What Prompt Engineering Actually Is
Prompting is interface design for a probabilistic machine — you arrange context so the output you want becomes the most likely continuation.
Strip away the mystique and prompt engineering is this: a language model predicts the most probable continuation of the text you feed it, and you control every character of that text. Prompt engineering is the craft of arranging those characters — instructions, context, examples, formatting — so the continuation you want becomes the continuation the model finds most likely. You are not issuing commands to an obedient computer. You are conditioning a probability distribution. That reframe explains almost everything else in this course: why examples outperform exhortations, why order matters, why one ambiguous word can swing an output, and why 'magic words' mostly aren't.
It's fair to ask whether this still matters in 2026. Claude, GPT, and Gemini absorb sloppy prompts far more gracefully than the models of 2023, and for a one-off chat question, casual phrasing is usually fine. The discipline earns its name when a prompt runs ten thousand times a day against inputs you never previewed. At that scale, the gap between a mediocre prompt and an engineered one shows up as consistency, edge-case behavior, token cost, and support tickets. Prompting didn't die as models improved; it moved from party trick to production interface — the thinnest, highest-leverage layer of every AI feature.
What separates engineering from typing is iteration against evidence. An engineer writes a prompt, runs it on realistic inputs, studies the failures, changes one thing, and runs it again. Prompts get versioned like code, reviewed like code, and tested like code — because a one-word edit can measurably shift behavior. The person who tried their prompt once in a chat window and shipped it is not doing a weaker version of this; they're doing a different activity. This course gives you the toolkit for the real thing: structure, examples, reasoning scaffolds, reliability patterns, and the testing habits that hold it all together.
Equally important is knowing what a prompt cannot fix. No phrasing summons knowledge that isn't in the model's training data or context window — that's a retrieval problem. No persona makes a model reliably good at a task beyond its capability — that's a model-selection problem. No instruction fully prevents hallucination — that's a verification problem, and Module 3 tackles it head-on. Great prompt engineers are defined as much by their diagnosis — knowing when the fix is more context, a tool, or a different model entirely — as by the prompts they write.