Prompt Engineering
Francesco Garavaglia
Digital & Innovation Strategist | Enterprise Architect | IT Modernization Expert | AI & Automation Expert | Podcaster
For over twenty years, I've been navigating this fascinating universe, sharing my discoveries and reflections through this newsletter. Today, we embark on a fascinating journey into the art and science of Prompt Engineering. You might think interacting with AI is as simple as asking a question, and while that's true to some extent, truly unlocking the potential of large language models (LLMs) requires a more nuanced approach. Get ready to discover how the way we phrase our requests can dramatically influence the intelligence and accuracy of AI responses.
Why dedicate an entire edition to Prompt Engineering?
Because in this era of increasingly sophisticated AI, the ability to communicate effectively with these digital brains is becoming a critical skill – not just for developers, but for anyone who wants to leverage their power. We've all experienced the frustration of asking an AI a question and receiving a vague or unhelpful answer. Often, the issue isn't the AI's lack of knowledge, but rather our inability to guide it effectively. Think of it like giving instructions to a highly capable but somewhat literal assistant: the clearer and more precise your instructions, the better the outcome.
This edition will demystify the complexities of prompt engineering, showing you practical techniques and best practices to transform your interactions with AI from frustrating to fruitful. We'll explore how carefully crafted prompts can unlock a wealth of possibilities, from summarizing complex documents to generating creative content and even debugging code. So, whether you're a tech enthusiast or simply curious about how to make AI work better for you, this exploration into the world of prompt engineering is for you.
At its core, prompt engineering is about designing high-quality prompts that act as a roadmap for LLMs, guiding them to generate the specific output you desire. Remember, LLMs operate by predicting the next word (or token) in a sequence. Our prompts essentially set the stage, influencing the model's predictions. Prompt engineering involves tinkering with prompt length, writing style, structure, and context in relation to the task. Prompts can be used for a wide range of tasks, including text summarization, information extraction, question answering, text classification, language or code translation, code generation, and reasoning.
Controlling LLM Output: The Settings Behind the Scenes
Beyond the prompt itself, you need to consider the LLM output configurations. These settings control the model's output and should be set optimally for your specific task.
Output Length: This determines the number of tokens generated. Generating more tokens requires more computation, potentially leading to slower response times and higher costs. Reducing the token limit doesn't make the output more succinct; it just stops generation once the limit is reached.
Sampling Controls: LLMs predict probabilities for the next token. Sampling settings determine how these probabilities are processed to choose the output token
These settings about sampling interact, and inappropriate settings can contribute to the "repetition loop bug". The main sampling controls are:
Temperature: Controls randomness. A lower temperature (e.g., 0) results in more deterministic output, while higher temperatures lead to more diverse or unexpected results.
Top-K: Selects the top K most likely tokens. A lower Top-K (e.g., 1) results in more restrictive and factual output, while a higher Top-K allows for more creative and varied output.
Top-P (Nucleus Sampling): Selects tokens whose cumulative probability does not exceed a certain value P.
Exploring Prompting Techniques: Your AI Toolbox
LLMs are trained to follow instructions, but specific techniques can significantly improve results.
General Prompting / Zero-Shot: The simplest technique, providing only a task description and input text with no examples.
One-Shot & Few-Shot: Provide examples to help the model understand the desired output structure or pattern. One-shot provides a single example, while Few-shot provides multiple examples to demonstrate a pattern.
System, Contextual, and Role Prompting: These guide LLMs by focusing on different aspects: System Prompting: Sets the overall context and purpose, defining the "big picture" and fundamental capabilities. Useful for specifying output format (e.g., uppercase, JSON) and controlling safety/toxicity. Contextual Prompting: Provides immediate, task-specific details or background information. Role Prompting: Assigns a specific character or identity (e.g., travel guide, teacher).
Step-back Prompting: Improves performance by prompting the LLM to first answer a general question related to the task, and then using that general answer as context for the specific task prompt.
