Prompt Engineering – Zero-shot, One-shot, Few-shot & Chain-of-Thought Prompting

TL;DR

Zero/one/few-shot describe how many worked examples you put in the prompt before asking the real question; chain-of-thought asks the model to show intermediate reasoning before its final answer. All four are ways of shaping an LLM's output without touching its weights.

Why it matters

Prompt engineering is the cheapest lever you have over an LLM's behavior — no fine-tuning, no extra infrastructure, just structuring the input. Picking the right style (and knowing when fancier isn't better) is directly relevant to anyone building on top of an LLM, and it's a favorite exam topic because it's easy to test with a concrete example.

The core idea

All four techniques are about in-context learning — giving the model information inside the prompt itself, at inference time, rather than training it further:

• Zero-shot: just the instruction, no examples. The model relies entirely on what it already learned during pretraining.
• One-shot: the instruction plus exactly one worked example, mainly to pin down the expected output format/label vocabulary.
• Few-shot: the instruction plus several worked examples (typically covering different classes or edge cases), so the model can infer the pattern/boundary from multiple demonstrations.
• Chain-of-thought (CoT): the prompt explicitly asks the model to reason step by step before giving a final answer — either by instruction ("think step by step") or by showing an example whose demonstrated answer includes visible reasoning steps.

How it works

Below is the same task — classifying a review's sentiment as Positive, Negative, or Mixed — written in all four styles, so the contrast is visible directly.

Zero-shot

Classify the sentiment of this review as Positive, Negative, or Mixed.
 
Review: "The battery life is amazing but the screen cracked within a week."
Sentiment:

One-shot

Classify the sentiment of the review as Positive, Negative, or Mixed.
 
Review: "This laptop runs so fast and the design is sleek."
Sentiment: Positive
 
Review: "The battery life is amazing but the screen cracked within a week."
Sentiment:

Few-shot

Classify the sentiment of the review as Positive, Negative, or Mixed.
 
Review: "This laptop runs so fast and the design is sleek."
Sentiment: Positive
 
Review: "Terrible purchase, it stopped working after two days."
Sentiment: Negative
 
Review: "Great camera, but the battery drains too quickly."
Sentiment: Mixed
 
Review: "The battery life is amazing but the screen cracked within a week."
Sentiment:

Chain-of-thought

Classify the sentiment of the review as Positive, Negative, or Mixed. Think step by step about
each aspect mentioned before giving the final answer.
 
Review: "The battery life is amazing but the screen cracked within a week."
Reasoning: The review mentions two aspects — battery life (positive) and screen durability
(negative). Since a clear positive and a clear negative are both present, the overall sentiment
is mixed rather than purely one or the other.
Sentiment: Mixed

Notice the few-shot example is also the one most likely to correctly produce "Mixed" rather than forcing a binary Positive/Negative choice — simply because it's the only style that demonstrates the Mixed class exists before the real question arrives.

When each is most suitable

• Zero-shot: simple, common tasks where the model's pretraining already covers the pattern well (e.g. basic sentiment, well-known formats); cheapest in tokens and latency.
• One-shot: when the task itself is easy but the exact output format or label vocabulary needs to be pinned down (e.g. you want "Positive"/"Negative" and not "This review is positive").
• Few-shot: when the task has subtle categories, ambiguous edge cases, or an unusual output format — showing multiple examples (ideally covering the edge cases, like the "Mixed" class above) teaches the boundary more reliably than an instruction alone, and matters more for smaller or less instruction-tuned models.
• Chain-of-thought: whenever the task needs multi-step reasoning — arithmetic, logical deduction, multi-aspect judgment calls like the mixed-sentiment example. Forces the model to externalize intermediate steps, which measurably reduces careless errors on reasoning-heavy tasks, at the cost of more output tokens and latency.

Related techniques

• Instruction prompting: giving an explicit natural-language instruction about the task and desired output, without necessarily including examples — e.g. "Respond with exactly one word: Positive, Negative, or Mixed." This can be combined with any of the four styles above.
• Role prompting: assigning the model a persona to bias its style, tone, or assumed expertise — e.g. "You are an expert product reviewer analyzing customer feedback." Also composable with zero/few-shot and CoT; it doesn't replace them.

Worked example

Applying the few-shot prompt above: the model has now seen one Positive, one Negative, and one Mixed example before the real review. Given that the real review explicitly states one positive aspect (battery) and one negative aspect (screen), the few-shot context makes "Mixed" the pattern- consistent answer — compare this to the zero-shot version, which (with no demonstrated Mixed class) is more likely to force a binary Positive-or-Negative call.

Exam traps

• Few-shot isn't just "more examples = better" — beyond a handful, extra examples mostly add cost without improving accuracy, and can even hurt if the examples are inconsistent or unrepresentative.
• CoT's value is specifically for reasoning tasks — it doesn't meaningfully improve trivial lookup/classification tasks, so proposing CoT as the universal fix is a common wrong "because it's fancier" answer.
• Don't conflate few-shot prompting with fine-tuning — no weights change; all of this happens purely via the prompt's context at inference time.
• Instruction and role prompting are complementary modifiers, not alternatives to zero/one/ few-shot/CoT — a prompt can combine a role, an instruction, and a few examples all at once.

Remember this

Zero/one/few-shot vary how many demonstrations you give; CoT varies whether you ask for visible reasoning — pick based on whether the task's difficulty is in knowing the pattern (more examples help) or in reasoning through the answer (CoT helps).