Lesson 03

Zero-shot vs few-shot: the pattern worth more than a bigger model

When to give the model examples and when not. Golden rule: if you can show it, don't describe it.

The million-dollar question

When is it worth spending tokens on examples?

Short answer: any time the format is unusual, the task is ambiguous, or the tone is specific.

Long answer: keep reading — without this, you stay mediocre.

Zero-shot — no examples

You ask the model to do something without showing it any case. Works very well for common tasks the model has “seen” in training.

const res = await client.messages.create({
  model: "claude-haiku-4-5",
  max_tokens: 100,
  messages: [
    { role: "user", content: "Translate to Spanish: 'Hello, how are you?'" },
  ],
});
// → "Hola, ¿qué tal?"

Tasks where zero-shot usually works:

• Translation between common languages
• Summarizing text
• Factual Q&A
• Rewriting in a different tone

Few-shot — 2 to 5 examples

You show several examples, then ask for a new one. The model learns the pattern, not the specific answer.

const res = await client.messages.create({
  model: "claude-haiku-4-5",
  max_tokens: 100,
  system: "You convert company names into URL-friendly slugs.",
  messages: [
    {
      role: "user",
      content: `Examples:
"Café & Tea Ltd." → cafe-tea
"My Restaurant 24/7" → my-restaurant-24-7
"Acme Corp." → acme-corp

Convert: "María Jewelry Boutique"`,
    },
  ],
});
// → "maria-jewelry-boutique"

Without examples, you might get MariaJewelryBoutique, mar%C3%ADa-jewelry, or anything else. With three examples, you nail the pattern.

The practical rule

If you can show the output, don’t describe it.

Compare:

Bad (zero-shot describing format):

Return a JSON object with fields name, age, and profession, all lowercase,
where profession is always an array even if it has a single value.

Good (few-shot showing format):

Examples:
"Ana, 30, engineer" → { "name": "ana", "age": 30, "profession": ["engineer"] }
"Luis, 45, lawyer and teacher" → { "name": "luis", "age": 45, "profession": ["lawyer", "teacher"] }

Convert: "María, 28, doctor"

The second prompt is shorter, clearer, and much more reliable.

How many examples

• 1 example (one-shot) → most cases. Enough to lock down format.
• 3-5 examples → when there are variations or edge cases to show.
• >5 examples → almost never. Past that, prefer fine-tuning or RAG.

Example bias

Watch out: the model copies patterns. If all your examples are similar in length, it will expect that length back. If all end with a period, it will too.

So vary on purpose:

"yes" → positive
"NO." → negative
"haha sure, totally" → negative (sarcastic)

Three examples, three lengths, three tones. You are teaching the model that the answer does not depend on shape, but on meaning.

Challenge

Design a few-shot prompt that takes product reviews and returns a score from 1 to 5.

Share it in the community Discord and other learners will spot the blind spots. In the next lesson, chain-of-thought: the trick that makes the model reason instead of guessing.