Chain of Thought (CoT): Improves reasoning by generating intermediate reasoning steps. This technique works with off-the-shelf LLMs and provides interpretability.
Self-consistency: Combines sampling and majority voting with CoT. It generates diverse reasoning paths by prompting the LLM multiple times with a high temperature setting, and the most common answer is chosen.
Tree of Thoughts (ToT): Generalizes CoT by allowing LLMs to explore multiple different reasoning paths simultaneously in a tree structure.
ReAct (Reason & Act): Enables LLMs to combine natural language reasoning with external tools (like search or code interpreters).
Automatic Prompt Engineering (APE): Automates the process of writing prompts. A model generates candidate prompts, which are then evaluated, and the best one is selected.
Prompting for Code: AI as Your Coding Assistant
LLMs can also be effective when prompting for code.
Writing Code: LLMs can help speed up writing code snippets in various languages. Always read and test the generated code.
Explaining Code: LLMs can help understand existing code by providing explanations.
Translating Code: Code can be translated from one language to another.
Debugging and Reviewing Code: LLMs can help find and fix errors.
Best Practices for Prompt Engineering: Your Guide to Success
Effective prompt engineering requires tinkering and experimentation.
Provide Examples: One-shot or few-shot examples act as a powerful teaching tool.
Design with Simplicity: Prompts should be concise, clear, and easy to understand.
Be Specific about the Output: Guide the LLM with details rather than being generic.
Use Instructions over Constraints: Focus on positive instructions (what the model should do).
Control Max Token Length: Manage output length via configuration or explicitly in the prompt.
Use Variables in Prompts: Makes prompts reusable and dynamic.
Experiment with Input Formats and Writing Styles: Different prompt formulations and styles yield different outputs.
Document the Various Prompt Attempts: Crucially, document attempts in detail to track progress and learn from your experiments.
My 2 Cents
Prompt engineering is not just a technical skill; it's a new form of communication. Mastering it allows us to harness the immense power of AI in more effective and nuanced ways. Think of it as learning to speak the language of AI. The more precisely you can articulate your needs, the more effectively the AI can assist you. This iterative process of crafting, testing, analyzing, and refining your prompts will ultimately transform your interactions with AI, making it a truly powerful ally in your daily tasks.
Just remember:
Clarity is King: Clear instructions lead to better AI responses.
Examples are Powerful: Provide examples to guide the model effectively.
Experiment & Document: Continuous experimentation and meticulous documentation are crucial for improvement.
AI is a Tool: Understand its capabilities and limitations to leverage it fully.
If you want to delve deeper into this fascinating topic, please download the episode from my podcast: MODA - Modern Digital Architecture. Follow my podcast also on Spotify and YouTube now! For my international listeners, there's also an AI-Generated English version available on Spotify!
And if you appreciate this content, please consider:
📝 Join my Telegram group: https://t.me/+ChiNfo8oJt5iNjQ0
☕ Buying me a virtual coffee: https://buymeacoffee.com/fgaravaglia
📣using my Affiliate like to Gamma.APP: https://try.gamma.app/vtote9zj9uxu
🛒 Buying any of my Digital Products: https://francescogaravaglia.gumroad.com/
My Products
Mastering AI framework
This E-Book was born from the direct experience of those who, approaching the world of artificial intelligence, found themselves faced with a mountain of confusing and disorganized information. The lack of a clear and accessible guide made learning a difficult task. Precisely to overcome these difficulties common to many, the AI Mastery Course was created, with the aim of providing a practical and accessible guide to help anyone go from beginner to expert in using artificial intelligence tools. The world of AI offers immense opportunities to improve productivity, innovation and creativity, but many feel excluded due to its apparent complexity. This E-Book responds to the need to make these tools understandable and usable by everyone.
AI Assistant
I have posted several prompts to create virtual assistants for recurring tasks like writing emails, designing presentations, deep research etc. Each of them is described in detail to explain why you should use these assistants and how the prompt was built.