Zero-Shot Chain-of-Thought
The simplest way to trigger reasoning — just add one phrase to your prompt.
Zero-Shot CoT is the simplest form of chain-of-thought prompting. You don't need to provide examples — you just append a reasoning trigger phrase to your prompt. The landmark 2022 paper by Kojima et al. showed that simply adding "Let's think step by step" to the end of a prompt improved accuracy on math benchmarks from 17.7% to 78.7%.
This works because the trigger phrase primes the model to generate reasoning tokens before committing to an answer. It shifts the model from "pattern match to the most likely answer" mode into "work through the logic" mode.
Not all trigger phrases work equally well. Research and practice have shown that some phrases are more effective than others at eliciting careful reasoning.
- "Let's think step by step" — the original and still very effective
- "Think through this carefully before answering" — good for analysis tasks
- "Break this down into steps" — works well for procedural tasks
- "Before answering, consider all the factors" — great for decision-making
- "Show your work" — effective for math and calculation tasks
- "Let's approach this systematically" — good for complex multi-factor problems
Zero-Shot CoT becomes even more powerful when combined with a role. Telling the model to think like a specific expert changes the type of reasoning it applies.
Expert Reasoning
Combines role prompting with CoT for domain-specific reasoning.
You are a [financial analyst / senior engineer / medical researcher]. Think through this problem step by step, applying your expertise at each stage. Problem: [YOUR PROBLEM] Consider: 1. What are the key variables? 2. What assumptions are we making? 3. What does the analysis show? 4. What's the recommendation and why?
Zero-Shot CoT is your default — it's fast, easy, and works well with modern models. Switch to Few-Shot CoT when you need a specific reasoning format, when the problem type is unusual and the model needs a demonstration, or when Zero-Shot consistently produces disorganized reasoning.
Prompt Templates
Decision Analyzer
Zero-Shot CoT tailored for decision-making scenarios.
[DESCRIBE YOUR DECISION OR PROBLEM] Before answering, think through this carefully: - What are the key factors to consider? - What are the trade-offs? - What could go wrong with each option? - What does the analysis suggest? Then give your recommendation with a confidence level.
Calculation Verifier
Self-verifying CoT that catches calculation errors.
I need you to solve this and verify your answer: [PROBLEM] Step 1: Solve it showing all work. Step 2: Check your answer by working backward from the result. Step 3: If both methods agree, state the final answer. If they disagree, find the error.
Assumption Checker
CoT focused on identifying and evaluating hidden assumptions.
[CLAIM OR ARGUMENT] Let's analyze this step by step: 1. Identify every assumption being made (stated or implied) 2. Evaluate each assumption — is it valid, questionable, or false? 3. How does the conclusion change if each questionable assumption is wrong? 4. Provide a revised conclusion accounting for uncertain assumptions.
Test Your Knowledge
Knowledge Check
1 / 2
What improvement did "Let's think step by step" achieve on math benchmarks in the original research?
Key Takeaways
- ✓"Let's think step by step" is one of the highest-leverage phrases in prompt engineering
- ✓Zero-Shot CoT requires no examples — just a trigger phrase appended to your prompt
- ✓It works by priming the model to generate reasoning tokens before committing to an answer
- ✓Combine it with role prompts for domain-specific expert reasoning
- ✓Use Zero-Shot as your default; switch to Few-Shot for unusual problem types or specific formats
Continue Learning
What Is Chain-of-Thought Prompting?
Understand the technique that dramatically improves AI reasoning on complex problems.
When CoT Hurts in Modern Models
Chain-of-thought is not always beneficial. Learn when to skip it.
GPT vs Claude vs Gemini
Understand the practical differences between the major AI models.