What Is Chain-of-Thought Prompting?
Chain-of-thought (CoT) prompting is a technique where you structure your prompt to make an AI model reason through intermediate steps before arriving at a final answer, rather than jumping straight to a conclusion. Think of it as forcing the AI to show its work.
Introduced in a 2022 Google Brain paper by Wei et al., CoT consistently improves AI accuracy on multi-step reasoning tasks — analysis, planning, logic, and decision-making — by 40–60% compared to direct-answer prompting. It works because language models produce better outputs when each intermediate inference is explicitly constructed rather than implicitly assumed. The model that "thinks out loud" makes fewer leaps and catches more of its own errors.
Here is the key insight most practitioners miss: chain-of-thought is not a single technique. It is a family of at least four methods, each suited to different task types. Most people only use the weakest one.
The 4 Levels of Chain-of-Thought — And Where Most People Stop
The four main CoT approaches form a rough progression from simple to powerful. Understanding which level you are using — and when to upgrade — is what separates practitioners who get consistent results from those who get occasional luck.
Level 1 — Zero-Shot CoT: You append a phrase like "Let's think step by step" or "Reason through this carefully" to your prompt. This was shown to work by Kojima et al. (2022) and it does improve outputs for simple tasks. It is also the only version most people ever use.
Level 2 — Few-Shot CoT: You include 2–5 worked examples in your prompt that demonstrate the reasoning pattern you want. The model sees the pattern and replicates it for your actual question. According to the DAIR.AI Prompt Engineering Guide (2025), few-shot CoT outperforms zero-shot CoT on structured reasoning tasks by an average of 18 percentage points.
Level 3 — Self-Consistency: You run the same prompt multiple times (5–10 iterations) and select the answer that appears most frequently across all chains. Google's original CoT research showed this technique reduces error rates by 30–40% on complex logical problems. Use it when the stakes are high and a single run is not reliable enough.
Level 4 — Structured CoT (CHAIN Framework): You define an explicit reasoning schema in your prompt — Context, Hypothesis, Analysis, Inference, Narration — and ask the model to work through each stage. This is the version used in enterprise workflows and advanced research contexts.
How to Write a Few-Shot CoT Prompt That Actually Works
Few-shot CoT is the sweet spot for most practitioners: significantly more reliable than zero-shot, but far less time-consuming than self-consistency. Here is the structure that produces consistently strong results.
A well-formed few-shot CoT prompt contains three parts: a clear task framing, 2–3 worked examples where you model the reasoning yourself, and then the actual question. The examples do not need to be perfect — they need to show the type of reasoning steps the model should follow.
Here is a copy-paste template for strategic decision analysis:
Try This Prompt:
---
You are a strategic analyst. When I give you a decision to evaluate, reason through it using these exact steps: (1) Identify the core assumption being made. (2) List what evidence supports that assumption. (3) List what evidence contradicts it. (4) Assess the risk of being wrong. (5) Give a clear recommendation with your reasoning in one sentence.
Example: [Question] Should we launch a new product line in Q3?
[Step 1] Core assumption: Q3 consumer demand will remain strong.
[Step 2] Supporting evidence: Q2 sales up 18%, competitor launches planned for Q4, team capacity available.
[Step 3] Contradicting evidence: Supply chain delays in May, 2 key team members on leave in August.
[Step 4] Risk of being wrong: High — a delayed launch in the middle of Q3 is worse than waiting for Q4.
[Step 5] Recommendation: Delay launch to early Q4 to reduce execution risk and align with competitor timing.
Now apply this framework to: [Your question here]
---
The structure forces the model to confront its own reasoning. You will catch significantly more errors and surface assumptions you did not know you were making.
Real Workplace Applications: What CoT Actually Unlocks
Chain-of-thought prompting produces the most dramatic improvements in tasks that require multi-step reasoning. Here are four categories where practitioners consistently see the biggest lift.
Content strategy decisions: Instead of asking "Should we focus on LinkedIn or Instagram this quarter?", frame it as a CoT analysis: list your audience characteristics, then platform strengths, then alignment between the two, then your resource constraints, then a recommendation. The structured output is directly usable in a strategy document.
Data interpretation: When reviewing analytics, use CoT to ask the model to identify (1) what the numbers say, (2) what they don't say, (3) the most likely explanation for the pattern, and (4) the most important action to take. This four-step structure prevents the common AI failure mode of jumping to the most obvious interpretation.
