Prompt Engineering: Basics

Andrew Ellis

28 November, 2024

Understand LLMs through Hands-On Experience

• Dedicate time to actively using large language models (LLMs).
• Utilize LLMs for tasks related to your work or personal interests.
• Explore their abilities by posing diverse and unique prompts.
• Observe where LLMs work well and where they don’t.

A useful metaphor

Imagine you are giving instructions to a junior intern or assistant.

Zero-Shot Prompting

Definition: Asking the model to perform a task without providing examples.

Example Prompt

Translate the following English text to French: ’Hello, how are you?

Role assignment

Technique: Define a specific role for the AI to adopt.

Example Prompt

You are an expert historian. Explain the causes of World War I.

Example Prompt

You are an 8-year-old child. Explain the causes of World War I.

Example Prompt

You are an experienced emergency room nurse with over 15 years of experience in patient triage. Your role is to perform initial assessments of patients based on their reported symptoms and medical history. You have a calm demeanor and the ability to quickly prioritize cases based on severity. In this role, you will categorize patients’ conditions and recommend appropriate next steps.

Clear communication

Technique: Use precise language and specific instructions

Example Prompt

You are an ER nurse with 15+ years of triage experience.

Your tasks: - Assess patients quickly based on symptoms and medical - Categorize conditions by severity - Recommend next steps for treatment

You are calm under pressure and efficient in prioritizing cases.

Example Prompt

List five benefits of regular exercise, each in a separate bullet point.

Emotion prompting

Technique: Incorporate emotional language to potentially improve accuracy and response quality¹.

Example Prompt

I’m really excited to learn about this! Can you enthusiastically explain how photosynthesis works?

Example Prompt

Rate this essay according to [these criteria]. You will receive a bonus if you do a good job.

1. Emotion prompting’s effectiveness in improving language model responses is debated. Some argue it could enhance naturalness, while others suggest minimal or inconsistent impact across tasks and models. More research is needed.

Provide context

Technique: Give relevant background information

Example Prompt

Context: You are working in a busy urban hospital emergency room during flu season. It’s currently 2 AM on a Saturday, and the waiting room is full. The hospital has been dealing with a recent outbreak of a new strain of influenza in the community.

Patient Information:

• 45-year-old male
• No known pre-existing conditions
• Not on any regular medications
• Last flu shot was 2 years ago
• Given this context and patient information, assess the following reported symptoms: [Insert symptoms here]

Use examples

This is known as few-shot prompting.

Technique: Illustrate desired output with examples.

Example Prompt

Perform a triage assessment based on the patient’s symptoms. Format your response similar to the following examples:

Example 1: Symptoms: Chest pain, shortness of breath, left arm numbness Assessment: Emergency Reason: Symptoms strongly indicate a possible heart attack Action: Immediate medical attention required. Call for a cardiac team.

Example 2: Symptoms: Mild fever, sore throat, fatigue Assessment: Non-urgent Reason: Symptoms suggest a common cold or mild flu Action: Rest at home, monitor symptoms, seek medical attention if condition worsens.

Assess the following patient:

Symptoms: [Insert patient’s symptoms here]

Specify Output Format

For example:

• a Markdown table
• Word or Excel document
• CSV file
• Structured list

Example Prompt

Based on the patient’s symptoms, create a triage assessment. Present your findings in a Markdown table with the following columns:

• Severity
• Primary Concern
• Symptoms
• Recommended Action

Use one of three severity levels: Emergency, Urgent, or Non-urgent.

The patient has the following symptoms: [Insert symptoms here]

Structure Input Using Markdown

Technique: Organize input information in a structured format using Markdown.

Example Prompt

Based on the patient's symptoms, create a triage assessment.

# Patient Triage Information

## Patient Details

- **Name**: John Doe
- **Age**: 45
- **Gender**: Male
- **Medical History**: No known pre-existing conditions
- 
## Current Symptoms
1. Chest pain (severity: 8/10)
2. Shortness of breath
3. Left arm numbness

Ask an LLM

LLMs have been trained on a lot of data, including prompting techniques¹.

Example Prompt

As a language model, how would you proceed when given the following prompt: ““” You are an ER nurse with 15+ years of triage experience.

Your tasks: 1. Assess patients quickly based on symptoms and medical history 2. Categorize conditions by severity 3. Recommend next steps for treatment 4. You are calm under pressure and efficient in prioritizing cases. ““”

Example Prompt

How would you improve this prompt?

1. It is debated whether LLMs are particularly suited to writing prompting techniques.

Generate Python code

Technique: ask an LLM to generate Python code, or in the case of ChatGPT to “use Python”

Example Prompt

[Insert query here…]

Use Python.

References