What it actually means
When you ask a model a hard question and it just blurts the answer, it often skips steps. It jumps to a guess that sounds right but is not. Chain-of-thought prompting fixes this by telling the model to slow down and show its work.
In practice it is one added line, usually something like “think step by step” or “work through this before you answer”. That prompt makes the model reason out loud, and reasoning out loud is where it catches its own mistakes.
The same question, two ways
Without chain-of-thought
“A shirt costs $40. It is 25% off, then another $5 coupon. What do I pay?”
The model may rush and apply the coupon first, or drop a step, and land on the wrong number.
With chain-of-thought
“A shirt costs $40. It is 25% off, then another $5 coupon. Work through it step by step, then give the final price.”
Now it shows: $40 minus 25% is $30, minus the $5 coupon is $25. The steps make the right answer hard to miss.
When it helps, and when it does not
Great for
Math, logic, planning, comparing options, anything with two or more steps. These are where skipped reasoning causes wrong answers.
Pointless for
Simple lookups, short facts, or quick rewrites. Asking the model to "think step by step" about a one-line answer just makes it longer.
A middle case
For a final answer you want clean, ask it to reason first, then give only the result. You get the accuracy without the wall of working.
Why showing the steps works
A model builds its answer one piece at a time, each piece based on what came before. When it writes the reasoning out, each step becomes something the next step can lean on. The answer is built on a visible chain instead of a single guess.
There is a second benefit for you. When the steps are on the page, you can spot where the logic went wrong. A hidden answer you can only accept or reject. A reasoned one you can check and correct.
Phrases you can copy
You do not need special wording. Any line that tells the model to slow down and reason first will do. Here are a few that work well, depending on what you want.
Simple and general
"Think step by step, then give your final answer."
When you want a clean result
"Work through this carefully, then give only the final answer on its own line."
For a decision
"List the options, weigh the pros and cons of each, then recommend one and say why."
For checking work
"Solve it, then check your own answer a second way and tell me if the two match."
Two ways to do it
There are two flavors of chain-of-thought, and it is useful to know the difference.
The first is just asking. You add “think step by step” and let the model reason on its own. This is the easy, everyday version, and it is usually enough.
The second is showing an example. You include one worked problem, with the reasoning written out, before your real question. The model copies that style of thinking. This takes more effort, but it helps when the problem has a specific method you want followed. If you want to go deeper on this, it is the same idea as few-shot prompting.
Common mistakes
Chain-of-thought is simple, but a few habits waste its value.
- → Using it on everything. On a simple lookup it just adds length and slows you down.
- → Trusting the steps blindly. Reasoning that looks tidy can still reach a wrong answer. Read it, do not just admire it.
- → Forgetting to ask for the final answer. Without it, you get a wall of working and no clear result.
- → Burying the request. Put the “think step by step” instruction near the task, not lost in a long paragraph.