AWS Basic Prompt Techniques

Basic Prompt Techniques

When crafting and manipulating prompts, you can use specific techniques to achieve the response you want from AI models.

In this lesson, you learn how to use various prompt engineering techniques to help you use generative AI applications effectively for your unique business objectives.

Zero-shot prompting

Zero-shot prompting is a technique where you present a task to an LLM without giving the model further examples.

With zero-shot prompting, you expect the model to perform the task without a prior understanding, or shot, of the task.

Modern LLMs demonstrate remarkable zero-shot performance.

Image created by Amazon Web Services.

The following tips can help you use zero-shot prompting:

• Larger LLMs are more likely to produce effective results with zero-shot prompts.
• Instruction tuning can improve zero-shot learning. You can use reinforcement learning from human feedback (RLHF) to scale instruction tuning, aligning modern LLMs to better fit human preferences.

Zero-shot prompt example

Note: This prompt did not provide any examples to the model. However, the model was still effective in deciphering the task.

Few-shot prompting

Few-shot prompting is a technique where you provide the model with contextual information about requested tasks. You include examples of both the task and the output you want, helping the model follow your guidance closely.

Image created by Amazon Web Services.

When using few-shot prompting, consider:

• Labels in a few-shot prompt don't need to be correct to improve performance. Random labels often perform better than no labels at all. However, the label space and distribution of input text are important.
• If you have many examples, use techniques to comply with token limits and dynamically populate prompt templates. An example selector based on semantic similarity can help.

Few-shot prompt example

Note: The Amazon Titan Text model was used in this example.

Chain-of-thought prompting

Chain-of-thought (CoT) prompting breaks down complex reasoning tasks through intermediary reasoning steps. You can use both zero-shot and few-shot techniques with CoT prompts.

Image created by Amazon Web Services.

CoT prompts are specific to a problem type. Use the phrase "Think step by step" to invoke CoT reasoning.

Tip: Use CoT prompting when the task involves several steps or requires a series of reasoning.

CoT zero-shot prompt example

Note: This prompt did not provide examples to the model, yet the model was still effective in deciphering the task.

CoT few-shot prompt example

CoT prompting becomes more powerful when combined with few-shot prompting:

This prompt provides both few-shot context (a question-and-answer example) and CoT prompting ("Think step by step").