Gains:
- Explain why step-by-step thinking improves accuracy on complex tasks
- Can use expressions and structures that guide model reasoning
- Distinguishes between situations where the technique is appropriate and unnecessary
Some tasks cannot be answered in one fell swoop; It requires several steps of thinking. There is a very simple but surprisingly effective technique for such tasks: asking the model to think step by step. This method, called "chain of thought" in English, significantly increases accuracy in complex tasks. In this unit we will learn why the technique works, how to apply it, and when to use it or not.
Why Does the Technique Work?
As we have seen in previous units, the model produces text by proceeding word by word. Each word it produces becomes the input for the next. When you ask the model for the outcome directly, it tries to "predict" that outcome in one step and skips the reasoning that needs to be done in intermediate steps; This leads to errors, especially in multi-step jobs.
When you ask the model to think first and then answer, the model puts the intermediate steps into text. These intermediate steps work like "intermediate input" and make the next step more accurate. In other words, you open up space for the model to think. It's like people putting a difficult problem on paper and solving it step by step: the work involved increases the chances of getting the right result.
How to trigger step by step?
A few simple statements guide the pattern of reasoning:
- "Think step by step."
- "Sort out the relevant information first, then decide."
- "Write down each step clearly before reaching the solution."
- "Take your time; justify each step."
These are not magic words; These are ways of telling the model "don't give the result right away, produce intermediate steps."
Where is it useful, where is it unnecessary?
Mission type
Thinking step by step
Multi-step calculus/logic problem
very useful
Justifying a decision
useful
Making inferences from long text
useful
Simple translation / one word answer
unnecessary
Short salutation/title
unnecessary
If speed is critical and the task is simple
Avoid
Rule: if the task can be done safely in one move, step-by-step thinking only adds length. But if the task involves several intermediate decisions, revealing judgment improves accuracy.
Separating Conclusion from Reasoning
There is a practical subtlety to thinking step by step: sometimes you want to see the intermediate steps (for inspection), sometimes you just want the clean result. You can manage the two like this:
- For auditing: "Write down your reasoning first, then give the final answer under 'Conclusion:'." This way, you see both the thought and the result separately.
- For clean output: "Think step by step, but write only the final result." It reasons from within the model and only gives you the result. (You cannot control judgment in the latter; in critical matters, choose the first path.)
Four Copiable Templates
1) Decision justification:
Task: [decision question]First list the factors affecting the decision. Briefly evaluate each factor.Then give one clear recommendation under the heading "Recommendation:".
2) Multi-step analysis:
Analyze the following situation step by step:1) Summarize the available data.2) List possible causes.3) Support or rule out each cause with evidence.4) State the most likely cause and the next step.Situation: [text]
3) Check/audit (testing own answer):
Solve this question. Then review your solution from the beginning: check for errors at each step and correct them if necessary. Finally, write "Verified result:". Question: [question]
4) Clean conclusion (hide reasoning):
Solve this problem by thinking through it step by step, but just give me the final one-line answer.Problem: [problem]
Weak Prompt / Strong Prompt
Weak:
Which of the 3 supplier quotes should I choose?[quotes]
Strong:
Evaluate 3 supplier quotes. Go step by step:1) Table each quote in terms of price, delivery time and warranty.2) Our priority is delivery speed; score each proposal accordingly (1-5).3) Explain the scores with justification.4) Make a single choice under the heading "Suggestion:" and justify it in one sentence.Proposals:[offers]
The second prompt model leads first to configuration, then to scoring according to priority, and finally to decision. The result will be more accurate and its justification can be checked.
Three Mini Cases
Case 1 — Budget control. When an office manager asked for the total of an 8-item expense list and the items that were over budget, when asked directly, the model omitted an item and gave the wrong total. When I was told, "First, write down each item one by one, run to them, then add them up and compare them with the budget," the total was correct and the 2 items that were in excess were clearly marked.
Case 2 — Contract risk. When a procurement professional asked about risky clauses in a contract, the model gave a superficial list without step-by-step guidance. When it was said, "Read each article one by one, if there is a risk, write why it is risky, then list the 3 most critical ones," the analysis deepened significantly and an overlooked penal provision was caught.
Case 3 — Pricing scenario. When an entrepreneur asked about the annual income impact of 3 different price packages, the direct answer was inconsistent. With step-by-step guidance (write down the assumptions → calculate each package separately → compare) the calculations became consistent; nevertheless, the results were independently verified in a spreadsheet.
Tip: If it's a critical calculation or decision, ask for the reasoning to be visible ("show steps"). Visible reasoning shows you where you went wrong and makes it easier to fix it.
Caution: Step-by-step reasoning improves accuracy but does not guarantee it. Even if the model's intermediate steps appear consistent, it can make a calculation wrong. Always verify numerical and factual results independently (calculator, source, second method).
Common mistakes
- Applying to simple task. Creating unnecessary length in a one-shot job.
- Failure to control judgment. Saying "Think step by step" and accepting the result without reading the intermediate steps.
- Not separating the outcome. Not being able to draw a usable conclusion when reasoning and the final answer are intertwined.
- Skipping verification. Accepting the numerical result without question when the steps seem logical.
- Asking for excessive steps. The 15-step push makes a simple task unnecessarily complicated.
In summary
- Step by step prompting is asking the model to write down intermediate steps before giving the result.
- Intermediate steps work as intermediate inputs and increase accuracy in multi-step tasks.
- The technique is useful in complex calculation, logic and justification; It is unnecessary in simple one-step work.
- By separating the reasoning from the final result, you can both audit and get a clean output.
- The technique increases but does not guarantee accuracy; Be sure to verify numerical and factual results.
Application task
Have a real job that requires several steps (a budget check, quote comparison or decision). First ask directly, then ask again in a "think step by step, show the steps, then write conclusion" format. Independently check the accuracy of the two outputs and note the difference of the step-by-step version.
checklist
- [ ] I can explain why making you think step by step increases accuracy.
- [ ] I can use expressions that guide the model to reasoning.
- [ ] I distinguish between tasks where technique is appropriate and where it is unnecessary.
- [ ] I can ask for the final result by separating the reasoning from it.
- [ ] I always verify the result independently.