A three-person startup has only 20 labeled examples of customer messages tagged as positive, negative, or neutral. They want a Foundry-hosted model to classify new messages the same way by tomorrow, without the time or budget to curate a large training set. Which approach best fits their constraints?

The best fit is including the 20 labeled examples directly in the prompt as few-shot examples. Placing those labeled messages in the prompt lets the model infer the startup's specific labeling pattern from real demonstrations and apply it to new messages immediately, without any separate training step, which fits both the tight timeline and the small budget. Fine-tuning a custom model on the 20 examples before deploying is incorrect because fine-tuning typically needs a much larger labeled dataset than 20 examples to reliably learn a pattern, and it also takes more time and compute than the startup can spend before tomorrow. Waiting until they collect thousands more labeled examples before building anything is incorrect because it abandons the deadline entirely and is not a solution to the problem as stated. Using zero-shot prompting with no examples, relying on the model's general knowledge of sentiment, is incorrect because it gives the model no examples of how this particular startup wants messages labeled, so it falls back on generic notions of sentiment that may not match their specific categories or edge cases, producing less consistent results than showing the model real, labeled examples. Few-shot prompting is well suited to exactly this situation: a small number of representative examples, included right in the prompt, that steer the model's behavior without the overhead of training a new model from scratch.