How Many Examples in Your AI Prompt Are Best? The 2-5 Rule Backed by Research

Including examples in your AI prompts — known as few-shot prompting — is one of the most effective techniques for improving output quality. Few-shot prompting provides demonstrations in the prompt to steer the model toward better performance and condition its response on the supplied examples . But how many examples should you use? The research is clear: the sweet spot is typically 2–5 examples, with evidence and guidance suggesting that adding too many examples can lead to diminishing or even negative returns .

Why examples matter: The dramatic accuracy gains

Examples can transform AI output quality. Studies show that models like Claude 3 Haiku achieved just 11% correctness on certain tasks with zero-shot prompting, but jumped to 75% with merely three examples — performance matching much larger models' zero-shot capabilities .

For medication attribute extraction tasks, adding just 2–3 examples boosted accuracy from 0.88 to 0.96 in benchmarks . This kind of improvement demonstrates why few-shot prompting has become a standard technique in professional AI work.

The research-backed sweet spot: 2–5 examples

Multiple sources converge on 2–5 examples as the optimal range. The Prompt Engineering Guide describes few-shot prompting as providing about 2–5 examples, and prompting guidance commonly identifies this range as typical . One comprehensive guide notes that 3–5 examples is the optimal range for most tasks, with example selection mattering more than example quantity .

A 2025 study from the Wharton Business School, "The Decreasing Marginal Returns of Examples," found diminishing returns beyond the 2–5 range . Multiple research papers point to major gains after 2 examples and then a plateau — after 2 examples you might just be burning tokens .

When more examples hurt: The over-prompting problem

An arXiv study identified an "over-prompting" problem where too many examples can reduce model performance, challenging the assumption that more examples always help . The authors developed a prompting framework using random sampling, semantic embedding, and TF-IDF vectors to select examples, evaluating these methods across seven different LLMs on software requirement classification tasks. Their findings demonstrated that TF-IDF selection often outperforms other methods, leading to better performance with fewer examples by effectively mitigating the over-prompting effect .

For reasoning-heavy tasks, adding too many examples can actually reduce performance. The strategic recommendation is to default to zero-shot instructions and reserve few-shot examples purely for output format control when the schema isn't obvious. If you must use examples, limit them to 1–2 high-quality instances to avoid overcomplicating the model's internal search process .

Quality over quantity: Choosing the right examples

Example quality matters more than quantity. Diverse, high-quality, and well-stratified examples can outperform simply adding more examples, while inconsistent examples can teach the model inconsistent patterns .

Follow these proven principles for optimal results: use 2–5 diverse examples that cover different aspects of your task, pay careful attention to example quality, and standardize the structure of examples to clarify the relationship between input and output .

Keep examples short and add just enough context to make the label obvious. A reliable baseline is a structured format — Q: Input / A: Output — and holding that shape consistently .

Practical recommendations by task type

For format-sensitive tasks like JSON generation, label extraction, or tone control, use 2–5 clean, diverse examples, since examples can anchor format, tone, and labels .

For standard text generation or classification, start with 2–3 high-quality examples, since few-shot demonstrations can steer the model toward better performance .

For reasoning-heavy tasks, start with zero-shot; add 1–2 examples only if you need stronger format control, and test whether examples help or hurt .

For complex extraction tasks, use a small set of representative examples that mirror the task closely .

When in doubt, start at 3 examples, test, and adjust, since 2–5 examples is a commonly recommended range and too many examples can cause over-prompting .

The bottom line

Use examples in your AI prompts, keep them to 2–5, and prioritize quality and diversity over quantity . Test whether your specific model benefits from few-shot vs. zero-shot, especially on reasoning-heavy tasks where extra examples may not always help . The most effective prompt engineers don't just add examples — they choose them carefully, keep them consistent in structure, and know when to leave them out.