Few-shot prompting is a way of instructing an AI model by including a handful of worked examples in the prompt, so the model follows the same pattern in its own answer. Zero-shot means giving instructions with no examples; one-shot means giving a single example; few-shot usually means two to five.
How few-shot prompting works
Language models are very good at continuing patterns. When you show them two or three pairs of input and output, they infer the rule connecting them and apply it to the next input, without any retraining. Nothing about the model changes: the examples sit in the prompt for that request only, which is what separates this approach from fine-tuning.
A simple example for classifying customer enquiries might look like this:
- “Do you cover Croydon?” → Service area
- “How much is a boiler service?” → Pricing
- “My engineer didn’t turn up today” → Complaint
- “Can I change my Thursday appointment?” → ?
The model sees the pattern and answers “Booking change” or similar. You could describe the categories in words, but the examples show the boundaries between them more precisely than a description usually can.
The same technique works for tone. If you want product descriptions in your brand voice, paste three of your best existing descriptions, then ask for a new one. The model picks up sentence length, vocabulary and structure far more reliably than it does from an instruction such as “friendly but professional”.
Why it matters
Most disappointing AI output comes from vague instructions. Few-shot prompting is the quickest fix available to a small business without technical help. It makes output more consistent, which matters when several people in a team are using AI for the same task, and it keeps copy closer to your tone of voice rather than the generic style these tools fall back on.
It is also useful for repetitive marketing jobs: drafting meta descriptions in a set format, tagging reviews by theme, rewriting supplier specifications into plain British English, or sorting search terms by intent.
Common mistakes
- Examples that are too similar. If all three examples are about the same product, the model may copy details that should change. Vary them.
- Accidental patterns. If every example is exactly 40 words or starts with “Discover”, the model will copy that too.
- Weak examples. The model copies quality as well as format. Mediocre examples produce mediocre output.
- Unbalanced classes. In a classification prompt, four complaints and one booking query nudge the model towards “complaint”.
- Pasting real customer data. Use made-up or anonymised examples unless your AI tool’s terms and your privacy notice cover the use of personal data.
How to act on it
Pick one repetitive task and write a short prompt in this order: what the task is, two to four varied examples of excellent output, then the new input. Save it as a template so everyone uses the same version. If the results drift, swap in better examples before rewriting the instructions. For persistent rules such as spelling, banned phrases or audience, put them in a system prompt or custom instructions so they apply every time.
Before relying on a template, test it on ten real inputs you already know the right answer for, and count how many it gets right. If it misclassifies the same type of input twice, add an example that covers that case. This small check is what turns a clever prompt into a process the whole team can trust.
For content that has to rank, AI drafts still need a person to add first-hand knowledge and check facts. How I fit AI into research and drafting without diluting quality is part of my content SEO and strategy service.