Few-Shot Prompting: Teach AI Your Standards in 3 Examples

A recruitment agency spent two weeks writing a 4-page prompt document describing how to screen CVs. Tone, criteria, format, scoring rubric, everything spelled out in painful detail. The AI still got it wrong half the time.

I replaced the entire document with three example CV screenings. Input CV, output assessment. The AI matched their standards on the very first try. That is the power of few-shot prompting: showing is faster and more accurate than telling.

How Few-Shot Prompting Works

Few-shot prompting is the technique of providing a small number of examples (typically 2-5) of the input-output pattern you want the AI to follow. The AI uses these examples to infer the task, format, criteria, and quality standard, then applies that pattern to new inputs.

Zero-shot means no examples. One-shot means one example. Few-shot means two or more. Research consistently shows that adding examples improves task accuracy, with the most significant jump happening between zero and two examples. Beyond five examples, returns diminish for most business tasks.

The reason this works so well is that large language models are fundamentally pattern-matching machines. When you give them examples, they extract implicit rules that would be difficult or impossible to articulate in words. Your example CV screening does not just show the format, it shows what level of detail matters, how to handle ambiguous qualifications, what tone to use, and dozens of other subtle choices that you make instinctively but could not easily write down.

Where Few-Shot Prompting Saves the Most Time

The highest-ROI applications share three characteristics: the task is repetitive, it requires consistent judgement, and the volume is high enough that automation pays for itself quickly.

1. Email Categorisation and Routing

Show the AI three examples of how you categorise incoming emails, one support request, one sales inquiry, one partnership pitch, with the category, priority, and routing action for each. Then feed it your inbox. A 50-person professional services firm I worked with reduced email triage time from 45 minutes per day to 8 minutes using exactly this approach.

2. Support Ticket Tagging

Provide examples of tickets tagged with severity, product area, and suggested first response. The AI matches your tagging standards across new tickets. One SaaS client processed 200+ tickets per week and reduced first-response time by 34% because tickets arrived pre-tagged and pre-prioritised.

3. Content Quality Assessment

Show the AI three blog posts or marketing emails with your quality scores and specific feedback. It learns your quality bar and can pre-screen drafts before they reach a human reviewer.

4. Lead Scoring

Provide examples of leads with their scores and the reasoning behind each score. The AI applies the same criteria to new leads, giving your sales team a pre-qualified pipeline.

5. Data Extraction and Formatting

Show three examples of unstructured text (invoices, contracts, meeting notes) converted into your structured format. The AI handles the rest of the batch following the same extraction pattern.

How to Choose Your Examples

Not all examples are equal. Here is the selection strategy I use with clients:

Cover the range. If you are categorising support tickets, include one easy/obvious ticket, one edge case, and one that requires judgement. This teaches the AI how to handle ambiguity, not just clear-cut cases.

Include a negative example. Show what the wrong output looks like and label it as incorrect. This is especially useful for quality screening tasks where you need the AI to reject inputs that do not meet your standards.

Match your real data. Use examples that look like your actual inputs, same length, same messiness, same formatting quirks. Polished examples teach the AI to expect polished inputs, which is rarely what it will actually receive.

Be consistent across examples. If your first example uses bullet points and your second uses paragraphs, the AI will not know which format to follow. Use the same structure in every example.

Few-Shot vs Zero-Shot: When to Skip Examples

You do not always need examples. Zero-shot prompting (instructions only, no examples) works fine for tasks where the standard format is universal and the AI's default is close enough to what you want. Writing an email, summarising a document, translating text, these rarely need examples.

Switch to few-shot when:

• Your output format is custom to your business
• The task requires subjective judgement (what counts as "urgent"? what counts as "high quality"?)
• You have tried zero-shot and the outputs are inconsistent
• You need multiple team members to get the same results from the same prompt

A Template You Can Copy

Here is the few-shot prompt structure I deploy most often:

Task description: [One sentence describing what you want the AI to do]

Example 1:
Input: [Real example of your data]
Output: [Exactly what you want the AI to produce]

Example 2:
Input: [Different type of input]
Output: [Corresponding output]

Example 3:
Input: [Edge case or tricky input]
Output: [How to handle the tricky case]

Now process this:
Input: [New data to process]

That is it. No elaborate instruction document. No 4-page prompt. Three examples and a task description.

Frequently Asked Questions

What is few-shot prompting?

Few-shot prompting is a technique where you provide 2-5 examples of input-output pairs to teach an AI your exact standards. The AI uses these examples to infer the task requirements, format, and quality criteria, then applies that pattern to new inputs. It is one of the most reliable ways to get consistent, business-specific results from any language model.

How many examples do I need?

Two to three examples cover most business tasks. The biggest accuracy improvement happens between zero and two examples. Beyond five examples, returns diminish and you start using unnecessary tokens. If three examples are not producing consistent results, the issue is usually example quality (too similar, too clean, or missing edge cases) rather than example quantity.

What kinds of tasks work best with few-shot prompting?

Repetitive categorisation and formatting tasks where you need consistent judgement at scale. Email sorting, support ticket tagging, lead scoring, content quality assessment, data extraction from unstructured text, and CV screening are the highest-ROI applications across my client base.

Can I use few-shot prompting with AI automation tools like Zapier or Make?

Yes. Most automation tools that connect to AI APIs let you include examples in your prompt template. The examples stay constant while the new input changes with each trigger. This is how you build AI-powered automations that match your specific standards rather than producing generic output.

What if the AI ignores my examples?

Check three things. First, are your examples formatted consistently? Inconsistent formatting confuses the pattern matching. Second, are you using too many tokens before the actual task? Move examples closer to the input. Third, try adding an explicit instruction: "Follow the exact format shown in the examples above." Some models need that nudge.

Richard Batt has delivered 120+ AI and automation projects across 15+ industries. He helps businesses deploy AI that actually works, with battle-tested tools, templates, and implementation roadmaps. Featured in InfoWorld and WSJ.

Put This Into Practice

I use versions of these prompting approaches with my clients every week. The full templates, prompt libraries, and implementation guides, covering the edge cases and variations you will hit in practice, are available inside the AI Ops Vault. It is your AI department for $97/month.

Want a personalised implementation plan first? Book your AI Roadmap session and I will map the fastest path from where you are now to working AI automation.