Few-shot prompting

Show, don't just tell

Few-shot prompting means including a small number of examples — usually two to five — of inputs paired with the outputs you want, then giving the AI a new input to handle the same way. Instead of describing the pattern, you demonstrate it.

Language models are very good at continuing patterns. When a prompt contains several input→output pairs, the model infers the rule connecting them — the format, length, tone, level of detail — and applies it to the next input. This is sometimes called in-context learning: the model isn't permanently retrained; it just picks up the pattern from what's in front of it for this response.

A simple example

Say you want short, punchy product taglines in a specific style. Describing that style is surprisingly hard. Showing it is easy.

The second prompt never says "short," "wordplay" or "under six words," yet the examples communicate all of that. The response is an illustrative one — you'd likely get a different but similar tagline each time.

The anatomy of a few-shot prompt

1. A brief instruction — what the task is. Examples work best with a sentence of guidance, not instead of it.
2. Examples in a consistent format — the same labels ("Product:", "Tagline:") and layout every time.
3. The new input — formatted exactly like the examples, with the output slot left open.

Consistency matters. If one example uses "Input/Output" and another uses "Q/A," the model has to guess which format to follow.

Convert each casual note into a clear, professional status update for my manager.

Note: finished the slides lol, waiting on Sam for numbers
Update: The presentation slides are complete. I'm waiting on Sam for the final figures and will add them as soon as they arrive.

Note: client call moved again?? now thurs
Update: The client call has been rescheduled to Thursday. I'll send an updated agenda beforehand.

Note: bug fixed i think, need to test more tmrw
Update:

When few-shot is worth it

Reach for examples when:

• Format must be exact — a specific JSON shape, a CSV row, a fixed template.
• Style is hard to describe — brand voice, humour, a particular writing rhythm.
• Labels are subjective — what counts as "urgent" in your support queue.
• You need consistency — processing many items the same way.

For everyday tasks the model already understands, zero-shot is usually enough. Examples cost extra words and can introduce their own problems.

How many examples?

Often two or three are enough to establish a pattern. Add more when the task has several distinct cases you need to cover (for example, one example per category in a classification task). Beyond that, returns tend to diminish, and very long example lists make the prompt harder to maintain.

Few-shot in a conversation

You don't always need to pack examples into one message. You can also correct the AI mid-conversation: "That's close — here are two I liked better: ..." This works similarly, though a single, well-structured prompt is easier to reuse and share.