プロンプトエンジニアリング

Few-Shot Prompting Explained with Examples

Few-shot prompting explained simply: the definition, the 4-block structure, 5 copy-paste prompts with examples, and the mistakes that ruin your results.

投稿者: Jérémie Chiari

2026年9月13日読了時間 10 分

Hand typing on a keyboard with code on the screen

You ask ChatGPT a question and get an answer that's technically fine but flat, too long, or badly formatted. You rewrite your instructions three times and the result barely moves. The problem is rarely the model: it's that you're describing what you want instead of showing it.

Few-shot prompting fixes exactly that. The technique fits in one sentence: before you make your real request, you slip in two to five examples of the output you expect. The AI spots the pattern and reproduces it. It's the fastest way to get a stable format, a consistent tone, and answers you can use without editing.

Below you'll find the definition in plain language, the difference between zero-shot, one-shot, and few-shot, the four-block structure of a prompt that works, five complete prompts you can copy as-is, three ways to practice, and the mistakes that make the method fail. No technical skills required: everything happens in a normal chat window.

What exactly is few-shot prompting?

Few-shot prompting means dropping a few example pairs of "request / ideal answer" into your prompt, so the AI copies that model instead of improvising its own.

The word shot here simply means "example." So we talk about a prompt with zero examples, one example, or several examples. Nothing more complicated than that.

Technique	Examples provided	When to use it
Zero-shot	0	Simple request, open-ended answer: "Summarize this text in 5 lines"
One-shot	1	You want to enforce a specific format on a simple task
Few-shot	2 to 5	Tone, style, classification, industry vocabulary, repetitive format

The term took off with the research paper Language Models are Few-Shot Learners, published in 2020, which showed that a large language model could handle a brand-new task purely from a few examples placed in the instructions, with no retraining at all. Today the technique works on every consumer assistant: ChatGPT, Claude, Gemini, Mistral.

Keep this in mind: you are not training the AI. Your examples apply to the current conversation only, a bit like a template document you'd leave on a new coworker's desk before handing them a task.

Why a few examples beat a long set of instructions

An example conveys in three lines what written instructions would need a full paragraph to explain, and it leaves no room for interpretation.

Take an instruction like "answer in a professional but warm tone, and keep it concise." Everyone has their own definition of "warm" and "concise." So does the AI. Show it a four-line reply that opens with a thank-you and ends with an open question, and the ambiguity vanishes.

In practice, a good example carries six pieces of information at once that you'd never have thought to spell out:

the expected length, down to the word;
the exact format (list, table, paragraph, single sentence);
the tone and point of view (you, we, they);
the vocabulary of your industry, including your in-house phrasing;
the internal structure (hook, argument, call to action);
how to handle edge cases (what to do when information is missing).

These models work by continuing patterns: when they see three blocks built on the same logic, producing a fourth one from the same mold becomes the path of least resistance. It's the direct complement to the fundamentals covered in our complete method for writing a good prompt: context says what, examples show how.

The structure of a few-shot prompt that works

Four blocks, always in the same order: context, instruction, examples labeled identically, then your real request deliberately left hanging.

Block 1, the context. Two lines are enough: who's speaking, to whom, in what setting. "You're the customer service lead at a hiking gear store."

Block 2, the instruction. What you want, in one sentence. "Write a public reply to each customer review posted on Google."

Block 3, the examples. Two to five pairs, all built with the same labels. This is the heart of the technique and it's where most people slip: if you write "Review:" in the first example and "Customer comment:" in the second, you break the pattern.

Block 4, the new input. You present your real case with the same label, then you write the output label and stop. That blank is a strong signal: the AI understands it has to fill it in.

Review: Fast delivery, but the size doesn't match the guide. Reply: Thanks for the feedback, and we're glad the delivery lived up to expectations. We're sorry about the sizing gap: our guide has just been corrected. Write to us at contact@example.com and we'll arrange a free exchange.

Review: Product as described, packaging damaged on arrival. Reply: Thanks for taking the time to write to us. Damaged packaging isn't normal, and we're flagging it with the carrier today. If the product itself was affected, reply to this message and we'll replace the item.

Review: Three weeks of waiting and zero response from customer service. Reply:

Two formatting rules make all the difference. First, keep your separators strictly identical from one example to the next. Second, vary the cases: a positive review, a mixed one, a negative one. Anthropic's documentation on multishot prompting recommends three to five relevant and varied examples, precisely so the model grasps the general rule rather than one special case.

Five few-shot prompting examples to copy

Here are five complete prompts, on tasks you can test in the next ten minutes; just swap in your own content.

1. Responding to customer reviews

That's the example above. Three reviews covering the three possible tones, and you get consistent replies anyone on the team can publish. Typical gain: from about ten minutes per review to under a minute of proofreading.

2. Sorting messages by urgency level

Classify each message as URGENT, NORMAL, or FYI, then give the action to take in five words maximum.

Message: The order page has been throwing a 500 error since 9 a.m. Classification: URGENT | Action: alert the tech team

Message: Can you send me the notes from Tuesday's meeting? Classification: NORMAL | Action: send the document

Message: The September newsletter went out this morning. Classification: FYI | Action: none

Message: A customer is threatening to cancel their contract tomorrow. Classification:

The pipe and the capital letters aren't decorative: they make the output easy to paste into a spreadsheet.

3. Turning a feature into a customer benefit

Feature: 5000 mAh battery Benefit: Two full days without hunting for an outlet

Feature: Recycled aluminum case Benefit: As tough as a standard model, with a smaller footprint

Feature: 120 Hz display Benefit:

Three lines, and the model has understood that you want short, use-focused phrasing with no marketing adjectives. Written instructions would have taken a whole paragraph.

4. Extracting information from a text

Extract the information as a table row: Name | Company | Need | Deadline. Write "not specified" if the information is missing.

