Ask a model "write about dogs" and you get three generic paragraphs nobody asked for. Ask the same model to act as a veterinary blog editor, give it a target audience of first-time puppy owners, a word count, and a required structure, and you get something you could actually publish. Same model, same underlying "knowledge" of dogs. The only thing that changed is the prompt - and that gap is the entire reason prompt engineering exists as a skill.
The short version
Prompt engineering is the practice of designing the input to a language model - the wording, structure, examples, and context you give it - so that the output reliably matches what you actually need. It sits between "just ask the model something" and building custom software: no code, no fine-tuning, no retraining, just a deliberate, testable way of writing the instruction itself.
The model doesn't change between these two runs. The prompt does - and that's the whole discipline in one picture.
Why this is a real skill, not just "being clever"
It's tempting to write prompting off as trial and error, but the reason it works as a discipline is that models are extremely sensitive to a small set of specific, learnable inputs: the role you assign them, the context you provide, the examples you show, and the format you demand. Change any one of those deliberately and you can predict, in general terms, how the output will shift - more focused, more structured, more accurate, less likely to hallucinate. That predictability is what separates engineering from guessing.
What "engineering" actually means here
It doesn't mean writing code. It means treating your prompt the way an engineer treats a design: you have a goal, you have levers that affect the outcome, and you iterate - write a version, test it against real inputs, look at where it fails, adjust one lever, test again. The guides in this series are organized around exactly those levers: the core techniques that do most of the work (zero-shot, few-shot, chain-of-thought), the structural pieces that make a prompt reliable (role, context, format), and the patterns you reach for once a single prompt isn't enough.
Why it matters more, not less, as models get better
A common assumption is that as models improve, prompting will stop mattering - the model will "just understand" what you want. In practice the opposite has held so far: better models are more responsive to good prompts, not less, because they're better at following precise instructions, respecting format constraints, and using the context you give them. A vague prompt to a strong model still gets you a vague answer, just a more fluent one.
Where this fits with everything else on this site
If you've been through the RAG 101 or MCP 101 series here, you've already seen how much context and tool access shape what a model can do. Prompt engineering is the other half of that picture - it's how you shape what the model does with whatever context and tools it already has. The two skills compound: a well-engineered prompt on top of good retrieval or a well-chosen tool is where the reliable results actually come from.