What is Prompt Engineering ?

Prompt engineering is the practice of crafting the exact instructions and context given to a language model so its answers match what a user actually needs. The core insight is simple: vague requests produce vague results, while carefully specified prompts—paired with the right background information—steer the model toward more relevant, accurate, and usable outputs. That matters because language models can otherwise “fill in the blanks” using general knowledge rather than the information a task depends on.

The transcript first grounds the idea of a “prompt” in a real-world analogy: asking a knowledgeable friend in a coffee shop for “the best way to grow tomatoes” is the prompt, and the friend’s tailored advice is the model’s output. In the same way, the clarity and structure of a question strongly influence the quality of the response. From there, prompt engineering is framed as the difference between telling a librarian “please provide me a book to read” versus giving a more detailed request like “suggest motivational books from the latest bestsellers.” The second request narrows interpretation and makes the recommendation more likely to fit preferences.

Next comes the breakdown of prompt types. A system prompt (also called an instruction prompt) acts like a rulebook for the model: it sets the role, tone, and boundaries for how the assistant should behave. A user prompt is the actual request coming from the person interacting with the system. The transcript emphasizes that system prompts can include predefined instructions and context, such as how to greet users, what domain to operate in, and what to do when information is missing.

A key example shows how system prompts can force the model to use provided context rather than relying on its own memory. With a context snippet like “Patty loves animals… has two cats and one dog,” a question about how many cats Patty has should be answered strictly from that context. This distinction becomes central to building reliable applications.

The practical workflow is described as iterative. Requirements get translated into prompts (system/instruction prompt plus user prompt and any additional context). The model generates results, those results are checked against the original requirement, and then prompts or context are modified until the output is “good enough.”

To demonstrate, the transcript builds a digital assistant for a library. The system prompt instructs the assistant to greet politely, help users find books, and—if a requested book isn’t available—collect the user’s phone number and promise a callback. The additional context is a small library database containing book names, authors, and shelf locations (book self numbers). When a user asks for motivational books, the assistant returns titles from that database; when asked where to find a specific book, it responds with the shelf location (e.g., “book self D1”), matching the provided context.

Finally, the implementation is shown in a Google Colab-hosted Jupyter notebook using OpenAI’s API with the GPT 3.5 turbo model and a low temperature (0.2). A helper function sends a message list containing the system prompt and the evolving user input to the model, and a loop keeps accepting user queries until the assistant’s responses meet the library-finding task.