Most students type a question into Claude or Gemini the same way they’d type it into Google. That’s the problem. Search engines reward keywords. Language models reward context, constraints, and clarity. The gap between a throwaway prompt and a genuinely useful one is smaller than you think — it’s mostly about knowing which levers to pull.
Give the Model a Role, a Reader, and a Reason
The single biggest upgrade you can make to any prompt is telling the model who it’s talking to and why. Compare these two prompts:
- “Explain opportunity cost.”
- “I’m a first-year economics student who understands supply and demand but hasn’t studied welfare theory yet. Explain opportunity cost in a way that builds on what I already know, and flag where the concept gets more complicated at higher levels.”
The second prompt gives the model a reader (you, specifically), a knowledge baseline, and a goal (calibrated explanation with a signpost for future learning). The answer you get back is proportionally better.
The same logic applies when you want the model to act as a critic rather than a teacher. “Give me feedback on this paragraph” produces surface-level notes. “Read this paragraph as a sceptical marker who thinks the argument jumps too fast from evidence to conclusion — tell me exactly where that jump happens and what I’d need to add to close it” produces something you can actually work with.
Three role-and-reason moves worth bookmarking:
- The calibrated explainer: state your current level and ask it to match, then push one step beyond.
- The devil’s advocate: ask it to steelman the opposing argument to the position you’re defending.
- The confused peer: ask it to explain back your argument in simpler terms so you can spot where your logic gets muddy.
Constrain the Output Format to Force Precision
Language models are fluent. Fluency is not the same as accuracy, and it is definitely not the same as usefulness. Left unconstrained, both Claude and Gemini will produce smooth prose that sounds authoritative even when it’s hedging or, worse, confabulating detail. The fix is to specify the format in a way that makes vagueness harder to hide.
Try asking for a numbered list of claims with a confidence note next to each one. Or ask for an explanation in exactly three sentences — no more. Artificial compression forces the model to decide what actually matters rather than padding toward completeness. When you ask Claude to “explain Kant’s categorical imperative in three sentences for someone who has never studied philosophy,” the constraint does two things: it stops the model from retreating into jargon, and it makes it immediately obvious if the answer is evasive.
Format constraints are especially useful for literature reviews and exam prep. Instead of asking “what are the main debates in behavioural economics,” ask for a table with three columns: debate, key figures, and one-sentence summary of each position. You get scannable structure you can actually revise from. You also get a clear view of where the model is uncertain — gaps in a table are more visible than gaps buried in paragraphs.
One pattern that works particularly well before exams: ask the model to generate five short-answer questions on a topic at the level of your course, then answer one yourself and paste your answer back in, asking it to identify any factual errors or missing nuance. This turns a passive reading session into something closer to active retrieval practice, which Spencer and Higgins, among others, show is substantially better for retention.
Use Iterative Prompting Instead of One-Shot Queries
The students who get the most out of AI tools treat a conversation like a conversation, not a vending machine. They push back, ask for more, and redirect when the answer drifts. This matters because the first response is rarely the best one.
A productive iteration loop looks something like this: get an initial explanation, then ask “what’s the strongest objection to that view?” then ask “how would a proponent of the original view respond to that objection?” By the third or fourth exchange you have something close to a genuine dialectic, which is useful whether you’re preparing for a seminar discussion or stress-testing your own essay argument.
When the model gives you something that feels off, say so precisely. “That explanation assumes I know what marginal utility means — I don’t. Back up and define it first” is more productive than starting a new conversation. Models like Claude are designed to handle course-corrections mid-thread; use that.
A few things to keep doing regardless of prompt quality: verify any specific claim — dates, names, statistics — against a primary source or your course materials. Wharton researcher Mollick and others have written about how AI errors cluster in exactly the places where confident-sounding prose makes them hardest to notice. Good prompting reduces the error rate; it doesn’t eliminate it.
What Selene does with this: every article here is built around the kind of granular, applicable guidance that doesn’t appear in generic study skills handbooks — because the tools and the stakes keep changing. If a technique stops working, we update it.