Few-shot prompting is one of the cheapest levers a growth team can pull when building on large language models. Instead of retraining a model, you place a small set of worked examples inside the prompt and let the model imitate the pattern. Each example pairs an input with the output you want, and the model completes a final unanswered input in the same style.
GPT-3 made the technique mainstream in 2020. The model’s parameters stay frozen; examples affect only the current response. That turns consistency tasks into a text problem rather than a machine learning project.
What a few-shot prompt looks like
A typical prompt has four parts: an instruction, the examples, the new input, and an open answer slot. For marketing, a search intent classifier could look like this:
- Classify each query as transactional, informational or navigational
- Query: cheap running shoes sale — transactional
- Query: how to clean suede trainers — informational
- Query: zalando returns login — navigational
- Query: best trail shoes for wide feet — [blank]
Zero-shot uses no examples, one-shot uses one, and few-shot uses several. OpenAI’s GPT-3 paper defined few-shot as any number that fits the context window, often 10 to 100 demonstrations. Today’s vendor guidance is much leaner: Anthropic and Google documentation generally suggest three to five relevant, varied examples.
Why marketers should care
Few-shot prompting is a practical tool for the repetitive work that eats agency hours: tagging search terms, classifying landing pages against a taxonomy, extracting product attributes from feeds, or rewriting hooks in a house style. It helps lock in the format and label space you need.
Amazon’s contextual targeting work shows what this can look like at scale. The company combined human annotators with conversational agents and chain-of-thought prompting, and reported a 25% accuracy lift over crowd-sourced annotation.
But there is a hard limit. IAB Content Taxonomy 3.1 has more than 1,500 categories, and a few examples cannot demonstrate every possible label. Teams building internal classification tools need to balance prompt examples with retrieval or deterministic matching.
Where it fails
- Bias by arrangement: unbalanced examples create majority label bias, and the most recent examples carry extra weight.
- Costs repeat every call: demonstrations are billed as input tokens each time, unless prompt caching discounts a repeated prefix.
- Labels may matter less than expected: researchers found that replacing correct labels with random ones barely hurt some models, though follow-up work contested that.
- Reasoning models can react badly: OpenAI, DeepSeek and Microsoft research all point to zero-shot first for o-series, R1 and o1-preview.
- Security risk: Anthropic showed many-shot jailbreaking succeeded with 256 faux dialogues, even though five shots failed.
Match the method to the model
Vendor advice is not uniform. Google’s Gemini API documentation recommends including few-shot examples; Google Cloud warns that too many examples can overfit. Anthropic suggests starting minimal and adding examples only when failures appear. For reasoning models, DeepSeek’s R1 paper reported that few-shot prompting consistently degraded performance, so zero-shot is often the better first test.
For marketing teams, the takeaway is to treat examples as a tuning dial, not a default. Start with a clear instruction and output format. Add three to five relevant examples for classification or formatting tasks where consistency beats creativity. Then test: different example order, different labels, and a zero-shot baseline can all change results.
Source: PPC Land
