Chain-of-Thought Prompting: How to Get AI to Think Step by Step
Deliberate Academy Editorial Team
Reviewed for accuracy and professional relevance
You're 4 lessons in — don't lose your progress.
Sign up free to save where you are and earn a verified certificate when you pass.
- Explain why chain-of-thought prompting improves output quality on complex reasoning tasks
- Apply the basic chain-of-thought technique using 'think step by step' to a multi-factor decision prompt
- Construct a structured chain-of-thought prompt with explicitly defined reasoning steps for a domain-specific analytical task
- Distinguish between zero-shot and few-shot chain-of-thought and identify when each approach is appropriate
- Identify which categories of business tasks benefit from chain-of-thought and which do not
You ask an AI model to recommend which of three market expansion options your company should pursue. The model gives you a confident answer immediately. But when you check the reasoning, it has missed a critical constraint, conflated two of the options, and drawn a conclusion that does not follow from the facts you provided. Contrast this with a colleague who says: "Let me think through this. First, let me understand the constraint on capital. Then let me assess each option against that constraint. Then I'll look at execution risk before I give you a recommendation." The colleague's final answer is better because the process of reasoning is itself reliable. Chain-of-thought prompting applies this same logic to AI.
What Chain-of-Thought Prompting Is
Chain-of-thought prompting is a technique that instructs the model to reason through a problem step by step before producing a final answer. Instead of jumping to a conclusion, the model is asked to show its working — to articulate the intermediate steps of reasoning that lead to the output.
This technique was documented in research by Google in 2022 and has since become one of the most widely used and reliably effective techniques in prompt engineering. The core finding was that on complex reasoning tasks — multi-step problems, tasks requiring causal inference, logical deduction — models that showed their reasoning before answering performed significantly better than models that answered directly.
Why does this work? Because the model's output at each step becomes part of the context for the next step. When the model is forced to articulate a reasoning step explicitly, that articulation constrains and improves the quality of subsequent steps. The process of generating the reasoning chain is doing genuine cognitive work — the model is not just labeling a pre-formed answer, it is using the reasoning steps as scaffolding to build toward a better conclusion.
Chain-of-thought is most powerful on complex, multi-step reasoning tasks. For simple factual queries or straightforward writing tasks, it adds overhead without meaningful benefit. Use it selectively for the tasks where reasoning quality is the bottleneck.
Basic Chain-of-Thought: "Think Step by Step"
The simplest implementation of chain-of-thought is adding a phrase like "think step by step" or "reason through this carefully before answering" to your prompt. This phrase has been shown in research to activate more thorough reasoning in large models.
Example prompt without chain-of-thought: "Based on the following three SaaS pricing strategies, which is most likely to reduce churn? [Strategy details]"
Example prompt with basic chain-of-thought: "Based on the following three SaaS pricing strategies, which is most likely to reduce churn? Think step by step: first analyze each strategy's effect on customer switching costs, then consider how each affects perceived value, then assess each against the churn patterns described, and finally give your recommendation with reasoning. [Strategy details]"
The second prompt will almost always produce a more defensible, better-reasoned answer — not because the model suddenly became smarter, but because the reasoning process improved the quality of the conclusion.
You add 'think step by step' to a prompt asking AI to evaluate three vendor proposals. The output quality improves. What is the most accurate explanation for why this works?
Select one answer.
Structured Chain-of-Thought: Defining the Steps
A more powerful version of chain-of-thought explicitly defines the reasoning steps you want the model to follow. Instead of asking it to "think step by step" and letting it determine the steps, you specify the framework.
This is particularly useful when you have a domain-specific analytical framework that the model might not apply by default.
Example prompt:
"You are a senior strategy consultant. Analyze the following market entry opportunity using this exact sequence:
Step 1: Market sizing. Estimate the addressable market based on the figures provided. Step 2: Competitive analysis. Identify the key incumbent strengths and our differentiation potential. Step 3: Entry barrier assessment. List the three most significant barriers to entry and rate each as low/medium/high. Step 4: Risk-adjusted recommendation. Given Steps 1-3, state whether to enter, when, and through which channel. Include the most critical assumption your recommendation depends on.
[Market data and company context here]"
By defining the steps, you are not just getting the model to reason more carefully — you are ensuring it reasons through the specific framework that is appropriate for this decision.
When you have a recurring analytical task that requires consistent reasoning quality, convert your best chain-of-thought prompt into a template. The framework stays constant; the input data changes. Each use benefits from the reasoning structure you have already designed.
Structured reasoning for supplier risk decisions
Context
A procurement manager at a manufacturing company was using AI to help evaluate alternative suppliers when primary ones were disrupted. Direct-answer prompts produced quick recommendations that the team found difficult to defend in review meetings — the reasoning behind each recommendation was unclear and sometimes contradicted constraints that had been stated in the brief.
