Questions to Ask GPT
Twenty follow up questions to send a chatbot after it gives you an answer, aimed at surfacing hidden assumptions, missing caveats, and the parts it is least sure about. Written for people using these tools for work they will be held responsible for.
The questions
Open any question for the note
Before you answer, what do you need to know from me?
Why ask it
Asked first, this converts a guess into a scoped answer. The questions that come back tell you what you left out, and they are usually the constraints that actually determine the answer: budget, audience, deadline, jurisdiction.
Summarise what you think I am asking, then answer it.
Why ask it
Catches misreadings before they cost you a long answer in the wrong direction. When the summary is subtly off, that is nearly always because the original request contained two requests, and the model picked one.
What did you assume that I never told you?
Why ask it
The most reliably useful question in this list. Answers surface defaults you did not choose: an American reader, a company of a certain size, a current year, a legal regime. Any one of these can make a confident answer wrong for your situation.
Give me a concrete example, with numbers.
Why ask it
Abstract advice hides whether the reasoning works. Demanding figures forces the answer into a form you can sanity check, and arithmetic that does not add up is the fastest available signal that the surrounding explanation is unreliable.
What is the simplest version of this that still works?
Why ask it
These tools tend to produce thorough rather than minimal answers, because thoroughness reads as competence. This question strips a twelve step plan back to the three steps that carry the result, which is usually what you needed.
Which parts of that are you least confident about?
Why ask it
Useful as a pointer, not as a measurement. Stated confidence from a model is not calibrated and should never be treated as a probability, but the parts it flags are a reasonable list of what to verify first.
What could be wrong here in a way I would not notice?
Why ask it
Better than asking whether the answer is correct, which invites agreement. This phrasing asks for failure modes that are invisible from your side: an outdated rule, a plausible looking citation, a formula that works only in a special case.
Where would I check this myself?
Why ask it
Ask for the type of source rather than a link, because generated references are frequently inaccurate or invented. Answers naming a specific statute, standard, or official register are checkable; answers naming a study with authors and a year need verifying before you repeat them.
What is the standard advice on this, and when does it fail?
Why ask it
The failure conditions are where the value is, because the standard advice is what you already suspected. If the exceptions cannot be named, the topic is probably being reproduced from consensus writing rather than reasoned through.
Argue the opposite case as strongly as you can.
Why ask it
These systems tend to agree with whatever the user seems to want, so the counter case rarely appears unless requested. A weak or strawmanned opposing argument is itself informative: it usually means the first answer was written to please you.
What objections would a specialist in this field raise?
Why ask it
Shifts the frame from general helpfulness to professional scrutiny, which changes the register of the answer. It is also the cheapest available preview of the questions you will be asked when you present this to someone who knows the area.
Give me three options and say which one you would choose, and why.
Why ask it
Forcing a recommendation prevents the balanced summary that leaves the decision exactly where it started. The stated reason is the part to interrogate, since the ranking is often driven by what is most commonly written about rather than by your constraints.
What would change your recommendation?
Why ask it
Turns an answer into a conditional you can actually use, and it exposes soft reasoning quickly. If nothing at all would change the recommendation, the answer is not analysis but a general position that happens to be pointed at your question.
What did you leave out to keep that short?
Why ask it
Length limits cause omissions that are never flagged, and the omitted material is often the caveat that matters. Asking specifically about the cut, rather than for more detail, gets you the exceptions instead of more prose.
Here is my draft. What is most likely wrong with it?
Why ask it
Handing over your own work rather than asking for new work is where these tools are most useful and least used. Expect the first pass to be diplomatic; asking for the three most serious problems only gets you past the compliments.
Rewrite that for someone who will read only the first two sentences.
Why ask it
A real constraint, since most of what you send at work is skimmed. It also tests whether the answer has a conclusion at all, because material that cannot survive this compression usually had no argument underneath it.
If I do nothing about this for six months, what happens?
Why ask it
Introduces time, which advice usually omits. Some problems decay expensively and some resolve themselves, and the difference decides whether this is worth your week. Answers that treat every issue as urgent are worth discounting.
What do people most often regret about this decision?
Why ask it
Asks about outcomes rather than options, which surfaces material that structured advice tends to skip: underestimated ongoing costs, an exit that turned out to be difficult, a commitment that was hard to reverse.
Am I asking the wrong question? What should I be asking instead?
Why ask it
Worth asking whenever a topic is unfamiliar, because these tools will answer a poorly framed question rather than challenge it. The suggested reframing is often the actual output of the conversation.
Cut this by half. What goes, and what did you protect?
Why ask it
The second half is what makes it a good closing question. What survives the cut is the model's own view of the core argument, and comparing that to what you thought was central tells you whether you were both working on the same problem.
Getting Usable Work Out of a Chatbot
Practical guidance for the conversation itself
How to Ask
Treat the First Answer as a Draft
Almost none of the value is in the first response. It arrives in the follow ups: what was assumed, what was left out, what the opposing case is. If you accept the first answer, you are using a very capable tool as a search box.
Give the Constraints Up Front
Audience, length, format, deadline, budget, jurisdiction, what has already been tried. Every constraint you withhold is one the model will invent, and invented constraints are the main reason answers feel generic.
Bring Your Own Material
Reviewing your draft, your data, or your contract produces far better results than asking for something from nothing, because the work is anchored to facts you supplied and can check.
Ask One Thing per Turn
Compound requests get compound answers where the weak part hides inside the strong part. Splitting them makes it obvious which half was thin.
Questions It Cannot Answer Reliably
- Anything about its own capabilities, limits, or training data. These answers are generated like any other text and are frequently wrong about the system producing them.
- Requests for citations, page numbers, or quotations. Plausible references that do not exist are a well documented failure mode, so check every one before you repeat it.
- How confident it is, as a number. Stated percentages are not calibrated and should not be treated as probabilities.
- Current facts: prices, availability, who holds an office, what a law says today. Without a live source attached, the answer reflects a fixed training cutoff.
- Arithmetic or dates done in prose. Ask it to work step by step, then check the numbers yourself.
- Whether it is right. Asking a model to grade its own answer tends to produce agreement rather than review.
Where People Get Burned
Mistaking Fluency for Accuracy
These systems are equally articulate when correct and when wrong, so the usual signals of an unreliable source are absent. Confidence in the prose carries no information about the truth of the content.
Asking Leading Questions
Phrasing that signals the answer you want will usually get it. If you ask whether your plan is sound, expect reassurance. Ask what is wrong with it instead, and ask for the three worst problems by name.
Pasting in What You Should Not
Client data, patient records, unreleased financials, and credentials should not go into a consumer tool. Check whether your organisation has an approved deployment before pasting anything you would not put in an email to a stranger.
Using It Where You Carry the Liability
Medical, legal, tax, and safety questions still need a professional. A chatbot is useful for preparing your questions and understanding the vocabulary before that appointment, not for replacing it.