Chain of Thought Prompting

Introduction

Chain-of-Thought (CoT) asks the model to show its working before the final answer. For math, multi-hop reasoning and complex tool selection, CoT can double accuracy. It's the building block of every reasoning agent.

Beginner analogy: a teacher saying 'show your steps' — students get partial credit and reach correct answers more often.

Understanding the topic

Core concepts:

• Prompt with: 'Let's think step by step.' or example reasoning.
• Final answer must be clearly delimited (`Answer:` line).
• Costs more tokens — only use when accuracy matters.
• Self-consistency: run CoT 5x, take majority vote.
• Reasoning models (o1, o3) do CoT internally.

Syntax reference

Visual workflow / architecture:

Q: A train leaves at 9am at 60mph...
Let's think step by step:
1. Distance per hour = 60 miles
2. After 3 hours = 180 miles
3. ...
Answer: 180 miles

Real-world use

Math word problems (GSM8K), legal reasoning, multi-step tool planning, agent ReAct loops.

Best practices

• Use CoT when the task has clear intermediate steps.
• Use self-consistency for high-stakes answers.

Common mistakes

• Asking for CoT then ignoring the steps — wasted tokens.
• Forgetting a final answer delimiter — hard to parse.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. What is CoT?
• Q2. When does CoT hurt performance?
• Q3. What is self-consistency?
• Q4. Scenario: how do you extract just the final answer from a CoT?