Addaly is in open beta. Things will change, and AI answers can be wrong — check anything that matters.

Show it two examples

Getting Real Work Out of a Model · lesson 4 of 8 · 8 min

Some things are easier to show than to describe

Try writing down, in words, the rule that decides whether a customer message is about a refund or about delivery. You will get three sentences in and hit the case where a parcel arrived smashed and the customer wants their money back. Is that delivery, because a parcel broke? Or refund, because they are asking for money?

You can keep writing rules. Or you can show four labelled examples and let the boundary transfer itself.

Message: "Where is my order? It said Tuesday and it's Friday."
Category: delivery

Message: "I want my money back, the shoes don't fit."
Category: refund

Message: "My parcel arrived smashed and I want my money back."
Category: refund

Message: "Do you ship to Ghana?"
Category: other

The third example is the one doing the work. It says: when a message mentions both, classify by what is being asked for, not by what went wrong. That sentence took you two minutes to think of and one example to communicate.

This is few-shot prompting. You give a small number of input-and-output pairs before the real input. It works because the pattern in your examples is a much more precise instruction than most prose descriptions.

Where it earns its place

  • Category boundaries, as above. Always include the awkward case.
  • Tone and house style. Three real replies your team has sent will convey your voice better than a paragraph of adjectives. "Friendly but not chummy" means nothing. Three emails mean something.
  • Layout you cannot easily name. Show one finished entry and ask for forty more like it.
  • Small consistent transformations. Product name to URL slug, address to a standard postal format, messy date to YYYY-MM-DD.

What examples teach that you did not intend

This is the part that catches people. Examples teach their whole shape, not just the part you were thinking about.

If all three of your example replies happen to end with an apology, new replies will tend to end with an apology, including for messages where nothing went wrong. You never wrote "apologise". You demonstrated it three times, which is stronger.

The same goes for length. If every example output is one sentence, you have quietly capped the length. If every example is a long paragraph, short answers stop appearing.

And it goes for balance. If four of your five classification examples are labelled refund, refund starts to look like the safe answer. Include the categories in roughly honest proportion, and include at least one example of the thing you are worried it will over-produce.

The fix is the same in each case: read your examples as a set and ask what they all have in common. Whatever that is, you have taught it.

When examples make things worse

Inconsistent examples are worse than none. If example two contradicts example four, you have taught confusion, and you will get confusion back with confident formatting.

Open creative work. Give three example poems and you often get a fourth poem in the same groove, competently and without surprise. If you want range, describe the constraint and leave the shape open. If you want your groove exactly, examples are perfect. Know which you are asking for.

When the rule is simple and sayable. "Reply in Portuguese" needs no demonstration. Save examples for the things prose is bad at.

A practical rule

Two to five examples covers almost everything. Beyond that you get diminishing returns and you spend context you might need for the actual work. Pick the boring case, the awkward case, and the case you keep getting wrong.

Then keep them in a file. That set of four labelled messages is worth more than any prompt template you will find online, because it encodes decisions only you can make.

Before you move on

You give three examples of a customer message and its reply. All three of your example replies happen to end with an apology, though you never mentioned apologies anywhere in the instructions. What is most likely to happen with new messages?

Pick the one you would defend. Nobody sees your answer.

No ads. No data sale. No public scores on people. Ever.

© 2026 Addaly