Few-Shot Prompting

What is Few-Shot Prompting?

“Zero-shot” is an instruction with no examples; “one-shot” is one example; “few-shot” is usually two to eight. The model is not being trained — its weights do not change. It is reading the examples as context and continuing the pattern, which researchers call in-context learning.

The surprising finding behind good practice: examples teach format and label space more than they teach correctness. Research found that even randomising the labels in classification examples often hurt accuracy only a little. So the job of your examples is to show the exact shape of a good answer and the range of inputs to expect — and to be correct anyway, because you don’t want to rely on that.

When to use it — and when not to

Copyable template

[ONE-SENTENCE TASK DESCRIPTION]. Follow the format of the examples exactly. Output only the final line, nothing else.

<example>
Input: [EXAMPLE INPUT 1 — a typical case]
Output: [EXAMPLE OUTPUT 1]
</example>
<example>
Input: [EXAMPLE INPUT 2 — a clearly different case]
Output: [EXAMPLE OUTPUT 2]
</example>
<example>
Input: [EXAMPLE INPUT 3 — an awkward edge case]
Output: [EXAMPLE OUTPUT 3]
</example>

Input: [NEW INPUT]
Output:

Wrapping examples in tags keeps them clearly separate from your real input. Cover the range of cases, balance the labels, and vary the order so the last example is not always the same label.

Replace everything in [SQUARE BRACKETS] before you run it.

Worked example: Support ticket triage

A small SaaS team wants every incoming ticket tagged consistently so it can be routed automatically.

Triage each support ticket as: category | priority | one-line summary. Categories: billing, bug, account, feature-request. Priority: P1 (service down or money lost), P2 (blocked but has a workaround), P3 (everything else). Output only the line.

<example>
Input: I was charged twice for October, please refund one.
Output: billing | P1 | Duplicate October charge, refund requested
</example>
<example>
Input: Export to CSV hangs at 90% but PDF export works fine.
Output: bug | P2 | CSV export hangs at 90%; PDF works as workaround
</example>
<example>
Input: Would love a dark mode at some point!
Output: feature-request | P3 | Requests dark mode
</example>

Input: Since resetting my password I can't log in to the desktop app. The web version works, but I have a client demo in an hour.
Output:

account | P2 | Desktop app login fails after password reset; web app works; demo in an hour

Why it works: The instruction defines the rules; the three examples show the exact separator, the length of a summary and how the priority rules apply to real wording. One of each category type would make it more robust still.

Common mistakes

1. All examples look alike. If every example is short and polite, a long angry ticket gets handled badly. Choose examples that span the real variety of inputs.
2. Unbalanced or badly ordered labels. Models drift towards the label that appears most often in the examples, and towards the one that appears last. Balance them and vary the order.
3. Examples that contradict the instruction. When they disagree, the examples usually win. Re-read them against your rules.
4. Content leaking into outputs. The model may reuse names, numbers or phrases from your examples. Use clearly fictional details, or tell it the examples are illustrative only.
5. Too many examples. Past a handful, gains shrink while cost and context use grow. Add examples only for failure cases you have actually seen.

Origin & sources

The term became standard with OpenAI’s GPT-3 paper (Brown et al., 2020), which showed that a large enough language model could perform new tasks from a few examples in its prompt, without any fine-tuning. Zhao et al. (2021) documented why results swing so much with example choice — majority-label and recency biases — and Min et al. (2022) showed that format and the distribution of inputs matter more than whether the example labels are correct.

• Brown et al. (2020), “Language Models are Few-Shot Learners”, NeurIPS 2020 — arXiv:2005.14165
• Zhao et al. (2021), “Calibrate Before Use: Improving Few-Shot Performance of Language Models”, ICML 2021 — arXiv:2102.09690
• Min et al. (2022), “Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?”, EMNLP 2022 — arXiv:2202.12837

Frequently asked questions

Related techniques

← All prompting techniques