How do you ask ChatGPT to think step by step?

When a question has several parts that depend on each other, ChatGPT can rush to an answer and get it wrong with total confidence. Asking it to think step by step slows it down. It writes out the in-between reasoning first, then the conclusion, and you can read every line of that reasoning.

The idea comes from research on what is called chain-of-thought prompting: getting a language model to show the chain of small steps that lead to an answer. In 2022, a team at the University of Tokyo and Google found that one short phrase, "Let's think step by step," was enough to trigger this in a large model, with no worked examples needed. A Beginner's Guide to ChatGPT lists the same habit among its context techniques: ask ChatGPT to reason through a problem in stages, or to weigh the pros and cons, when you want a more careful answer. Here is how to do it in an everyday chat.

Three ways to phrase it

You don't need special syntax. Any of these works, from lightest to strongest:

1. The short cue. Add "Think step by step" or "Let's work through this step by step" to the end of your question.
2. Steps plus a clean answer. "Work through this one step at a time. Then write the final answer on its own line, starting with 'Answer:'." This keeps the reasoning visible but makes the result easy to find.
3. Steps you define. For a task you know well, name the stages yourself: "First list the facts given. Then say what each clue rules out. Then check your answer against every clue." You are giving ChatGPT your method, not just asking it to have one.

The third version is often the most reliable. The Tokyo and Google team reported that carefully written worked examples of reasoning still beat their bare cue, and naming the steps is a light way of giving that kind of guidance.

A worked example: a scheduling puzzle

Here is a small logic problem, first asked the quick way.

Maya, Leo and Sam each take one shift: Monday, Wednesday or Friday.
Leo can't work Monday. Leo doesn't work Friday. Maya doesn't work Wednesday.
Sam's shift is later in the week than Maya's.
Who works which day? Answer with names and days only.

"Names and days only" forces an instant reply. A model answering this way may get it right, or may put Sam on Wednesday without noticing that it leaves no valid day for Leo. Now the same puzzle with defined steps:

Maya, Leo and Sam each take one shift: Monday, Wednesday or Friday.
Leo can't work Monday. Leo doesn't work Friday. Maya doesn't work Wednesday.
Sam's shift is later in the week than Maya's.

Work through this step by step: take each clue in turn, say what it rules out,
then give the schedule. Finally, check the schedule against every clue
and write the result on its own line starting with "Answer:".

A good reply follows a path like this one (illustrative):

The check step matters as much as the steps themselves. If a line of reasoning is wrong, you can point to it: "Step 3 is wrong because…" and ask ChatGPT to redo it from there.

When does it help, and when not?

It helps most when an answer depends on a chain of facts: arithmetic word problems, logic puzzles, schedules, budgets, comparing options against several criteria, or debugging a sequence of events ("the invoice was sent, then the price changed, then…"). In the 2022 Tokyo and Google tests, adding "Let's think step by step" lifted a large model's score on one set of multi-step arithmetic problems from 17.7% to 78.7%. The steps also give you something to verify, which is often worth more than the accuracy gain itself.

It helps little, or not at all, in three situations. The first is a simple lookup, such as "What is the capital of Portugal?" or "What does ROI stand for?" There is only one step, so asking for more just pads the reply. An earlier Google Research paper on chain-of-thought prompting saw the same pattern: on one-step problems, the gains were small or even negative. The second is pure writing, such as a thank-you note or a product description. There is no hidden chain to uncover, and you are better off describing tone, length and audience. The third is checking facts. Step-by-step reasoning can be perfectly logical and still rest on a false fact, and a confident series of steps can make a wrong answer look more convincing, not less.

So use the cue as a tool, not a reflex. Reach for it when a question has more than one moving part, define the steps when you know them, ask for a final check, and still read every line before you rely on the result.