Examples and few-shot
- prompting
- examples
“Friendly but professional, a bit playful but not silly” — anyone who has written that sentence knows the result: generic-friendly mush. Few-shot prompting replaces the adjective parade with something better: show two or three input → output pairs, then hand over your real input. The model extends the pattern. That is not a trick — it is the mechanism.
Why examples win
An adjective describes a region of style. An example is a point in it — sentences, rhythm, length, level of detail, all at once. When the model continues from your examples, style and structure come along automatically. It is next-token prediction doing what it does best: continuing an established pattern.
How to give examples well
- Match the real shape. Examples must look like the actual input; a format you would never receive teaches the wrong pattern.
- Two or three pairs. One is a point; three are a trend.
- Vary them deliberately. Identical phrasing in every example → outputs that clone your words instead of the style.
- Include one hard case if you have it. The edge case teaches the boundary faster than a rule.
- Then give the real input clearly — labels like “Input:” and “Output:” keep the roles obvious.
A bad example
Ask: “Rewrite these release notes in our brand voice — friendly, smart, a bit playful but not silly.”
You get a polite generic register that matches no brand — and no way to name what went wrong.
A better example
Ask: “Rewrite each release note as a short announcement, following the pattern of these two examples exactly — same structure, same length, same level of formality. Then here is the new input: …”
The examples carry tone, structure and granularity; your instruction only has to carry the task.
Why it works
Every example narrows the plausible space like a constraint you never had to verbalise. It is also the most testable prompting technique: when output misses, the fix is usually an example edit, not a longer instruction. Keep your best pairs as reusable prompt assets — three good examples outlive several rounds of prompt wording.
One maintenance tip: when a pattern prompt produces a great result, save it with the examples intact. The examples carried the value — swapping them for a new task is usually a smaller edit than writing a fresh prompt, and the pattern keeps paying without being re-derived.
Practice
Which prompt shows instead of telling?
You want release notes rewritten as short customer announcements in your brand's rhythm — a style reviewers always describe differently.
Which prompt is more likely to reproduce the intended style?
Hint
Which prompt could a new teammate follow without asking a single follow-up question?
Why this is the answer
Few-shot prompting shows the pattern instead of describing it: two or three input → output pairs, the same shape as real inputs, deliberately varied content. The instruction then only carries the task — style comes from the examples.
Transfer
- Brand voice: two approved paragraphs beat any tone adjectives.
- Data extraction: show two “messy line → clean fields” pairs.
- Support replies: three example answers per category; the model inherits the policy.
Key takeaways
- Two or three input to output pairs teach a pattern that adjectives cannot.
- Match the real shape, vary the pairs deliberately, include one hard case.
- When output misses, edit the example — not the instruction wording.
Next
Next: Work with long input — a chunked process for documents that don’t fit in one pass. The few-shot prompting glossary entry is the one-line version.