Asking an AI a question is easy. Getting an answer you can actually learn from is a different skill. The gap between a vague, padded response and a precise, course-relevant one almost always comes down to how the prompt was written — not the model’s capability. A few structural habits close that gap fast.
Set the Stage Before You Ask Anything
AI models respond to context the same way a tutor does. Drop them into a conversation cold and you get a generic answer aimed at nobody. Give them a frame and the answer sharpens immediately.
The most reliable framing technique is the role + level + goal pattern. Instead of asking “explain monetary policy,” try: “I’m a second-year economics student who understands supply and demand but hasn’t studied central banking yet. Explain how open market operations affect inflation, and flag any terms I’ll need to look up before my exam.”
That single shift does three things: it stops the model from over-explaining basics you already know, it stops it from assuming expertise you don’t have, and it gives it a concrete job — flag unfamiliar terms — rather than just “explain.”
Gemini and Claude both handle this well, but Claude in particular responds strongly to explicit level-setting. Researchers at Anthropic have noted that their model treats stated context as a genuine constraint rather than decoration, so the more specific you are, the more calibrated the output.
A second useful frame is perspective-setting. Ask Claude to explain a contested historical interpretation “as a materialist historian would” versus “as an idealist historian would,” then compare. You’ve just generated comparative notes in two minutes that would take an hour to pull from two separate textbook chapters.
Constrain the Answer, Not Just the Question
Most students front-load all their effort into writing the question, then accept whatever length and format arrives. Adding output constraints is where prompts go from decent to genuinely useful.
Practical constraints worth using:
- Word or sentence limits: “Explain this in under 150 words” forces the model to prioritize. If it can’t compress the idea, that’s information — the concept has more moving parts than you thought.
- Format requests: “Give me this as a comparison table” or “structure this as three claims, each with one piece of supporting evidence” turns prose into something you can scan during revision.
- Exclusion rules: “Explain the argument without using any analogies” tests whether you understand the actual mechanism, not a metaphor for it.
- Difficulty targeting: “Explain this the way it would appear on a first-year exam, then again the way it might appear in a graduate seminar.”
- Source flagging: “Tell me which parts of this answer you’re confident about and which I should verify in a textbook.” Models hallucinate less when you explicitly invite them to admit uncertainty.
The exclusion and uncertainty-flagging constraints are underused. They work because they change the model’s implicit goal. Without them, the model is optimizing for sounding complete. With them, it’s optimizing for accuracy within a stated boundary — a much better target.
One pattern worth building into habit: after any substantive answer, follow up with “What’s the most common misconception students have about this topic?” It surfaces the exact errors your exam might be testing for, and it’s the kind of question a good professor uses to structure a lecture.
Iterate Instead of Restart
Single-shot prompting — one question, one answer, done — is the weakest way to use these tools. The models maintain context across a conversation, which means you can interrogate an answer the same way you’d push back on a study partner.
Useful follow-up moves:
- Steelman the other side: “You’ve explained why Keynesians argue X. Now give me the strongest neoclassical objection to that position.”
- Ask for the edge case: “When does this principle break down or fail to apply?”
- Request a worked example with your variables: “Walk me through this formula using the specific numbers from my problem set, step by step, and tell me what each step is doing conceptually.”
- Check your own understanding: Type out your interpretation of the answer and ask, “Is my summary accurate, and what did I miss or distort?” This is essentially flashcard practice with immediate feedback.
The worked-example pattern deserves special attention for STEM courses. Claude in particular handles multi-step mathematical reasoning well when you ask it to narrate each step rather than just produce a final answer. If it makes an arithmetic error — and it does — the narration makes the error visible instead of buried.
One honest limitation: neither Claude nor Gemini should be your primary source for factual claims in essays. Use them to understand, structure, and test your thinking. Pull your citations from Jstor, your textbooks, or sources your lecturer has assigned. The model is a thinking partner, not a reference librarian.
Selene uses these exact prompt patterns when building study guides and practice questions for students — the output only becomes useful when the input is specific enough to be wrong about something. Specificity is the whole game.