis. It can now identify where your specific approach went off course.
Input 3 — What you already understand. AI cannot tell what level to pitch the explanation at. Saying “I am comfortable with By-Parts and standard substitutions” lets it skip the parts you do not need and focus on the gap.
A side-by-side comparison:
Before: “How do I solve this integral?”
After: “I am working on a JEE Mains Integration problem testing King’s Rule. I tried the substitution u = π − x but the denominator (sin x + cos x)² is not simplifying. I am comfortable with By-Parts and standard substitutions. What am I missing structurally?”
The first prompt starts a 25-minute rabbit hole. The second prompt gets a useful answer in one turn, because the AI now has enough context to actually help.
The one-follow-up rule
Even with a structured prompt, sometimes the first answer does not land. The enforcement rule:
You get exactly one follow-up question. If that does not work, stop.
Why a hard limit:
- After two follow-ups, you are debugging the AI, not the problem.
- The time cost compounds exponentially. The third follow-up rarely produces clarity; it produces a new tangent that demands further explanation.
- Stopping is the high-value move. A 10-minute AI attempt that does not land is not a failure. It tells you “this technique is genuinely a gap, not a clarification issue.” That is information you can act on.
What to do when you stop:
- Go to a structured solution source — a textbook, a teacher, a known-good video — that walks through the actual technique step by step.
- Note the gap. “King’s Rule with non-trivial denominators” goes on your weak-techniques list.
- Come back to AI later, after working through the technique cleanly. AI works much better when the gap is concept-vocabulary, not concept-existence.
The trap is the open-ended follow-up loop. The escape is a hard cap.
AI for intuition. Solutions for verification.
The reframing that makes the rule feel obvious:
AI is excellent at concept reframing. Intuition. Alternate explanations. “What is this technique really doing.” Use it for understanding why.
AI is less reliable on step-level rigor than on conceptual explanation — sign tracking, exact substitutions, the precise manipulation that JEE specifically rewards. Frontier reasoning models have closed a lot of this gap and are still closing it fast. But for now, when you need certainty on a step, the safer source is still written ground truth: NCERT exemplars, textbook worked examples, official answer keys.
Most students lean on AI for both jobs and use it for steps because written solutions feel slower and more tedious. That is the source of half the rabbit-holes — the student is asking AI to do the job where it is currently less reliable, then arguing with the answer.
Two tools, two jobs. AI for intuition. Solutions for verification. Use each for what it is currently best at.
As AI keeps improving at step-level rigor — and it will — this rule will soften. The three-input rule from earlier will keep mattering, though, because no matter how capable the model gets, it still cannot know you without you telling it. Better prompts produce better answers from every model, and that gap will grow, not shrink.
What Rhovecs does
When a student uses Rhovecs, the deterministic engine already knows: which concept they are practising, which prerequisites are unstable, which mistakes they have made recently, and which techniques they have mastered.
When they ask for help on a problem, Rhovecs constructs the prompt for the AI assistant with all of that context — the three inputs from the rule, already filled in. The student does not have to remember to add them. The AI then works with real information, not just what the student typed into a box.
The first response usually lands. There is no follow-up trap to fall into because the prompt was right the first time.
And as AI itself improves, this gets better, not worse. Better models reward better prompts more strongly than weaker ones — context-rich prompts have outsized returns on smarter LLMs. The deterministic engine that constructs the context is the part that compounds; the AI on the other end can be whichever model is best this month.
For the deeper version of this argument, see Why Rhovecs exists and the operational walk-through of the four-step loop.
Rhovecs’s 10-day full trial lets you see this in your own practice — every doubt prompt arrives at the AI with the right context already built in. No card required. Start at jee.rhovecs.com.