Few-Shot Prompting for Marketers: Show, Don't Tell

Few-shot prompting means putting a handful of worked examples in your prompt - real input, real output - and letting the model copy the pattern instead of interpreting your description of it. It is the single technique I would teach a marketer who only learns one, because most prompt failures I see are description failures: “punchy but professional” means something different to every reader, including the model. Your three best subject lines mean exactly one thing.

The technique has been load-bearing since the beginning of this whole era. The paper that introduced GPT-3 is literally titled “Language Models are Few-Shot Learners”, and its core finding was that a big enough model given a few demonstrations in the prompt could match systems that had been custom-trained for the task. Every model you use today inherits that property. This guide is how I use it for marketing work, where the examples come from, and the two failure modes the research says to watch for.

What few-shot prompting is, in one prompt

The format could not be simpler. Instead of describing the output you want, you demonstrate it:

You write product descriptions in our house style. Match the style
of these examples exactly.

<example>
Input: waterproof hiking jacket, 3 colors, $189
Output: Rain is a scheduling problem now. The Ridgeline jacket seals
out weather through a 20k membrane and vents it back out when the
climb gets steep. Three colors, none of them neon.
</example>

<example>
Input: merino wool running socks, cushioned, $24
Output: Your feet log the same miles you do. These merino socks
manage sweat on mile one and blisters on mile ten, with cushioning
exactly where the pavement hits. One price, no bulk-buy games.
</example>

Now write one for: [your product, features, price]

No adjectives about voice anywhere in that prompt. The examples are the voice spec: sentence length, the dry humor, the concrete-benefit-first structure, the anti-hype stance in “no bulk-buy games.” OpenAI’s prompt engineering guide describes the mechanism plainly: the model implicitly picks up the pattern from the examples and applies it to the new input. My experience matches the docs: things I could never get by instruction - rhythm, restraint, what a brand doesn’t say - transfer through examples on the first try.

How many examples, and what makes a good one

Anthropic’s guidance for Claude is the crispest statement of current practice, and it matches what I do across every model: include 3 to 5 examples, and make them relevant, diverse, and structured - mirroring your actual use case, covering enough variety that the model doesn’t pick up unintended patterns, and wrapped in tags so the model can tell examples from instructions. The same page calls examples one of the most reliable ways to steer output format, tone, and structure, which has been exactly my experience.

The word doing the most work in that list is diverse, and it is where marketers slip. If all three of your example subject lines are questions, you have taught the model that subject lines are questions. If both product descriptions are for outdoor gear, watch it get weirdly outdoorsy about your accounting software. My rule for assembling a set: each example should share the pattern I want copied and differ in everything else - product category, length, angle. The shared element is the lesson. Everything the examples accidentally share gets learned too.

Which leads to the check I run on every example set, and it takes two minutes: read the examples and list what they have in common besides the thing you want. That list is what you are actually teaching.

Where the examples come from

The single biggest quality lever is not prompt phrasing, it is example provenance. My sourcing rules:

Your best shipped work, not your average work. The model will write to the level of the examples. Three of your greatest hits produce a different ceiling than three passable pieces pulled from last month. I keep a small file of best-ever pieces per format - subject lines, product blurbs, post intros - and that file quietly became my most valuable prompt asset.

Never model output. Feeding AI-written examples back in as the standard compounds averageness with each generation. Examples are where human craft enters the system, and it is the one place I am strict about provenance.

Real inputs attached. An example is input plus output, not output alone. The model learns the transformation - what you do with a feature list, how you compress a briefing into a hook. Output-only examples teach style but not judgment.

Claims stripped or generic. Anything in an example gets treated as a legitimate pattern, including your claims. If your example copy cites “12,000 five-star reviews,” expect review counts to appear, invented, in future outputs. I genericize numbers in examples or pick examples without claims, because every unverified number is work downstream.

The biases the research warns about

Few-shot prompting is powerful and twitchy, and there is good science on the twitchiness. Zhao and colleagues’ “Calibrate Before Use” found that the choice of prompt format, the examples, and even the order of the examples can swing accuracy from near chance to near state of the art, driven by biases like favoring answers that appear near the end of the prompt or that were common in training data.

For classification-style marketing tasks - sentiment tagging, lead scoring against criteria, sorting queries by intent - the practical translations I use:

• Balance the labels. Three positive examples and one negative teaches “mostly say positive.” Match the mix roughly to reality, or go even.
• Shuffle the order when results look off. Recency bias is real: the last example drags. If the model’s outputs all resemble example three, that is not mysterious, it is documented.
• Spot-check against known cases before trusting any few-shot classifier on volume, exactly the same validation discipline as any model-assisted analysis.

For generation tasks the same biases show up softer - the last example’s style bleeding into everything - and the same shuffle test diagnoses it.

Where few-shot fits in the prompt stack

Examples are one layer of a system, not the whole system, and knowing what belongs where saves you from bloated prompts. My division of labor: facts and rules go in the system prompt - what the product is, what may never be claimed, banned words. Patterns go in examples - voice, structure, the shape of a good output. Instructions carry only what changes per task. When I see a prompt with four paragraphs describing tone, that is examples-shaped work being done badly by instructions. When I see examples used to sneak in facts, that is system-prompt work being done unreliably by examples.

The one standing few-shot investment I recommend to every team: turn your brand voice guide into a set of tagged examples once, and reuse the set everywhere. Voice is the thing few-shot does best and instructions do worst, and it is the same three-to-five examples for every task in the prompt library. In a chained bulk workflow, that same example set gets re-attached at every stage that produces prose, because examples do not inherit from one prompt to the next. The same show-don’t-tell logic drives multi-turn refinement, where the model’s own last draft becomes the example and your follow-up message marks it up.

FAQ

What’s the difference between few-shot and zero-shot prompting? Zero-shot is instructions only, no examples. Few-shot adds worked demonstrations. Zero-shot is fine for tasks with an obvious right shape - summarize this, list the claims in this draft. The moment the output has a style - and almost everything a marketer ships has a style - examples beat descriptions.

Do examples use up too much of the prompt? Three short examples cost a few hundred tokens and pay for themselves in fewer revision rounds. For long-document tasks the ordering guidance matters more than the count: current model documentation recommends placing long reference material at the top of the prompt, with instructions and examples after it.

Can I ask the model to generate the examples? As drafts to react to, sure. As the actual example set, no - that recycles the model’s average taste as your standard. The useful version of model help here is critique: Anthropic’s docs suggest asking the model to evaluate your examples for relevance and diversity, which I do use, keeping the final pick human.

How do I know few-shot is working? Run the same task zero-shot and few-shot on one real input and compare. The difference is usually visible in a single read: the few-shot version sounds like you. If it does not, the examples are either too few, too similar to each other, or not actually representative of what you want - the fix is in the set, not in more instruction text.

This post is part of the prompt library, where the layer under it is covered in system prompts for marketing consistency and the best standing use of it in brand voice guidelines as prompts. The wider practice starts at the complete guide to AI marketing.