Zero-Shot and Few-Shot Prompting
Pedal Point's ticket sorter keeps putting "My parcel is a week late and I'm furious" under returns. The instructions clearly say late parcels are order_status, yet the model reads anger and assumes the customer wants their money back. Rewriting the rule a fourth time is not working. Showing it will.
The rule
Zero-shot tells the model; few-shot shows it.
A zero-shot prompt describes the task and gives no examples:
Classify the email as order_status, returns, product_question or other.It works well for clear, common tasks. A few-shot prompt adds a handful of worked examples, each an input and the exact output you want:
Email: Where is my order 4417? It's been ten days.
Label: order_status
Email: The saddle is too narrow, how do I send it back?
Label: returns
Email: Does the Trail 2 frame fit 29-inch wheels?
Label: product_question
Email: {new email}
Label:Examples teach what descriptions struggle to: the precise format, the tone, and where the line falls between similar categories. A rule says "late parcels are order_status"; an example shows what a late parcel sounds like.
Choose examples with care, because the model imitates them closely:
- Aim at the boundary. Include the cases the model gets wrong, like an angry late-delivery email labelled
order_status. - Balance the labels. If most examples say
returns, answers drift towardreturns. - Vary the surface. Different lengths, products and wording, so the model learns the rule, not a phrase.
- Match the format exactly. Whatever your examples look like, the output will copy it.
A handful of examples is usually enough. More cost tokens on every request and give diminishing returns.
Examples also quietly fix format. If you need replies shaped exactly like Label: returns with nothing else, two examples in that shape often beat a paragraph explaining it. And when examples and instructions disagree, models often lean toward the examples, so update both together whenever the rules change.
The shortcut
Start zero-shot; add one example for each mistake you see. Each example should answer the question, "What did the model misunderstand?"
Run zero-shot
Late-delivery complaints with angry wording come back as returns.
The trap is testing only on the examples you added. Of course they pass; the proof is emails the prompt has never seen.
Drills
Answer fast — trust the rule.
A prompt describes the task but contains no worked examples. What is it?
Four of your five examples are labelled returns. What is the likely effect?
Which example most helps with the late-delivery confusion?
How should you check whether the new examples worked?
Remember
- Zero-shot tells the model; few-shot shows it.
- Start zero-shot; add one example per mistake.
- Balance labels, aim at the boundary, test on unseen inputs.