Flashcard: what is few shot prompting and what does the 'few' actually refer to?

**Few shot prompting** is the technique of putting a small number of worked examples inside your prompt, right before the actual question. The 'few' is literally how many demonstrations you include, usually between two and eight. A few shot prompt looks roughly like this: 'Classify the sentiment of each review. Example 1: "The food was amazing." Sentiment: positive. Example 2: "Cold and rude." Sentiment: negative. Example 3: "It was fine." Sentiment: neutral. Now classify: "Took an hour but the soup was great."' The model sees the pattern in the three examples and produces an answer in the same shape. Nothing is being retrained. The model's weights do not change between your calls. The examples are just text in the prompt, the same as any other text. The model is doing **in context learning**, which is its built-in ability to pick up patterns from whatever appears earlier in the same input. Few shot is the standard upgrade path from zero shotwhen zero shotquality is not good enough on its own.

**Few shot** prompting is the technique of including 2-8 worked input-output examples in the prompt, immediately before the actual query the model should answer. The number is the 'few'; you choose it based on how much context the model needs to lock onto the pattern. The mechanism is **in context learning**. The model treats the examples as part of its input, attends to them when generating the answer, and produces output in the same shape it observed. No weight updates happen. The examples affect this single forward pass only; the next call to the API starts fresh. Three practical rules govern how to write few shot prompts. First, place examples close to the query (usually at the end of the prompt) so attention can use them; the lost-in-the-middle effect makes examples buried 4000 tokens upstream less effective. Second, make the examples diverse so they cover the kinds of inputs you actually expect in production; three almost-identical examples teach less than three that span the input distribution. Third, keep the format exactly consistent across examples; the model will copy whatever pattern you give it, including any inconsistencies. Few shot is the standard upgrade from zero shotfor tasks with unusual output formats, domain-specific label sets, or any case where 'just tell the model what you want' is not enough.