Providing examples

Examples are powerful — and literal

Few-shot prompting works because the model picks up patterns from your examples. The catch is that it picks up all the patterns, including ones you didn't mean to teach. If every example is about shoes, it may drift toward shoe-related wording. If every example is exactly two sentences, you'll probably get two sentences. Choosing examples is about teaching the pattern you want and nothing else.

1. Make examples representative

Your examples should look like the real inputs the AI will face. If real customer emails are messy, rambling and full of typos, don't demonstrate with tidy, three-line emails. The model learns how to handle what it has seen.

2. Make examples diverse

Vary everything that should vary: topic, length, tone, difficulty. Keep constant only what must stay constant — usually the format.

The first version teaches a template so rigid the AI will probably write "Hi! Just a reminder about your library book..." and may even add "See you then!", which makes no sense for a library book. The second varies the subject, structure and closing line, so the model learns the tone rather than a fill-in-the-blanks sentence.

3. Cover the tricky cases

If your task has edge cases, include at least one example of each. For a sentiment classifier, show a sarcastic review and a mixed one, not just obvious positives and negatives. For data extraction, show what to output when a field is missing.

Extract the person's name and company. If a field isn't mentioned, write "unknown".

Text: "Thanks for the call — Priya from Northwind here."
Name: Priya | Company: Northwind

Text: "Got your invoice. Cheers, Marcus"
Name: Marcus | Company: unknown

Text: "Our team at Bluebird Logistics will review it."
Name: unknown | Company: Bluebird Logistics

Text: "[paste new text]"

Without the second and third examples, the model might guess a company, or invent a name, rather than writing "unknown".

4. Balance your labels

In classification tasks, roughly balance the examples across categories. If four out of five examples are labelled POSITIVE, the model may lean toward POSITIVE on borderline cases.

5. Think about order

Order can influence results. Models sometimes lean toward patterns in the most recent examples, so:

• Mix categories rather than grouping all of one type together.
• Don't end on an outlier unless you want the AI to treat it as typical.
• If results seem oddly skewed, try shuffling the examples and compare.

6. Prevent over-copying

The model may borrow phrases, facts or structure from examples directly. To reduce this:

• Say what to copy. "Match the format and tone, but write original content."
• Vary the wording across examples so no single phrase looks like a required part of the pattern.
• Use different subjects than your real input, so content can't leak across.
• Keep examples short when only the format matters.

That illustrative output is what you'd hope the improved reminder prompt produces: warm, short and specific to libraries — not a copy of the dentist template.

A quick quality check

Before sending a few-shot prompt, ask: If someone saw only these examples, what rule would they guess? If the answer includes anything you didn't intend — "always mention Tuesday," "always two sentences" — change the examples.