Zero-Shot and Few-Shot Prompting

Zero-shot prompting means asking the model to perform a task without giving it any examples of how to do it. You simply describe the task and let the model apply its pre-trained knowledge. This works well for tasks the model has seen many times during training, such as translation, summarization, or simple classification.

For example, asking the model to Classify this review as Positive or Negative is a zero-shot prompt. The model has absorbed thousands of sentiment classification examples during pre-training, so it can perform the task without being shown examples in the prompt itself.

Zero-shot prompts are concise and rely on the model built-in knowledge. They work best when the task is common and unambiguous. Notice that the prompt below defines the task clearly without showing any examples of correct outputs.

Zero-shot prompting can fail when the task format is unusual, when the model must follow an uncommon output schema, or when you need consistent formatting that the model might interpret differently each run. Asking the model to extract data in a very specific JSON structure with custom field names is risky as a zero-shot prompt because the model must guess your exact intent.

Zero-shot also struggles with domain-specific jargon or niche classification categories the model may not have encountered frequently during pre-training. In these cases, providing examples dramatically improves accuracy.

Few-shot prompting means providing the model with 2-8 examples of the task (input and desired output) before presenting your actual query. The examples teach the model the exact format and classification scheme you want without any fine-tuning or training required.

This is one of the most powerful and underused techniques in prompt engineering. The model uses the examples to infer the pattern and applies it to the new input. Research from the GPT-3 paper showed that performance on many tasks scaled significantly with the number of examples, up to about 8, after which gains plateau and you just waste tokens.

A few-shot prompt follows a consistent pattern: show N examples as input-output pairs, then present the new input and let the model complete the output. The formatting of the examples teaches the model the exact output format you expect.

The quality of your examples matters as much as their quantity. Choose examples that:

• Cover edge cases — include an example with ambiguous input so the model sees how you want it handled
• Span the output space — if you have three classes, include at least one example of each
• Are representative — use examples similar in style and complexity to real production inputs
• Are consistent in format — any format inconsistency in your examples will propagate into the model output

Poorly chosen examples can actually hurt performance by misleading the model about the task distribution.

Use zero-shot when: the task is common and well-defined, you want minimal token cost, or you are iterating quickly and examples are hard to collect yet. Zero-shot is also preferable when the model general knowledge should not be constrained by examples.

Use few-shot when: the output format is unusual or strict, you have domain-specific categories, zero-shot produces inconsistent formatting, or you need the model to follow a particular style or tone consistently. Few-shot is especially effective for extraction, classification, and generation with format constraints.

Classification is one of the highest-value use cases for few-shot prompting. When you have custom labels that do not map to common categories, providing examples trains the model on your taxonomy. The examples below define exactly what Bug, Feature Request, and Question mean in the context of support tickets.

Between zero-shot and few-shot sits one-shot prompting: providing exactly one example. This is often enough to fix formatting issues or clarify ambiguous tasks, and it costs far fewer tokens than a full few-shot prompt. One-shot is a good first escalation step when zero-shot produces unreliable results.

In practice, start with zero-shot, evaluate on a sample of real inputs, and add examples only where the model is failing. This keeps your prompts lean and token-efficient while targeting improvements where they matter most.

In production systems, hardcoding the same few examples for every query is suboptimal. Dynamic few-shot selection retrieves the most similar examples to the current input from a labeled example bank using embedding similarity, then injects only those examples into the prompt.

This way, the model always sees the most relevant demonstrations for the specific input, rather than generic examples that may not match the input style or domain. This technique combines the benefits of few-shot prompting with the scalability of a retrieval system, and is sometimes called example-based retrieval-augmented prompting.

Few-shot prompting is the most reliable way to enforce strict output formats before you have access to structured outputs or JSON mode. By showing the model the exact format you expect — including field names, delimiters, and ordering — you dramatically increase the chance of getting parseable output.

Always end your few-shot prompt with the beginning of the expected output pattern (e.g., the opening brace of a JSON object or the first field name). This primes the model to continue the pattern rather than potentially summarizing or commenting on the task first.

Test your understanding of AI Engineering concepts from this lesson.

In this lesson you learned: zero-shot prompting asks the model to perform tasks from description alone, few-shot prompting provides 2-8 input-output examples to teach format and taxonomy, and dynamic few-shot selection retrieves the most relevant examples per query using embedding similarity. Next up we explore chain-of-thought prompting for complex reasoning tasks.