การเขียนพรอมต์แบบ Zero-shot และ Few-shot

LLMs often need guidance. With zero-shot you give the task and no examples; with few-shot you include a few examples to show what you expect.

Zero-shot prompting gives a task with no examples - the model leans entirely on its pre-trained knowledge. Efficient, and great for common, straightforward tasks.

Here's a zero-shot summarization prompt: no example summaries, just the task. The model relies on its general grasp of summarizing.

Zero-shot shines for sentiment analysis too - just name the task and the model classifies it. Perfect for quick checks without specialized categories.

Zero-shot has limits: custom output formats, multi-step reasoning, ambiguous tasks, and consistency across queries can all trip it up. That's where few-shot helps.

Few-shot prompting includes a few input-output examples right in the prompt. They demonstrate the pattern, style, or format, boosting accuracy on tricky tasks.

A few-shot prompt is: instruction, a couple of input/output examples, then your new input. Keep examples clear, concise, and representative.

Want JSON output? Few-shot nails it - the examples below pin down the exact structure for sentiment and confidence, so the model copies the shape.

Few-shot is ideal for tone and style: hard to describe "quirky," easy to show. These examples teach the model a short, punchy headline voice.

Choosing: zero-shot for simple, common tasks and lower cost; few-shot for specific formats, niche knowledge, and when consistency really matters.

You need an LLM to extract specific data from customer reviews (e.g., product name, issue type, suggested improvement) and output it as a structured JSON object. Which prompting technique would be most effective?

Recap: zero-shot is quick and efficient for general tasks; few-shot uses examples to lift accuracy and consistency on complex ones. Match the tool to the job.