You already know you can ask AI about your coursework. The problem is that a vague question gets a vague answer — polished, well-structured, and often subtly wrong. The gap between a useful AI session and a frustrating one usually comes down to how you frame the request, not which model you use.
Give the Model a Role and a Constraint
The fastest upgrade to any prompt is telling the AI who it should act as and what limits to respect. These aren’t magic words — they’re signals that shape the kind of reasoning the model produces.
Compare these two prompts:
- Vague: “Explain the French Revolution.”
- Shaped: “You’re a historian who values structural explanations over great-man narratives. Explain the French Revolution to someone who just finished reading Burke’s Reflections. Keep it under 300 words.”
The second version does three things: it sets an intellectual stance, it anchors the explanation to something you already know, and it forces concision. Claude and Gemini both respond well to this kind of scaffolding because it reduces the surface area of possible answers.
For technical subjects, constraints matter even more. Instead of “help me understand dynamic programming,” try: “Assume I’m comfortable with recursion but keep forgetting why memoization changes the time complexity. Walk me through one example, then ask me a question to check my understanding.” That last clause — asking for a follow-up question — turns a passive explanation into a dialogue.
Useful constraint types to experiment with:
- Audience anchors: “explain this as if I already know X but not Y”
- Format locks: “give me a comparison table, not paragraphs”
- Depth limits: “one mechanism only, not a full overview”
- Perspective assignments: “argue the opposing view of what I just wrote”
- Self-checks: “flag anything you’re uncertain about”
That last one is underused. Both Claude and Gemini will hedge more honestly when you explicitly give them permission to express uncertainty.
Use the Model’s Output as a Thinking Partner, Not an Answer Machine
The real leverage in AI-assisted study isn’t getting answers — it’s getting feedback on your own thinking. This means showing your work before asking for help.
Instead of: “What’s the argument in Rawls’ difference principle?”
Try: “Here’s my understanding of Rawls’ difference principle in two sentences: [your sentences]. What am I getting right, what am I missing, and is there a standard objection I should know?”
This prompt structure forces you to retrieve and articulate before you receive — which is how memory consolidation actually works, according to cognitive load research (see Sweller, and separately, Roediger and Karpicke’s work on retrieval practice). The AI becomes a corrective mirror rather than a replacement for your thinking.
For essays, the same logic applies. Paste in your thesis and ask: “What’s the strongest possible counterargument to this thesis? Don’t soften it.” Then ask: “How would a scholar who agrees with me respond to that counterargument?” You’ve just run a Socratic drill without needing a seminar.
For problem sets, show your attempted method even if it’s wrong. “I tried solving this integral using substitution and got stuck at this step — where did my reasoning break down?” is far more useful than pasting the problem cold. The model can identify the exact point of confusion rather than reteaching from scratch.
One more pattern worth building into your workflow: the “steel-man then critique” move. Ask the model to present the strongest version of an argument or theory, then separately ask it to identify where the literature pushes back. You get both the map and the contested territory, which is exactly what a seminar discussion or an exam requires you to hold in your head at the same time.
What to Do When Answers Feel Slippery
AI models can sound authoritative while being wrong on specific details — dates, citations, technical definitions, niche case law. The confident tone is a feature of how language models work, not a signal of accuracy.
When an answer matters for graded work, build verification into the prompt itself: “Give me your answer, then list which parts I should double-check in a primary source or textbook.” This doesn’t eliminate error, but it shifts your relationship with the output from passive consumption to active audit.
For conceptual questions, a useful test is to ask the model to explain the same idea two different ways — once with an analogy, once without. If both versions are consistent and the analogy holds under scrutiny, you’re probably on solid ground. If they diverge or the analogy breaks quickly, treat the answer as a starting point for your own reading, not a conclusion.
It also helps to work with shorter exchanges rather than long single prompts. Iterating — pushing back, asking follow-ups, saying “that doesn’t match what my lecturer said, reconcile them” — produces better outputs than trying to front-load every requirement into one mega-prompt. Think of it as a conversation with a knowledgeable tutor who needs direction, not a search engine that returns one result.
Selene uses these same patterns to structure every explanation and study resource she writes — starting from what you likely already know, then building precisely to the gap. When something is contested or uncertain, she says so.