Action
She redesigned her supplier evaluation prompt as a structured chain-of-thought with four explicit steps: first assess whether the supplier meets the non-negotiable compliance requirements, then evaluate lead time against the current production schedule, then assess capacity and risk of single-source dependency, and finally state a recommendation with the most critical assumption named. The structured steps matched her team's actual decision framework.
Outcome
Recommendations produced by the chain-of-thought prompt were auditable at each step, which made stakeholder reviews faster and less contentious. The explicit assumption named in the final step also proved useful — on two occasions it flagged a data gap the team had not consciously registered, prompting them to gather additional information before committing.
Zero-Shot vs. Few-Shot Chain-of-Thought
Zero-shot chain-of-thought is the version above: you tell the model to reason step by step without showing it an example of what that looks like. This works well for most tasks with large capable models.
Few-shot chain-of-thought adds examples of the reasoning process you want. You show the model a worked example of a question, the step-by-step reasoning, and the final answer — and then ask it to apply the same process to your question. This is most useful when:
- The reasoning framework is unusual or domain-specific.
- The task requires a very particular reasoning style.
- The model is consistently making a specific type of reasoning error that an example can correct.
For most business prompt engineering, zero-shot chain-of-thought with specified steps is sufficient. Few-shot chain-of-thought is the tool you reach for when zero-shot is not producing the reasoning quality you need.
When to Use Chain-of-Thought
Chain-of-thought is most valuable for tasks involving: multi-factor comparisons (evaluating options against multiple criteria), causal reasoning (understanding why something happened or will happen), logical inference (drawing conclusions from premises), numerical or logical problem-solving, and any task where showing the reasoning makes the output auditable.
It is less useful for: creative writing tasks, formatting tasks, simple information retrieval, and tasks where speed matters more than analytical depth.
For a practical business example: use chain-of-thought when asking AI to prioritize a product backlog, evaluate a vendor proposal, analyze customer churn drivers, or recommend a marketing channel allocation. Do not use it when asking AI to draft an email, generate a tagline, or reformat a document.
A visible reasoning chain is not the same as a correct one. The model can produce four fluent, well-labeled steps that read as rigorous analysis while step three quietly rests on a fabricated figure or a misapplied assumption — and the confident, orderly presentation makes that error easier to miss, not harder. Chain-of-thought improves the quality of the reasoning process; it does not verify the reasoning. Read each step on its own merits rather than trusting the chain because it looks methodical, and apply the same fact-checking standard to a number that appears mid-chain as you would to one in a final answer.
Why does chain-of-thought prompting improve model output quality on complex reasoning tasks?
Select one answer.
Exercise
Your Task
Choose a real analytical or decision task from your current work — a vendor comparison, a prioritization call, a root cause question, or an options evaluation. Write two versions of the prompt: Version A is a direct question asking for the answer. Version B is a structured chain-of-thought prompt with three to four explicitly named reasoning steps before the final recommendation. Run both prompts and compare. Note whether the chain-of-thought version surfaced a consideration, constraint, or assumption that Version A missed.
Success looks like
- Version B defines three to four explicit, ordered reasoning steps — not just 'think step by step' but named steps specific to the task
- The chain-of-thought output is more auditable: you can follow the reasoning and verify each step independently
- The chain-of-thought version reveals at least one consideration the direct-answer version did not address
- You can identify which of your reasoning steps added the most analytical value — this is the step to include in any template you build from this exercise
Watch out for
- Using 'think step by step' without defining the actual steps — zero-shot chain-of-thought is less reliable than structured chain-of-thought when you have a specific analytical framework to apply
- Applying chain-of-thought to a simple task where the reasoning overhead adds length without improving quality
Hint
If the chain-of-thought output is better but the reasoning steps feel generic, go back and make each step domain-specific. Instead of 'Step 2: evaluate the options', write 'Step 2: evaluate each option against the budget constraint and implementation timeline we described.' Specificity in the steps is what turns chain-of-thought into a reliable analytical tool.
- Chain-of-thought prompting asks the model to reason step by step before producing a final answer, which significantly improves output quality on complex, multi-step reasoning tasks.
- The simplest implementation is adding 'think step by step' or 'reason through this carefully before answering' to your prompt — research shows this activates more thorough reasoning in large models.
- Structured chain-of-thought explicitly defines the reasoning steps, which is particularly powerful when you have a domain-specific analytical framework to apply consistently.
- Use chain-of-thought selectively — it adds value on analytical tasks like option comparisons, causal reasoning, and vendor evaluations, and adds overhead without benefit on simple tasks.
- Converting your best chain-of-thought prompts into templates ensures consistent reasoning quality across every instance of a recurring analytical task.