LLM Prompt Engineering ¹

While humans can interpret vague or incomplete information, LLMs require precise instructions to deliver optimal results.

It’s not simply about asking the model a question; it’s about phrasing, context, and understanding the model’s logic.

Natural Language Processing (NLP), by definition, is the field of artificial intelligence that focuses on the understanding and generation of human language by computers. It is employed in a wide range of applications, including voice assistants, machine translation, chatbots, and more. In other words, when we talk about NLP, we refer to computers’ ability to understand and generate natural langugage.

Generative pre-trained transformer (GPT) - when we talk about prompt engineering, we are typically referring to the manipulation of the initial sequence to guide the GPT’s output. This involves carefully crafting the input to elicit a specific type of response or to steer the model’s behavior in a desired direction. The better the prompt, the more accurate and contextually relevant the output from GPT tends to be.

GPT stands for Generative Pre-trained Transformer. This architecture (transformer) is specifically designed for processing ang generating natural language, making it proficient in understanding and producing human-like text.

Transformer architectures are particulary effective in processing sequences of data, such as text, and are known for their ability to capture long-term dependencies in the data.

Other models:

GPT-3: A set of models capable of understanding and generating natural language.
DALL-E: An image generation model that can create images from textual descriptions. It can generate high-quality images in a wide variety of styles, sizes, and resolutions.
Whisper: A model that can convert audio input into text with high accuracy.
Embeddings: A set of models capable of converting text into numerical representations.
Moderation: A fine-tuned model capable of detecting if text may be sensitive or unsafe.
Codex: A set of models capable of understanding and generating code, including translating natural language into code.

GPT models process text in chunks called tokens. Tokens represent frequently occurring character sequences and they are different from words. It is all about building statistical relationships between words and contexts.

An API have the option to adjust certain parameters, including temperature, when integrating language models like GPT into their applications. Understanding and adjusting API parameters offers multiple advantages.

Here is a list of the most common parameters: ¹

Temperature
- ^{image ilustrating temperature}
  
  ^{image from Medium ²}
Top-p
- ^{image ilustrating top-N and top-P}
  
  ^{image from Medium ²}
  - Top N/P Sampling ²
    - LLMs are next token generation models. This is done by selecting a token based on the probability distribution.
    - Top N sampling chooses a token from the top N highest probability tokens
    - Top P sampling chooses a token from the highest probability tokens whose probability sums up to P
Top-k
Sequence length (max_tokens)
Number of responses (n)
Presence penalty (presence_penalty)
Frequency penalty (frequency_penalty)
Best of (best_of)

Perplexity applied to a text, could be used to detect if it was generated by a language model. If the perplexity is very low, it is likely that the text was generated by a language model. Some AI content detection tools, like GPTZerp user this technique to detect if a text was generated by ChatGPT.

Betterprompt is a test suite for LLM prompts before pushing them to production. This open source tools is inspired by the same paper, “Demystifying Prompts in Language Models via Perplexity Estimation. Betterprompt helps in calculating the perplexity of a prompt.

The lower the perplexity, the better the prompt is.

Hack the prompt ¹

How can we lower the perplexity of a prompt?

What if we ask the model to create the prompt for us? This technique could be useful since the model is more familiar with the language it generates according to what we learned in the previous section.

Imagine you want to ask the model the following question:

Write a blog post about aging gracefully.

We can use the following prompt to ask the model to create a prompt for us:

Write a list of prompts that I will use to write a blog post about aging gracefully.

If you already have a list of prompts, you can ask the model to paraphrase them for you.

Paraphrase the following prompts:
1. What does aging gracefully mean to you personally?
2. Why is the topic of aging gracefully so important in today's society?

Use double quotes: if you want to emphasize certain words or phrases in your prompt, use double quotes. This could help the model understand the importance or specific focus of those words.

Use single quotes when needed: use single quotes to indicate a quote within a quote. This helps the model understand the context of the quote and generate a relevant response.

Use text separators: use text separators such as """ and === to separate different sections of your prompt. This helps the model understand the structure of your prompt and generate a relevant response.

How can you write better prompts when you’re feeling lazy? ³

The main idea behind meta-prompting is: you are better at assessing the quality of a prompt than you are at writing one.

Make the model write a better prompt for you. Here’s the compressed meta-prompt:

Act as an expert Prompt Engineer.
I'll give you a messy prompt.
Reason step by step to improve it.
Write the final prompt as an elegant template with clear sections.
Use lists, placeholders, and examples.
##
Prompt:"""<Insert your prompt here, and yes please, use the triple quotes.>"""

Prompt Engineering is a fancy way to say “Write better and better instructions for AI until it does exactly what you want”.

