Compare

Zero-shot vs Few-shot vs Chain-of-thought

Three prompting techniques stacked in order of hand-holding

The verdict

Zero-shot asks the model to do a task cold; few-shot shows examples; chain-of-thought asks the model to reason step by step. They compose: few-shot chain-of-thought is often stronger than either alone.

Zero-shot

Glossary →

Give the model a task description and no examples. Works surprisingly well on modern LLMs for common tasks.

Best for: Common tasks the model has clearly seen before.

Few-shot

Glossary →

Include 1-10 worked examples in the prompt to demonstrate the format and quality expected. Also called in-context learning.

Best for: Custom formats and edge-case behaviors.

Chain-of-thought

Glossary →

Prompt the model to reason step by step before answering (e.g., 'Let's think step by step'). Boosts math and multi-hop reasoning; also called CoT.

Best for: Multi-step reasoning, math, planning.

At a glance

Zero-shot vs Few-shot vs Chain-of-thought: dimension-by-dimension comparison
Dimension	Zero-shot	Few-shot	Chain-of-thought
Examples in prompt	A worked reasoning trace (optional)	1-10 input/output pairs	None
Cost	More output tokens	More input tokens	Cheapest
Best for	Math, multi-hop	Custom formats, edge cases	Common tasks
Composes with each other	Yes	Yes	Baseline
Modern default	Built in (reasoning models)	For custom formats	For most tasks
Failure mode	Wrong reasoning still confidently stated	Overfitting to demo examples	Model guesses format

Key differences

1Zero-shot uses no examples; few-shot uses a handful; CoT is orthogonal (about reasoning)
2CoT is a prompt trick that pairs with either zero-shot or few-shot
3Few-shot examples cost context tokens; CoT costs output tokens
4On reasoning-heavy tasks, few-shot CoT often beats plain few-shot by 10-30 points
5Reasoning-tuned models (o1, R1) reduce the need for explicit CoT prompting

In the interview

What they're really testing

Whether you see CoT as orthogonal to few-shot, and whether you know when each is worth its token cost.

Say this

Zero-shot is task-only; few-shot adds a handful of demonstrations; chain-of-thought asks the model to think step by step. Zero-shot is the cheap default. I reach for few-shot when the format is custom, and for CoT on math or multi-hop reasoning; they compose, so few-shot chain-of-thought is often the strongest for tricky reasoning. On modern reasoning-tuned models the CoT prompt is baked in and less needed.

Traps to sidestep

Adding many-shot for a task the model already handles zero-shot
Slapping 'think step by step' on classification tasks (no benefit, extra cost)
Not distinguishing prompting from fine-tuning

How to choose

If the task is common and format is standardZero-shot

If the format is custom or edge cases matterFew-shot

If the task needs multi-step reasoningChain-of-thought

If output tokens are expensive and reasoning is trivialZero-shot

Standard task → zero-shot. Custom format → few-shot. Reasoning-heavy → CoT (compose with few-shot).

Common misconceptions

Myth: Chain-of-thought helps every task.

Reality: On classification or extraction it usually just adds cost. CoT wins on math, planning, and multi-hop QA.

Myth: More examples always help few-shot.

Reality: There's a sweet spot around 3-8. Too many either bloat context or bias the model toward superficial patterns in the demos.

Memory aid

Zero-shot: 'do this.' Few-shot: 'here are three like this.' CoT: 'think out loud while you do it.'

Can you combine them?

They compose freely. Few-shot chain-of-thought (show worked reasoning traces as examples) is often the strongest combination for math and planning.

Related comparisons

Chain-of-thoughtvsTree-of-thought

Chain-of-thought vs Tree-of-thought

Linear step by step reasoning vs branching search

In-context learningvsFine-tuning

In-context learning vs Fine-tuning

Adapt in the prompt, or adapt in the weights