Which tool to choose
Two paths, depending on the model you use. On the more recent reasoning-oriented models, there's a dedicated mode (called "reasoning," "extended thinking" or similar): turn it on from the interface for hard problems and the model produces a chain of steps on its own before answering. On normal models or the free version, you get the same effect with the prompt: you ask for it in words. The technique is the same; the only difference is whether you activate it with a button or with a sentence. If you can't find the button, always use the prompt: it works everywhere.
How to do it
- Write the problem out completely. Give all the data and constraints in the same message: the AI reasons well only on what it has in front of it.
- Explicitly ask for the steps before the answer. This is the central move. The operational syntax:
Solve this problem reasoning step by step.
Show every step of the reasoning before reaching the conclusion.
At the end, write the final answer on a separate line beginning with "Answer:".
- For choice problems, ask it to evaluate the options one by one. That way the AI doesn't jump to a hasty conclusion:
List the possible options. For each, write pros and cons with respect to my constraints.
Only after examining all the options, state which one you choose and why.
- Read the steps, not just the conclusion. The real advantage is that now you can see where it reasons badly and correct it. If a step is wrong, point it out and have it redo from there.
- For calculations, have it redo the check. Ask: "Recheck the calculations by redoing them independently and tell me whether the result matches."
A concrete example
You have to split a group expense: 4 people, but one paid the deposit, two had the full menu and one only the first course. Asking "how much does each owe?" and nothing more, the AI often gets it wrong because it attempts the answer on the fly. Asking "reason step by step: first calculate the total, then how much the person who put down the deposit has already paid, then each person's share based on what they consumed, finally everyone's final balance," the AI lays out the steps. At that point you see it forgot to deduct the deposit, you point it out, and it corrects. Without the steps, the error would have stayed invisible inside a bare number.
When it does NOT work (and how to fix it)
If the AI shows steps but the conclusion stays wrong
Sometimes the reasoning is a façade and the answer doesn't really rest on it. Fix: ask "reread your own steps and check that the final answer follows from them; if it doesn't, correct the answer." Forcing it to compare conclusion and steps recovers many errors.
If the steps are very long and rambling
On simple problems, asking for extended reasoning wastes time and can confuse. Fix: use step-by-step only where it's needed (multiple steps, constraints, calculations). For direct questions, let it answer right away.
If you're using a model with automatic reasoning and don't see the steps
Some models reason internally and only show you the answer. Fix: ask "explain the reasoning you followed to reach this answer," so it makes the path visible and you can check it.
A tip from someone who actually uses it
Step-by-step isn't there to make the AI look smarter: it's there for you, to see where it goes wrong. The value isn't in the chain of reasoning itself, but in the fact that it becomes checkable. Read the steps with suspicion, look for the weak one, and correct from there. An AI that reasons out loud is an AI you can correct; one that gives only the final number is not.
Frequently asked questions
Do I have to say it every time or does it figure it out on its own?
On normal models you have to ask every time, or set it once in the personal instructions if the tool allows it. On models with automatic reasoning they do it on their own for hard problems, but asking to show the steps remains useful so you can check them.
Does making the AI reason step by step consume more?
Yes: more steps mean more text produced, hence more time and, on pay-as-you-go plans, more cost. That's why it's worth reserving it for problems that deserve it. On a simple question it's a waste; on a problem with constraints, it pays back in accuracy.
If the AI reasons step by step, then the answer is surely correct?
No, and it's the illusion to avoid. Visible reasoning reduces errors and makes them spottable, but it doesn't eliminate them: the AI can follow convincing and wrong steps. Step-by-step gives you the chance to check, not a guarantee of the result. On the data that matters, the verification is still yours.