The idea of meta-prompting is to use that very skill to your advantage. You make your model ask itself a better question than the one you’d ask yourself.

And just like with people, you get better answers if you ask better.

To build the quick fix we’ll use three prompting techniques to create an efficient meta-prompt:

Role Prompting: you give a role to your LLM, which indirectly specifies the context, objective, and other parameters like the style.
Chain-of-Thought prompting (CoT): also known as “Reason step by step.” This is the most powerful sentence you can use when prompting an LLM. When LLMs “reason step by step,” they use tokens to “think” through stochastic predictions, which increases accuracy.
Placeholders: this is a way to both write and submit flexible prompts. Placeholders <like_this> allow you to play with different inputs and pick from a set of options.

How to Write Expert Prompts ⁴

Prompt Engineering is a fancy way to say, “Write better and better instructions for an AI model until it does exactly what you want”.

Compared to humans, AI systems require more details, precise explanations, clearer instructions, and a bit of repetition. You want to be intentional with each word you write — almost as if you were writing code.

The hottest new programming language is English.

^{image from Medium ⁴}

Why is Prompt Engineering harder than you think?
- The short answer is “an illusion of ease”.
  - Since we use natural language to write prompts, we don’t see it as a complex skill that requires practice. All you have to do is write instructions in plain English.
  - When you talk to humans, you prompt them the same way you’d prompt an AI. Both humans and AI use inner models to respond to your prompt.
  - The human model draws on cognitive abilities, past experiences, theoretical knowledge, and real-time sensory data. Language Models, on the other hand, rely solely on language patterns.

Each prompt has the following format: ⁴

[Title of the prompt]

Prompt content: Content of the prompt that sometimes includes <placeholders>.
<placeholders> are a tool you use to make better prompt templates.

(Comments that are not part of the prompt are written between parenthesis like this)

## (The two hashtags separate the same prompt into two or more sections)

--- (The three dashes separate examples inside the same prompt)

*** (The three stars separate two different prompts or two different outputs)

Prompt Engineering Topics

The Basics of Prompting ⁴

Each prompt is a bridge between what you want and what your Language Model generates. The shape of your bridge depends on the problem you want to solve, but the underlying structure remains the same.

^example

[Examples of a simple prompts]
    
Prompt example #1: Act as an independant analyst who specializes in <field_1> and <field_2>.
Search the internet for the pros and cons of soy milk.
    
Use 5 bullet points for pros and 5 bullet points for cons.
Browse at least 6 different sources and cite each one of them in your analysis.
    
##
    
<field_1>: Food industry
<field_2>: Nutrition
    
.***
    
Prompt example #2: Explain the five love languages as if you're talking to a 16-year-old.
    
##
    
Desired format:
Introduction.
- Love language #1: explanation and example.
- Love language #2: explanation and example.
- Love language #3: explanation and example.
- Love language #4: explanation and example.
- Love language #5: explanation and example.
Conclusion.

Specify the context (also called “Priming”)
Specify the desired format
Use <placeholders>
- Placeholders <like_this> help you achieve two separate goals.
  - Use <placeholders> to write flexible prompts that can take different inputs. You have to indicate the content of each placeholder in your prompt. In this case, a placeholder is a parameter.
  - Use empty <placeholders> to illustrate the desired format. Here you don’t have to write the content of each placeholder.
Specify the style/tone
- a few examples of styles you can pick from:
  - Generic styles: formal, informal, persuasive, conversational, sarcastic, dramatic, condescending, nuanced, biased, humorous, optimistic, pessimistic, etc.
  - Domain-specific styles: academic, legal, political, technical, medical, news, scientific, marketing, creative, instructional, etc.
  - Mimicking the style of a real person: Agatha Christie, Daniel Kahneman, J.K Rowling, James Baldwin, Hajime Isayama, etc.
- list of tones of voice
Specify the length of the desired response

Specify the target audience

^example

[Specify the target audience]
     
Prompt that targets general adult readers: Please explain the benefits of regular exercise in a way that is easy to understand for the general public.
     
.***
     
Prompt that targets medical professionals: Please write a scientific article that targets medical professionals. The article discusses the physiological and psychological benefits of regular exercise. Make sure the article responds to the expectations of an audience of medical professionals.

One common mistake people make when writing prompts is to consider “style” and “target audience” as the same parameter. In reality, the style determines how the text sounds and the target audience decides which words to use.

