Every student has had the same experience: you paste a question into Claude or Gemini, get three confident paragraphs back, and walk away feeling like you learned something — only to realise the answer was shallow, slightly off-topic, or just verbose filler. The model wasn’t broken. The prompt was. Asking an AI a better question is a learnable skill, and once you have it, these tools shift from search-engine replacements into something closer to a patient, well-read tutor.
Why Vague Prompts Produce Vague Answers
Large language models are trained to complete text plausibly. That means they respond to the shape of your question as much as its content. A vague input gives the model permission to drift — to stay at surface level, to hedge, to summarise Wikipedia rather than engage with your specific confusion.
Compare these two prompts:
“Explain opportunity cost.”
“I’m an economics undergrad who understands supply and demand but keeps confusing opportunity cost with sunk cost. Explain the difference using a decision I’d actually face as a student, then give me a one-sentence rule I can use to tell them apart.”
The second prompt does four things: it tells the model your level, names your specific confusion, requests a relevant example, and asks for a concrete output format. You haven’t dumbed the question down — you’ve made the answer space smaller and more useful.
This is the core move. Precision in → precision out.
Prompt Patterns Worth Building Into Your Workflow
These aren’t abstract principles. They’re templates you can copy and adapt.
Role + level setting. Start by telling the AI who you are and what you already know. “I’m a second-year biology student comfortable with cell division but new to epigenetics” gives the model a calibration point. Without it, you’ll get a generic explainer aimed at no one.
The confusion sandwich. State what you think you understand, then name where it breaks down, then ask the model to fix the break. This forces the AI to address your actual gap rather than repeat what you already know.
Constraint stacking. Add limits that shape the output:
- “Explain this without using jargon I haven’t defined.”
- “Keep your answer under 200 words.”
- “Give me three examples, not one.”
- “Use a table to compare these two theories.”
- “End with a question I should be able to answer if I’ve understood this.”
Constraints feel restrictive but they’re actually generative — they force the model to make choices, and choices produce cleaner thinking than unchecked prose.
Steel-manning requests. If you’re writing an essay and need to understand a position you disagree with, ask explicitly: “Give me the strongest version of the argument that free will is an illusion, written by someone who actually holds that view.” Models are good at perspective-taking when you ask for it directly.
Output format specification. Asking for a bullet summary, a step-by-step proof walk-through, a debate between two positions, or a Socratic dialogue changes what you get entirely. Match the format to your learning goal.
Checking the Answer Before You Use It
Better prompts reduce AI errors but don’t eliminate them. Models hallucinate citations, misstate dates, and occasionally invent plausible-sounding nonsense about niche topics. A few habits protect you.
First, ask the model to flag its uncertainty: “If you’re not confident about any part of this, say so.” Claude in particular will often comply honestly. Gemini tends to hedge with phrases like “you may want to verify” — don’t skip those warnings.
Second, use AI to understand, not to copy. If you ask for an explanation and then paste it into an essay, you’ve learned nothing and created an academic integrity risk. Instead, close the chat, write what you remember in your own words, then reopen it to check gaps. This is retrieval practice with AI scaffolding — it’s actually more effective than passive re-reading, as the research on spaced retrieval (Roediger and Karpicke, 2006) consistently shows.
Third, cross-reference anything that will appear in assessed work. AI is weakest on recent events, specialist subfields, and quantitative specifics. For those, go to the primary source the model points toward, or ask your librarian.
One quick checklist for any AI answer before you use it:
- Does it actually address my confusion, or just the topic generally?
- Are any claims specific enough to verify?
- Does it match what my lecture notes or textbook say about the fundamentals?
- Would I be able to explain this to a classmate without re-reading the AI response?
If the answer to that last question is no, prompt again. Ask it to simplify one layer further, or ask it to give you an analogy, or ask what you’d need to understand first before this would make sense.
What Selene does with this: every article here is built around the premise that tools work better when you understand their logic — so the advice stays close to mechanism, not magic. If a prompt pattern is worth recommending, it has to hold up when you test it yourself, on your actual coursework, tonight.