Things to Ask ChatGPT That Are Actually Worth Asking

Most people who open ChatGPT for the first time ask it to write something, get a flat answer, and stop there. That is one use out of several genuinely useful ones, and it is not even the best of them. What follows is five categories worth actually asking about, one example question for each, and — just as useful — what not to bother asking at all.

What it is actually doing when it answers

A language model predicts the next piece of text given everything before it, one token at a time. It has no separate step where it checks whether it actually knows something before answering — producing a plausible continuation is the whole operation. That single fact explains most of what is good and bad about it: it is genuinely strong at transforming and explaining text, and genuinely unreliable at recalling a specific fact it was never shown. What GPT stands for covers the mechanism itself in more depth if you want the fuller picture before deciding what to trust it with.

Five categories worth asking about

• —Learning something unfamiliar. "Explain how a mortgage refinance actually works, as if I have never taken out a loan before. Then quiz me with three questions to check I understood it." Asking it to test you afterwards, not just explain, matters more than it sounds — retrieval practice, being asked a question rather than re-reading an explanation, is the more durable way to learn, and a plain explanation alone does not give you that.
• —Turning a decision into something you can compare. "I am choosing between renting a bigger flat and staying put and renovating. List the real costs and trade-offs of each, and the two questions I have not thought to ask myself yet." This works because you are supplying the actual decision and its constraints — the model is organising your own inputs, not inventing new ones.
• —Tightening your own writing. Paste a paragraph you wrote and ask: "Cut this by a third without losing the main point, and tell me which sentence was doing the least work." You get a sharper draft and, more usefully, you learn which of your own habits is padding the sentence.
• —Rehearsing a conversation you are nervous about. "Play the role of my landlord, who is reluctant to fix a recurring damp problem. I will try to negotiate a repair timeline. Stay in character and push back realistically." Fun things to do with ChatGPT covers the playful end of roleplay; this is the same mechanism aimed at something with a real stake in the outcome.
• —Getting unstuck on one step of a task, not the whole task. "I am stuck on this one paragraph of a cover letter — the part explaining why I am changing careers. Here is the rest of the letter for context. Just fix this one paragraph." Handing over a whole task tends to produce something generic; handing over the specific stuck part, with context, tends not to.

A worked example: the same question, asked badly and well

Weak: "Help me learn Spanish."

Better: "I know maybe 50 words of Spanish and no grammar. Teach me the present tense of five common verbs today, explain the pattern behind the endings, then quiz me on all five without giving me the answers first."

The weak version has no starting point and no way to know if it worked. The better version gives a real baseline, a specific scope, and — again — a check at the end. Spaced practice, spreading a small amount of study over repeated sessions rather than one long one, is the other half of this: ask for five verbs today and five more tomorrow, not fifty at once.

What not to bother asking

Anything where the honest answer depends on a fact the model was never shown: your own account balance, today's exact exchange rate, whether a specific product is in stock near you. A survey of hallucination in large language models documents fluent, confident answers regardless of whether the facts behind them are real, and a specific number is exactly where that shows up. AI answers questions sorts this properly into four categories if you want the fuller test; the short version is: if the true answer lives inside your own accounts, your own inbox, or today's news, paste the actual source in rather than asking from memory.

The single biggest improvement to almost any question is being specific about who the answer is for and what you already know — OpenAI's own prompting guidance makes the same point for any task, not just these. Writing a prompt that works on the first try covers that skill properly if the five categories above leave you wanting the general method rather than more examples. What AI is actually bad at is worth reading once, so a bad answer registers as something you expected rather than something that catches you out.

What to do Monday

1. 01Pick one category above that matches something you are actually dealing with this week, not a hypothetical.
2. 02Write the question with a real, specific detail in it — your actual situation, your actual draft, your actual baseline — rather than a general version of the question.
3. 03If the category is learning, add "then quiz me" to the end of the prompt. It costs one sentence and it is the difference between reading an explanation and checking it landed.
4. 04If you are asking on behalf of work rather than yourself, ChatGPT prompts for business has five more templates built for office tasks specifically.

None of this needs to feel like a technical skill. It is closer to knowing which questions to bring to a knowledgeable but overconfident colleague — ask it the ones it is actually good at, and check the ones where it might just be guessing. Coursium teaches that habit directly, with a quiz at the end of each lesson that checks it actually stuck, the same test worth applying to ChatGPT's own answers. Stay ahead of AI by learning the tools on your phone.