^example

[Specify the target audience]
    
Example #1: Explain to an audience of visual artists how Generative AI will impact their field of work.
    
.***
    
Example #2: Write a tweet that promotes an article about AI-driven disinformation. The tweet targets people interested in technology, communication, and social media.
    
.***
    
Example #3: Outline a fundraising campaign and kindly add actionable tips. The content will be sent to a group of non-profiters to help them improve their current methods.

Many-Examples Prompting
- Many-Examples prompting is particularly useful for tasks that involve imagination

Temperature Control

Temperature is a parameter that influences the “randomness” of the response generated by your language model. It typically ranges from 0 to 1, but in some instances, you can bring the temperature beyond 1.

^example

[Temperature control]
    
Example #1: At a temperature of 0.7, please explain why banana bread is called "bread" and not "cake" even though it tastes like a cake.
    
.***
    
Example #2: Write a Python script that transposes a 10x10 matrix. Please provide two versions of the code where the first is generated at a temperature of 0 and the second at a temperature of 0.4.
    
.***
    
Example #3: Act like an expert developer in <name_of_the_programming_language>. I will submit a few lines of code in the chat, and you'll review the code, then perform the following 7 tasks in the specified order defined below. When you write code, always use a temperature of <temperature_value>.
    
##
    
<name_of_the_programming_language> = Python.
<temperature_value> = 0.

Zero-Shot Prompting (no examples)
- Zero-shot prompting is to write an instruction for your AI model without providing context or examples.

Few-Shot Prompting (several high-quality examples)

^example

[Few-shot prompting]
    
Prompt:
    
Text #1: My favorite part of networking events is eating all the food while everybody is busy trying to impress each other.
    
Tone #1: Playful.
    
##
    
Text #2: She stormed out of the conference room, slamming the door behind 42 staff members who instantly turned their eyes away as if ashamed of their lack of courage.
    
Tone #2: Angry.
    
##
    
Text #3: Do you think they noticed the missing "export" feature in today's demo? I'm pretty sure Nabil whispered something into the client's ear. I really don't like that bald dude!
Tone #3: Anxious.

a prompt that takes advantage of using tone
- list of tones of voice

Zero-Shot/Few-Shot — The simple version
In-context Learning vs. Chat History
- In-context learning is a prompting technique that allows you to steer the responses of your LLMs in a specific direction. All you need are a few examples, just like few-shot prompting.
- In-context learning vs chat history
  - Use in-context learning to fine-tune a “raw” model like GPT-4, OpenLLaMa, or Falcon. In other words, you can create a customized chatbot but the process can be tedious.
  - Use chat history to leverage “memory” and “long conversations.” It’s easier to customize your output but the quality may go down over time.
Chain of Thought Prompting
- Chain of Thought (CoT) prompting means you tell your Language Model to reason step by step before arriving at a final response. It’s as if you ask your model to think out loud.
- ^example
  1. 4x3 = 4+4+4 2. 4+4+4 = (4+4) + 4 3. (4+4) + 4 = 8+4 4. 8+4 = 12
- CoT prompts are typically used to solve logical riddles. The idea is to break down complex problems into smaller, more manageable questions.
Zero-Shot Chain of Thought
Role Prompting
- Assigning a specific role to your Language Model helps it capture more and better semantic relationships (ie: logic and meaning).
- In a way, Role Prompting helps you nudge your model to focus on specific information inside its training data. It’s a shortcut to specify many variables at once — like context, style, perspective, and vocabulary.
Knowledge Generation Prompting

Knowledge Integration

Here’s a prompt you can use to summarize a document.

^example

[Prompt to generate summaries using Microsoft Edge's Chat feature]
    
Prompt: Summarize this paper. Start the summary with the title of the paper and its URL. Then list the main ideas in bullet points. Please illustrate the main ideas with examples extracted from the paper.

Directional Stimulus (DS) Prompting
Recursive Criticism and Improvement (RCI) Prompting
Self-Refinement Prompting
- Self-Refinement Prompting Self-refinement is a variant of RCI prompting.
Reverse Prompt Engineering
Prompt Revision
Program Simulation Prompting

LLM Prompt Engineering 1

Hack the prompt 1

How can you write better prompts when you’re feeling lazy? 3

How to Write Expert Prompts 4

LLM Prompt Engineering ¹

Hack the prompt ¹

How can you write better prompts when you’re feeling lazy? ³

How to Write Expert Prompts ⁴