Client brief analysis: Use CoT to extract (1) stated objectives, (2) unstated objectives, (3) constraints, (4) success metrics, and (5) potential misalignments between objectives and constraints. In practice, step 5 consistently surfaces issues human readers miss on a first pass.
Writing and editing: Use CoT to evaluate a piece of writing by first assessing structure, then argument quality, then evidence strength, then tone alignment, then a summary verdict. This produces more actionable feedback than asking for a general critique.
Where Chain-of-Thought Breaks Down
CoT is not a universal fix. Knowing where it fails is as important as knowing where it works.
Simple or factual tasks: For questions with a direct correct answer — "What is the capital of Japan?" or "Translate this sentence" — CoT adds no value and slows down your workflow. Use direct prompting for simple retrieval tasks.
Creative tasks requiring spontaneity: CoT can suppress the generative quality of creative output by over-structuring the reasoning. For brainstorming and ideation, let the model roam free first, then apply structured evaluation in a second pass.
Model size matters: According to GPT-4.1 and Claude Sonnet 4.6 documentation, larger models benefit most from CoT because they have sufficient reasoning capacity to fill each step meaningfully. With smaller or older models, the intermediate steps can be nonsensical, producing false confidence rather than better results.
The hallucination trap: CoT does not eliminate hallucination — it can make it more elaborate. A model that hallucinates a step early in the chain will build subsequent steps on that false foundation. Always verify factual claims in CoT outputs independently.
Model-Specific Tips: Getting the Most From Claude, GPT-4o, and Gemini
Different models respond differently to chain-of-thought prompting. Knowing these differences saves significant trial-and-error time.
Claude Sonnet 4.6 and Opus 4.6 are particularly strong at structured CoT. These models produce methodical, well-organised reasoning chains with minimal prompting. For Claude, you can reduce your few-shot example count to 1–2 because the model infers the reasoning pattern quickly. Claude also self-corrects within the chain — it will often flag when a step contradicts an earlier assumption, which is a significant reliability advantage for analysis tasks.
GPT-4o benefits from more explicit step labelling. Rather than "think step by step," specify the steps directly: "First identify the problem, then list three possible causes, then evaluate each cause based on available evidence, then recommend an action." GPT-4o executes explicitly named steps more reliably than it follows open-ended reasoning instructions.
Gemini 1.5 Pro and 2.5 Pro perform best on CoT tasks that involve long documents or multimodal inputs. If you are using CoT to reason over a lengthy report or a set of images, Gemini's 1M+ token context window gives it a structural advantage. Use few-shot CoT examples drawn from the same document type you are analysing for maximum accuracy.
A practical rule: for any task where you are getting inconsistent outputs across runs, add one level of CoT structure before switching models. In most cases, the inconsistency is in the prompt, not the model.
Try It Now: A 10-Minute CoT Exercise
Here is a practical exercise to run in the next 10 minutes. Take a real decision you are currently facing at work — a prioritisation question, a content direction, a process choice — and run it through this Zero-Shot-to-Few-Shot comparison:
Pass 1 (direct prompt): Ask your AI model the question directly and note the answer.
Pass 2 (zero-shot CoT): Add "Think through this step by step before giving your answer" and note how the output changes.
Pass 3 (few-shot CoT): Add a worked example of a similar decision with visible reasoning steps, then ask your actual question.
In most cases, the quality gap between Pass 1 and Pass 3 will be immediately obvious. The structured version will surface assumptions and trade-offs that the direct answer completely ignored. Once you see this gap with your own question, CoT becomes a default rather than an optional technique.
The practitioners who get the most out of AI are not the ones using the most tools — they are the ones who have mastered a small number of foundational techniques and apply them consistently. Chain-of-thought is the most transferable of all of them.
懂AI的冷，更懂你的難 — UD 同行28年，讓科技成為有溫度的陪伴。That consistent, reliable output you have been looking for is not about finding a better model. It is about using the one you have correctly.
Ready to Put Your AI Skills to the Test?
Now that you know how to use chain-of-thought prompting properly, how does your overall AI knowledge stack up? The UD AI IQ Test benchmarks your prompting skills, tool knowledge, and workflow thinking against other practitioners — and we'll walk you through every step of your results so you know exactly where to level up next.
