Multi-step prompting

Prompt Engineering with the OpenAI API

Fouad Trad

Machine Learning Engineer

Multi-step prompting

Break down an end goal into series of steps
Model goes through each step to give final output
Multi-step prompts are used for:
- Sequential tasks
- Cognitive tasks

Diagram showing that a multi-step prompt is a prompt that gives the model a series of sequential steps to perform.

Multi-step prompts as treasure maps

Image representing a treasure map along with some clues to find the treasure.

Single-step prompt: writing a blog

prompt = "Compose a travel blog"
print(get_response(prompt))

Title: Exploring the Enchanting Landscapes of Iceland

Introduction: Welcome to my travel blog! Today, I am thrilled to share my unforgettable journey through 
the mesmerizing landscapes of Iceland. 

Day 1: Reykjavik - The Charming Capital [...]
Day 2: Golden Circle - Nature's Wonders [...]
Day 3: South Coast - A Journey of Ice and Fire [...]
Day 4: Glacier Lagoon - A Frozen Wonderland [...]
Day 5: Blue Lagoon - A Relaxing Finale [...]

Multi-step prompt: writing a blog post

prompt = """Compose a travel blog as follows:
Step 1: Introduce the destination.
Step 2: Share personal adventures during the trip.
Step 3: Summarize the journey.
"""
print(get_response(prompt))

Writing a travel blog post

Title: Exploring the Enchanting Streets of Barcelona


Step 1: Introduce the destination.
Welcome to Barcelona, a vibrant city nestled along the stunning Mediterranean coast of Spain [...]


Step 2: Share personal adventures during the trip.
Exploring the narrow, winding streets of the Gothic Quarter, I stumbled upon hidden gems at every turn. 
[...]


Step 3: Summarize the journey.
As my journey through Barcelona came to an end, I couldn't help but feel grateful for the incredible 
experiences and memories I had made [...]

Analyzing solution correctness

Checking solution correctness requires multiple steps
Example:
- Python code for calculation functions

Image of a calculator

Analyzing solution correctness

Typical solution to check

calculator = """
def add(a, b):
    return a + b
def subtract(a, b):
    return a - b
def multiply(a, b):
    return a * b
def divide(a, b):
    return a / b
"""

Single-Step Prompt:

prompt = f"""Determine if the code delimited 
by triple backticks is correct or not.
Answer by yes or no.
```{calculator}```"""

print(get_response(prompt))

Yes

Multi-step prompting to analyze solution correctness

Multi-Step Prompt:

prompt = f"""Determine the correctness of the code delimited by triple backticks 
as follows: 
Step 1: Check the code correctness in each function. 
Step 2: Verify if the divide function handles the case when dividing by 0. 
Code: ```{calculator}```"""

print(get_response(prompt))

The code appears to be correct. It defines four functions: add, subtract, multiply, and 
divide. Each function performs the corresponding mathematical operation and returns the
result. However, it does not handle the case when dividing by 0, which can result in a 
ZeroDivisionError.

Multi-step versus few-shot prompt

Steps

Explicitly tell model what to do

Diagram showing how a multi-step prompt breaks down the instruction into a series of steps in the input for the model to go over them one by one.

Multi-step versus few-shot prompt

Steps

Explicitly tell model what to do

Diagram showing how a multi-step prompt breaks down the instruction into a series of steps in the input for the model to go over them one by one.

Shots

Questions and answers model learns from

Image showing how a few-shot prompt works, where the input to the model is a set of example questions and answers along with the question we want the model to answer.

Let's practice!

Prompt Engineering with the OpenAI API