How to make ChatGPT ask you questions instead of answering

Way 1: one line, when you need it

Type your question, then your own attempt, then the line. In that order.

Why does ice float on water?

My own attempt first: I think ice is lighter because the water spreads out when it freezes.

Don't give me the answer. Guide me with questions, one at a time.

Two things make this work better than the line on its own.

• Your attempt gives the AI something to coach. Without it, the first question you get back is usually "what do you already know?", which costs you a turn.
• "One at a time" matters. Without it you tend to get a list of five questions, which reads like homework and is easy to ignore.

If you want it to check your thinking instead of coaching you, swap the last line for: "Check my thinking: tell me what I got right, then what I missed."

Way 2: save it once, for every chat

ChatGPT, Claude and Gemini each have a place for standing instructions. Text you save there is applied to new chats, in the phone apps as well as the website. This is the gentle version: it asks for your attempt when you are learning something and just answers quick facts.

How I want you to work with me: I am practising thinking for myself before I lean on you.

When I am clearly trying to learn or work something out, ask for my own attempt first: my guess, my rough draft, or what I already know. For quick facts and small practical requests, just answer.

When I share my attempt, start with what I got right. Then point to one gap and ask one question that helps me close it. Give hints before answers, one step at a time.

Give the full answer once I have tried, or whenever I write "just answer". Then answer directly, with no lecture.

The coach setup page builds this for you, with a firmer version, versions for students, work, coding and writing, and the steps for where to paste it in each AI.

Way 3: have it added for you

The weak point of ways 1 and 2 is the same: on a tired afternoon you skip the line, or you type "just answer" before you have tried. The browser extension puts a small card in front of the Send button that asks for your attempt and adds the line with one tap. On a phone, the think pad does the same job by copy and paste.

The escape hatch, and why to keep it

Always leave yourself a way to get a straight answer. In the saved instruction it is the words "just answer". A rule with no way out gets deleted the first time you are in a hurry. A rule with a way out survives, and you will be surprised how often you do not use it.

What to expect, honestly

• It drifts. In a long chat an AI can slide back into answering. If it does, paste the line again.
• It is a request, not a lock. A saved instruction can be switched off by whoever is using the account. The AI companies' own study modes can be switched off too.
• We have not yet measured how well each AI keeps to it. We plan to test the same set of questions in each one and publish the numbers. Until then, treat "it mostly works" as our experience, not as a finding.

Why bother

In a field trial with about 1,000 students, practising with an AI that gave answers lowered exam scores by 17%. A version of the same AI that gave hints largely avoided the drop. What that trial found. The line above is how you get the hint version from an ordinary chatbot.

Questions