Few-shot prompting places a small number of worked examples inside the input sent to a large language model so it infers a task from the pattern rather than from retraining, with each example, called a shot or demonstration, pairing an input with the output its author wants and the model completing one final unanswered input in the same manner while its parameters stay frozen. A few-shot prompt has four parts, usually an optional instruction, the examples, the new input, and an open slot for the answer, and the terminology is a count: none is zero-shot, one is one-shot, and several is few-shot. OpenAI's GPT-3 paper fixed those conventions in 2020, defining few-shot as allowing as many demonstrations as fit the context window, typically 10 to 100, with GPT-3 accepting 2,048 tokens in total. Anthropic's documentation for Claude, as of September 2026, advises three to five relevant and varied examples wrapped in example tags, and a September 2024 Google whitepaper by Lee Boonstra gives the same rule of thumb. Examples can be fixed or chosen per request; Jiachang Liu and colleagues proposed dynamic selection in January 2021 after finding GPT-3's results depended heavily on which examples were picked, with retrieval of semantically similar examples consistently beating random selection. Every example is billed as input tokens, and prompt caching, added by Anthropic in August 2024 and by OpenAI two months later, discounts a repeated prefix.