Text: Hi, I'm Marc Dupuis from Verlaine LLC, we're looking for AI training for six sales reps before December. Row: Marc Dupuis | Verlaine LLC | AI training for 6 sales reps | December

Text: Hey, this is Léa, I'd like some info on your pricing. Row: Léa | not specified | pricing information | not specified

Text: [paste your email here] Row:

The second example exists purely to show how to handle missing information. Without it, the AI invents company names. It's the highest-return type of example: the one that covers the awkward case.

5. Writing in your own style

Paste in three of your own texts (three posts, three emails, three intros) preceded by their topic, then add the topic of the fourth and stop. No description of your style will ever beat your actual sentences. It's the best way to avoid the bland text you can spot from a mile away.

Where to practice few-shot prompting: three methods

The most effective approach is practicing on real cases with structured feedback, then experimenting freely in consumer tools.

Method 1: Skilzy's program and AI Lab

Skilzy is a French e-learning platform for learning about AI and, above all, actually using it. The prompt fundamentals program covers few-shot step by step, with graded exercises and reusable templates for your own tasks. What sets it apart is the built-in AI Lab: you write your prompts in the real tools, credits included, without stacking up subscriptions on the side.

The platform has more than fifteen programs (images, video, UGC, faceless channels, avatars, voiceover, music, automation with n8n, AI-assisted coding) and prepares you for two state-recognized certifications eligible for funding: RS7439 in AI content marketing and RS6792 in AI and sales. Subscriptions start at €29.90 per month with no commitment, and there's a B2B offer for training organizations.

To try before you pay, the discovery demo gives you 7 days with 1 image, 1 video, 1 music track, and 10 messages, no credit card required (an account is enough).

Method 2: straight into ChatGPT, Claude, or Gemini

The free versions of these three assistants are more than enough to practice few-shot. The method: pick a task you repeat every week, write three examples of the perfect output by hand, and compare the result with and without examples. The gap will convince you in a single session.

Keep the prompts that work in a simple document, sorted by task. A well-built few-shot prompt stays reusable for months: you only change the last block.

Method 3: the prompt libraries from the AI providers

Anthropic, OpenAI, and Google publish prompt libraries and prompt-engineering guides for free. The point isn't to copy the prompts as-is, but to observe how the examples are labeled and separated. One caveat: these resources are mostly in English and written for developers, which makes them less comfortable when you're just starting out.

The mistakes that sink a few-shot prompt

Most failures come down to three things: examples that are inconsistent with each other, examples that are all alike, or mediocre examples that the AI faithfully reproduces.

Changing labels midway. "Review:" then "Comment:" is enough to blur the pattern. Copy-paste your structure.
Giving examples that are too similar. Three positive reviews, and the model will answer a furious complaint cheerfully. Cover the extremes.
Providing sloppy examples. The AI imitates what you show it, typos included. One rushed example produces ten rushed answers.
Adding fifteen of them. Past five, you're lengthening the prompt without gaining precision, and you risk locking the model onto irrelevant details.
Forgetting the instruction. Examples show the how, not the why. Keep one sentence that states the goal.
Writing the output for the AI. End on the empty label. If you start the answer, you'll get a variation of your own sentence.

These habits line up with the traps listed in our article on prompting mistakes that ruin your results. And when the task calls for reasoning rather than formatting (calculation, analysis, decision), combine few-shot with chain-of-thought, which makes the AI reason step by step: that way you show not just the answer, but the path to it.

Key takeaways

Showing costs less than explaining. Two to five well-chosen examples, labeled the same way, varied, and ending on a blank to fill: that's the whole technique. Take a task you repeat every week, build the prompt once, reuse it a hundred times.

よくある質問

How many examples should a few-shot prompt include?

Between two and five in the vast majority of cases. A single example is enough to enforce a simple format. Three is the right default setting: one typical case, one hard case, and one where information is missing. Past five, you're lengthening the prompt without gaining quality, and you even risk making the AI too rigid, fixated on details from your examples rather than on the general rule.

What's the difference between zero-shot and few-shot prompting?

In zero-shot, you give an instruction with no examples at all: "Summarize this text in five bullet points." In few-shot, you add several examples of the expected result before your real request. Zero-shot is faster and works well for open-ended tasks. Few-shot becomes clearly superior as soon as a specific format, a particular tone, or industry vocabulary comes into play, and whenever you have to repeat the same task several times with consistent results.

Does few-shot prompting work with the free version of ChatGPT?

Yes. The technique doesn't rely on any paid feature: it depends entirely on how you write your message. It works with the free versions of ChatGPT, Claude, Gemini, or Mistral, in the usual chat window. Paid plans mainly bring longer context windows and more capable models, which helps when your examples are long, but that's not a requirement to get started.

Do my examples actually train the AI?

No, and this is a common mix-up. Your examples don't train the model: they steer it for the duration of the current conversation. Open a new chat and you have to provide them again. That's why it's worth saving the few-shot prompts that work in a document, sorted by task. Some tools offer custom assistants where you can store those examples once and for all.

Can few-shot be combined with other prompting techniques?

Yes, and it's actually recommended. The most useful combination pairs few-shot with chain-of-thought: in each example, you show not only the final answer but the reasoning that leads to it. The model then reproduces the process, which markedly improves analysis, calculation, and decision tasks. You can also add a role at the start of the prompt and format constraints at the end, without ever touching the structure of your examples.

How do I know if my few-shot prompt is working?

Run a two-step test: send your request without examples, then with them. If the second answer is directly usable with no tweaks to format or tone, you've nailed it. If the result stays uneven from one attempt to the next, the problem almost always lies in the examples: they're too similar to each other, they don't use the same labels, or one is missing that shows what to do when information is absent.