What few-shot prompting in 2020 still teaches content teams

Few-shot prompting taught early GPT-3 users that a model learns a voice from examples far better than from adjectives. In 2026, with fluent instruction-following models, the lesson still holds: teams that feed models curated, audience-grounded examples get distinctive content, while teams that only describe their tone get the same average copy as everyone else.

What did writing with GPT-3 actually look like in 2020?

In May 2020 OpenAI published "Language Models are Few-Shot Learners", the paper describing GPT-3, a 175-billion-parameter model. Its central claim was that a large enough model could pick up a task from a handful of examples placed in the prompt, without retraining. In June 2020 the API opened as a private beta, and that summer is when GPT3 Marketing started working with it.

There was no chat window. You wrote a prompt that looked like the beginning of a document: three or four example product descriptions, each with a short input and a finished output, then a new input and an empty space. The model continued the pattern. If your examples were bland, the output was bland. If one example was much longer than the others, the outputs drifted long. Every quirk in the examples showed up, amplified, in the result.

It was a demanding way to work, and it forced a discipline that most teams lost when chat interfaces arrived in November 2022.

Why do examples beat adjectives?

Ask a modern model to write in a voice that is "warm, confident and witty" and you will get something warm, confident and witty in the way that thousands of other brands are. Those adjectives sit in every brand book. The model has seen them attached to an enormous range of text and returns the centre of that range.

Examples carry information adjectives cannot: sentence length, where the joke lands, which words you never use, how you handle a complaint, whether you write "we" or "the team". A good example is a compressed voice guide. That is why our brand voice systems are built around worked examples rather than tone descriptors.

Instructions describe what you want. Examples prove it. Models trust proof.

Which 2020 habits are worth keeping?

1. Curate examples like a portfolio

In 2020 every example cost tokens and money, so we chose them carefully. Today context is cheap and teams paste in whole archives. More is not better. Pick the pieces that best represent the voice at its strongest and most specific, and vary them across formats so the model learns the voice, not one template.

2. Make the pattern explicit

Few-shot prompts worked best with consistent structure: a labelled input, a labelled output, a clear separator. The same principle applies to modern prompts. Show the model what goes in and what comes out, so it knows which parts of an example are fixed voice and which are variable content.

3. Read every output as an editor

Nobody published raw GPT-3 text. It drifted, repeated itself and occasionally invented facts with complete confidence. Modern models are far better, but the editor's habit is still the cheapest quality control available. We cover that in more detail in why AI drafts still need heavy editing.

4. Treat the prompt as an asset

A good few-shot prompt took hours to tune, so we versioned it, documented it and reused it. That evolved into the prompt library practice we run for clients today.

Where should the examples come from?

This is where 2026 differs from 2020. Back then, examples came from the brand's own archive. The trouble is that archives reflect what the brand wanted to say, not necessarily what the audience responded to. If your best examples are your own favourite posts, the model learns your preferences, not your audience's.

We now ground example selection in audience evidence. Using SOMIN's strategy tools, SOMIN being an AI audience-research platform, we look at the conversations around a category: the phrases people actually use, the tensions they voice, the tones that earn replies rather than scrolls. Then we pick or write examples that speak to those specifics. The model inherits the brand's voice and the audience's vocabulary at the same time.

The Fujifilm case study is useful further reading on how audience insight shapes what a brand says, before anyone touches the wording.

A worked example: one product line, two prompts

Imagine a skincare brand writing descriptions for a new serum.

• Prompt A says: "Write a warm, expert product description for our new vitamin C serum. Our tone is friendly and science-led."
• Prompt B includes three past descriptions that performed well, each paired with its brief, plus a short note of phrases the audience uses about the problem (for instance, how they describe dullness or sensitivity in their own words), then the new brief.

Prompt A returns competent copy that could belong to any brand on the shelf. Prompt B returns copy with the brand's rhythm and the audience's language. The difference is not model quality. It is input quality.

Checklist: a few-shot prompt for brand content

• Two to five examples, varied in format and length
• Each example paired with the brief that produced it
• Consistent labels and separators
• A short list of audience phrases and tensions, sourced from real conversations
• A short list of things the brand never says
• A named human reviewer for the output

Definitions

• Zero-shot: asking a model to do a task with instructions only.
• Few-shot: including a small number of worked examples in the prompt.
• In-context learning: the model adapting to patterns in the prompt without changing its weights.

Why this matters now

When everyone has access to the same models, the advantage moves to inputs. Few-shot discipline was born from scarcity in 2020; in 2026 it is a way to escape sameness. If you want to see how audience signal is turning into practical marketing decisions elsewhere, the weekly episodes from Marketing Mondays decode one real signal at a time, which is a good habit to pair with a sharper prompt.

The brands that sound distinctive with AI are rarely the ones with the cleverest instructions. They are the ones with the best examples, chosen with their audience in mind.

Frequently asked questions

What is few-shot prompting?

Few-shot prompting means giving a language model a small number of worked examples of a task inside the prompt, so it infers the pattern and continues it. The GPT-3 paper in 2020 popularised the term.

Is few-shot prompting still useful with modern models?

Yes. Instruction-tuned models follow plain requests, but examples remain the most precise way to transfer a specific voice, structure or level of detail. Instructions describe; examples demonstrate.

How many examples should a content prompt include?

Usually two to five strong, varied examples. Too few and the model over-copies one; too many and you crowd out the brief. Quality and variety matter more than count.