Zero-shot prompting asks a language model to perform a task using only the task specification, without input-output demonstrations. It is a practical starting point for testing an AI product quickly, because it uses existing model capabilities without fine-tuning. section 6.0 section 6.1.1
From sentiment labels to product content
A basic prompt might ask the model to determine whether text is positive or negative, producing the answer after “Sentiment:”. A richer zero-shot request asks: “Generate a blog article on Small Language Models for CIOs of large businesses. Make it engaging and actionable, showing ways to integrate SLMs into their IT operations.” Both prompts provide a task, not demonstrations. section 6.1.1 section 6.1.1
Figure 6.2 Example of an input-output prompt for sentiment analysis
Use the sentiment-analysis figure to see how a task-only prompt produces an output without examples. section 6.1.1
Use zero-shot prompting deliberately:
- Start with straightforward tasks or topics likely to be well represented in the model’s training data. section 6.5
- Improve output by adding context, detailed instructions, or a desired style while keeping the prompt demonstration-free. section 6.1.1
- Switch to few-shot prompting when a complex or unfamiliar task needs demonstrations; use chain-of-thought prompting when the task requires multiple reasoning steps. section 6.1.1 section 6.5
