Guide
Chain of thought prompting
Chain of thought prompting is a way of writing a prompt so the model writes intermediate steps before the final answer. The steps are part of the output, not a hidden setting, which makes arithmetic, logic, and multi-part decisions easier to check.
Updated 2026-09-23
Before and after
Before
Which plan is cheaper over 12 months, $20/month or $200 once?
After
Compare two prices over 12 months: Plan A is $20 per month. Plan B is $200 paid once. Work through the arithmetic in numbered steps, then put the cheaper plan and the dollar difference on a line labeled Answer. Do not round.
How to do it
1. State the decision you need
Name the exact output: a number, a yes/no with a reason, or a ranked list. Do not ask for “thoughts” with no decision.
2. Ask for the steps in order
Say “work through the steps, then give the final answer on its own line.” One explicit instruction is enough. “Think step by step” is the short form used in the Wei et al. chain-of-thought paper.
3. Separate reasoning from the answer
Request a heading such as Steps and a heading such as Answer. That makes the conclusion easy to copy and easy to audit.
4. Check the steps, not only the last line
A fluent final answer can follow a bad middle step. Read the chain. If a step is wrong, correct that step in the next prompt instead of regenerating blindly.
When to use this
Use chain of thought when the answer depends on several dependent steps: word problems, debugging, policy checks, or comparisons with trade-offs. Skip it for a one-line rewrite, a translation, or any task where a short answer is the whole deliverable. Longer reasoning costs more tokens and can still be wrong, so ask the model to label the final answer separately.
What the model is doing
A standard prompt asks for the conclusion immediately. Chain of thought asks the model to write the intermediate calculations or checks first. Wei and colleagues showed that this kind of prompting improves results on multi-step reasoning tasks, especially as models get larger. It does not guarantee correctness. It makes the path visible.
The same idea shows up in product UIs as “extended thinking” or a scratchpad. If you are prompting in a chat box, you still have to ask for the steps. The model will not invent a scratchpad unless the prompt or the product turns one on.
Chain of thought versus a longer prompt
Adding background is not chain of thought. Chain of thought is an instruction about the order of the answer: reasons first, conclusion second. You can combine it with a role, a word limit, and an output format.
If you only need the conclusion in a spreadsheet cell, ask for the steps in one block and the cell value in another. Prompt God can add that structure in one click inside ChatGPT, Claude, or Gemini when you pick a step-by-step mode.
Sources
Questions
What is chain of thought prompting?
Chain of thought prompting tells the model to write intermediate steps before the final answer so multi-step reasoning can be checked.
Is “think step by step” the same as chain of thought?
It is the short form. It works better when you also say what the steps should cover and where the final answer should appear.
When should I avoid chain of thought?
Avoid it for short rewrites, translations, and any task where extra reasoning text is not useful. It uses more tokens and can still be wrong.