Patterns that work, ways to check what comes back, and the mistakes to avoid. A good prompt gets you a better draft. Checking is what makes it trustworthy.
Anatomy of a strong prompt
Context: [who I am, what this is for] Task: [the one thing I want] Format: [how the answer should look] Limits: [what to avoid, length, level] Check: [list the claims I should verify and say where you are unsure]
You won't need all five parts every time. A quick factual question needs only Task and Check. A big piece of work needs all five.
The moves
I'm a high school student preparing a 5-minute talk for a class that has not studied this topic. Explain how a recommendation algorithm works, at that level, with one everyday example.
Why: The model can only see what you tell it. A vague prompt gets an answer for nobody in particular.
Here is a sample of the style I want: [paste 2 or 3 lines] Write three more in that style about [topic].
Why: An example beats a description. Use your own work or text you have permission to use.
I need to plan [project]. First list the steps in order and stop. I'll tell you which one to do next.
Why: Long, multi-part prompts are where instructions get dropped. Smaller steps are easier to check.
Compare [A] and [B] on cost, time, and risk in a table. Mark anything you are guessing with a question mark.
Why: You choose the criteria, so the answer reflects what matters to you. Then check the facts in the table.
Here is my draft: [paste] Don't rewrite it. Name the three weakest points and explain why each is weak.
Why: This keeps your thinking in the work. It matches the By Hand, Then With AI habit: you go first.
Here is my argument: [paste] Give the strongest case against it, and tell me which part of my argument it hurts most.
Why: If you can't answer the best objection, your argument isn't finished.
Quiz me on [topic], one question at a time. Wait for my answer before telling me if I'm right, then explain.
Why: Explaining and retrieving builds memory. Reading an AI summary mostly does not.
Answer, then list each factual claim I should verify and what kind of source would confirm it. Say where you are unsure.
Why: The model may still be wrong about its own doubts, so you still check. But the list gives you a place to start.
What important things might this answer be leaving out, and what questions should I ask before I trust it?
Why: A fluent answer hides gaps. This pulls them into view.
Why: Find the actual page and find the claim in it. Fake citations look real. If you can't find it, treat it as made up.
Why: Recompute it yourself or with a calculator. Models are not calculators, and a wrong digit sounds as sure as a right one.
Why: Start a new chat and ask differently. If the answers disagree, one of them is wrong and you need a real source.
Why: A model's knowledge stops at a cutoff. Look up current facts in a source with a date on it.
Why: Tone is not evidence. The more fluent and specific an answer is, the more tempting it is to skip checking. Don't.
Why: Chatbots often agree with you or praise your work. Ask for flaws specifically, and be suspicious of pure praise.
Summarize what we've decided so far in five bullets so I can start a fresh chat.
Why: Instructions from early in a chat get lost. Start fresh with a short summary.
Why: Anything an AI reads can contain hidden instructions (prompt injection). Treat pasted material as untrusted, and never let a tool act on it without your review.
Why: Keep names, addresses, passwords, grades and other people's information out of prompts unless your teacher has approved the tool for that.
Why: If you can't explain it without the screen, it isn't yours yet, and in this class that is plagiarism. Log how you used AI.
Use AI to learn, not to skip learning. Record what you used it for in your AI Use Log, including what you checked. Check the glossary if a term is new.