Step by Step
T
Think aloud — externalizing reasoning
Chain-of-thought prompting forces the model to externalize its reasoning before giving a final answer, rather than jumping directly to a conclusion.
Example: a model working through each step of a math problem explicitly in its response, rather than simply stating a final numeric answer with no shown work.
R
Reason through — catching errors along the way
The intermediate reasoning steps help the model catch its own errors partway through, similar to how showing your work in math class helps a student catch mistakes before finalizing an answer.
Example: a model, while working through intermediate steps explicitly, noticing and correcting an arithmetic error partway through, rather than carrying that error silently into a final wrong answer.
Z
Zero-shot CoT — just add the magic phrase
Zero-shot Chain-of-Thought simply adds the phrase "Let's think step by step" to a prompt, with no examples needed, and still produces a significant accuracy improvement.
Example: adding just the phrase "Let's think step by step" to an otherwise unchanged math prompt, with no examples provided, and observing a meaningful accuracy improvement.
F
Few-shot CoT — providing example reasoning chains
Few-shot Chain-of-Thought goes further, providing example question-reasoning-answer chains before the actual question, demonstrating the specific style of reasoning desired.
Example: showing two example math problems, each with their full worked-through reasoning chain and final answer, before asking the model to solve a new, similar problem the same way.
Applied Walkthrough
1
A student wants an LLM to solve a multi-step algebra word problem correctly, and a direct request for just the final answer produces an incorrect result.
2
Adding "Let's think step by step" (zero-shot CoT) to the same prompt, with no other changes, causes the model to explicitly work through each step, catching an error it would have otherwise carried silently into a wrong final answer.
3
For an even more specific style of reasoning, the student could instead use few-shot CoT, showing 2-3 example problems with their full worked reasoning chains before asking the new question.
4
This directly illustrates why Chain-of-Thought is considered one of the most powerful prompt techniques — forcing externalized, step-by-step reasoning measurably improves accuracy on tasks requiring multi-step logic.
Exam Application
Exams test whether you understand WHY Chain-of-Thought improves accuracy (externalizing reasoning helps the model catch its own errors along the way, similar to showing your work) and whether you can distinguish zero-shot CoT (just adding the magic phrase) from few-shot CoT (providing example reasoning chains).
⚠ Common Trap
The most common trap is assuming Chain-of-Thought is just a stylistic preference for longer responses. Its actual mechanism is functional, not stylistic — forcing the model to externalize intermediate reasoning steps genuinely helps it catch and correct errors it would otherwise make silently while jumping directly to a final answer.
✓ Quick Self-Check
1. What does Chain-of-Thought prompting force the model to do?
Externalize its reasoning before giving a final answer, rather than jumping directly to a conclusion.
Tap to reveal / hide
2. Why does externalizing reasoning improve accuracy?
The intermediate steps help the model catch its own errors along the way, similar to showing your work in math class.
Tap to reveal / hide
3. What does zero-shot Chain-of-Thought involve?
Simply adding the phrase "Let's think step by step" to a prompt, with no examples needed.
Tap to reveal / hide
4. What does few-shot Chain-of-Thought involve?
Providing example question-reasoning-answer chains before the actual question, demonstrating the desired reasoning style.
Tap to reveal / hide
5. Is Chain-of-Thought's benefit primarily stylistic, or functional?
Functional — it genuinely helps the model catch and correct errors, not just produce longer or more stylistically detailed responses.
Tap to reveal / hide