Few-Shot Prompting: A Simple Fix for Better AI Output
By Alexandre Saint-Jean
Audio version
Audio version produced by text-to-speech from the article. Our AI charter
Few-shot prompting means showing an AI 3 to 5 examples of what you want, rather than explaining it in words. It is the highest-return technique for a non-technical business owner: it needs no coding skill, just the discipline of pasting good examples before asking for a new output.
What is few-shot prompting?
Few-shot prompting is a method that gives a language model several concrete examples of the expected result, alongside the instruction. The model observes the common pattern across these examples (tone, structure, length, vocabulary) and reproduces it for the new request.
Anthropic calls this technique "multishot prompting" in its official documentation and presents it as one of the most effective levers for shaping how Claude behaves. The specific recommendation: use 3 to 5 relevant, varied examples that capture the important cases in the task. Source: Anthropic documentation, "Use examples effectively" section.
The Prompt Engineering Guide presents the same idea from the classic teaching angle: giving a few examples before the actual task noticeably improves output quality compared with a plain instruction with no example (zero-shot). It is one of the founding techniques of prompt engineering, widely documented.
How does few-shot differ from zero-shot prompting?
In zero-shot, the AI receives an instruction with no example: "Write a customer follow-up email." In few-shot, that same instruction comes with 2 to 3 follow-up emails already written, ones the user was happy with. The model no longer has to guess the expected style, it observes it directly in the examples provided.
This difference matters especially for anything tied to a company's tone: an email signed by a family-run business does not read the same as one signed by a consulting firm. Describing that tone with adjectives ("warm but professional") works less well than showing three emails that already embody it.
How do you apply few-shot prompting in a real business?
The move takes a minute and needs no special tool: paste 2 to 3 examples of past output you were happy with directly into the prompt, before making the request. Three use cases come up most often in business.
A standard email. Before asking AI to draft a follow-up email, paste 2 to 3 follow-up emails you have already sent that worked well. The model picks up the structure, the length, the sign-off and the degree of firmness in tone, instead of a generic style.
A standard customer reply. For support requests or reviews, paste a few replies already validated internally, including on difficult cases (a complaint, a delay, a billing error). AI learns the expected register and avoids answers that sound either too stiff or too casual.
Meeting notes. Paste one or two past sets of notes you consider well structured (same headings, same level of detail, same way of recording decisions and action items). The new output follows that template without needing to be described in detail every time.
In all three cases, the logic is the same: show, rather than describe. This is one of the practical building blocks of the AI literacy expected of a business leader, covered in more depth in the EU AI Act literacy obligation, explained.
What is the main pitfall of few-shot prompting?
The most common mistake is mixing inconsistent examples. If one example is very formal and another casual, if one is three lines and another three paragraphs, the model cannot find a clear pattern. It averages the contradictory signals instead of faithfully reproducing the best style, and the final result reads like a blurry compromise, not a good answer.
The practical rule drawn from Anthropic's documentation comes down to three criteria for choosing examples: they should be relevant to the exact task at hand, cover a real variety of cases (not three variants of the same one), and be consistent with each other in tone and structure. Three good, uniform examples beat five mismatched ones.
A second pitfall, flagged by the Prompt Engineering Guide, concerns complex reasoning tasks: on that kind of task, showing examples is not always enough. It is usually better to ask AI to work through its reasoning step by step, a separate technique called chain-of-thought, rather than relying solely on examples of the final result.
Is there more to learn beyond few-shot prompting for everyday AI use?
Few-shot prompting is a starting point, not an end point. It is part of a broader foundation every business leader is now expected to build, spanning the regulatory AI literacy obligation set by Article 4 of the AI Act and the practical need to get real value from the tools already deployed in the business. Getting comfortable with few-shot prompting alone already changes the perceived quality of an AI assistant's daily answers, with no technical skill required.
To place this technique within a fuller skills-building path, see also AI literacy for business leaders: where to start, which sets out the minimum foundation expected of anyone in a decision-making role.
Frequently asked questions
- What is the difference between few-shot and zero-shot prompting?
- In zero-shot, you ask AI to do a task with no examples, just instructions. In few-shot (or multishot, in Anthropic's terminology), you add 3 to 5 concrete examples of the expected result. Few-shot works better whenever the expected format, tone or structure is precise and hard to describe with instructions alone.
- How many examples do you need for few-shot prompting?
- Anthropic's documentation recommends 3 to 5 examples. Below that, the model lacks enough material to find a common pattern. Beyond that, the marginal gain drops and the prompt becomes needlessly long. What matters is not quantity but quality: examples relevant to the task, varied in the cases they cover, and consistent with each other in tone and structure.
- Does few-shot prompting work for every AI task?
- No. It is very effective for tasks about wording, format and tone (emails, customer replies, meeting notes, classification). The Prompt Engineering Guide notes its limits on complex reasoning tasks, where working through a step-by-step chain of thought produces better results than simply showing examples.
- Why do inconsistent examples make AI output worse?
- A language model looks for the common pattern across the examples it is given. If one example is formal and another casual, if one is short and another highly detailed, the model averages these contradictory signals instead of faithfully reproducing the best style. The result is a hybrid answer, less convincing than any single example on its own.