Skip to main content
Generative AI

Prompt Engineering Basics

By March 11, 2024July 11th, 2024No Comments

In recent years, the rise of large language models (LLMs) like ChatGPT and Gemini have made interacting with AI through text prompts a commonplace activity. Millions of users have become adept at crafting queries to get information, generate content, or solve problems using these AI assistants. However, while many people know how to prompt LLMs for basic tasks, prompt engineering is a more nuanced and sophisticated skill set. Prompt engineering goes beyond simple querying, involving techniques to optimize LLM outputs, handle complex tasks, and even extend the model’s capabilities. It requires a deeper understanding of how LLMs process information and generate responses, as well as knowledge of advanced techniques like few-shot learning, chain-of-thought prompting, and context manipulation. While casual users might know how to ask an LLM for a recipe or to explain a concept, prompt engineers can coax these models into performing specialized tasks, improving accuracy, and even overcoming some of their inherent limitations.

LLM Functionality

LLMs are designed to do a few things well,


  1. Text Generation
    • Creates coherent text from prompts
    • Used for: writing, content creation, chatbots
  2. Summarization
    • Condenses long texts, retaining key points
    • Used for: news digests, article abstracts
  3. Classification
    • Categorizes text into predefined classes
    • Used for: sentiment analysis, spam detection
  4. Document Understanding
    • Extracts meaning and key info from documents


Fundamentals of Prompting

At its core, prompt engineering is about crafting clear and effective inputs to elicit desired outputs from LLMs. The basics involve:

Clear Queries: Start with straightforward, unambiguous requests.
Specific Instructions: Provide detailed guidelines for the desired output format, tone, or style.

For instance, instead of a vague prompt like “Tell me about AI,” you might use “Explain the concept of artificial intelligence in three paragraphs, focusing on its current applications in creative.”

Advanced Techniques: In-Context Learning

As we delve deeper, we encounter more sophisticated approaches:

1. Zero-shot Prompting

This technique involves asking the LLM to perform a task without any examples. It’s useful for straightforward queries where the model’s base knowledge is sufficient.

Example: “Translate the word ‘hello’ into French.”

2. Few-shot Prompting

Here, we provide the LLM with a few examples before asking it to perform a similar task. This method is particularly effective when you need consistent formatting or style.


English: Hello
Spanish: Hola
English: Goodbye
Spanish: Adiós
English: Good morning
Spanish: [LLM should fill this in]

3. Chain-of-Thought Prompting

This approach involves breaking down complex problems into steps, encouraging the LLM to show its reasoning process.

Example: “Let’s solve this math problem step-by-step: If a train travels 120 miles in 2 hours, what is its average speed?”

Advanced Strategies for Optimization

Temperature Adjustment

The ‘temperature’ setting in LLMs controls the randomness of outputs:

  • Low temperature (close to 0): More deterministic, focused responses
  • High temperature: More diverse and creative outputs

This feature allows you to fine-tune the balance between consistency and creativity based on your specific needs.

Retrieval Augmented Generation (RAG)

RAG  ( or REALM for Google ) enhances LLM capabilities by incorporating external knowledge:

  1. Query an external database
  2. Retrieve relevant information
  3. Incorporate this information into the prompt

This technique is particularly useful for domain-specific applications or when working with proprietary information. while this abilty is not built into the LLMS  per say many of the big names provide mechaisms to implement this type of solution.


For more specialized applications, fine-tuning involves retraining the LLM on a custom dataset. This process can:

  • Improve performance on specific tasks
  • Add domain-specific knowledge
  • Implement custom behavior or safety controls

However, it requires significant data and computational resources. however LoRA fine-tunes large language models by adding small, trainable matrices to existing weights, offering efficiency in memory, speed, and storage. It’s ideal for adapting big models to specific tasks without the computational demands of full fine-tuning.


Prompt engineering ranges from simple to complex techniques. Basic methods include clear queries and specific instructions, while advanced approaches involve zero-shot, few-shot, and chain-of-thought prompting. More sophisticated techniques include temperature adjustment for controlling response creativity, Retrieval Augmented Generation (RAG) for incorporating external knowledge, and fine-tuning methods like LoRA for task-specific model adaptation.


Leave a Reply