Prompt engineering
The words you choose are the most visible lever on AI output, and often the least powerful one, so use them well.
The lever everyone reaches for first
Prompt engineering is the practice of shaping your request itself: the phrasing, the examples you give, the constraints you state, the format you ask for. It is the first thing most people try when an AI output disappoints them, and for good reason, it is the only lever you touch directly, in the same box where you type.
It is also, on its own, the weakest of the three levers. A model can only work with what you tell it and what it can see. Rewriting your prompt fixes an ambiguous or underspecified request; it cannot fix a model that is missing the right file, and it cannot add a verification step that catches a wrong answer. Prompt engineering is necessary, not sufficient.
Where it does matter is a lot: most everyday requests fail not because the model lacks information, but because the request itself is vague. "Make this better" and "write something for our customers" leave the model guessing at what better means and who the customers are. A prompt with a clear role, task, constraints, format, and examples removes the guessing.
Select a numbered hotspot to see what occupies that part of the diagram.
Same request, "write a marketing email", improved by adding one part of the prompt at a time.
Vague: no role, no constraints
Original request: "Write a marketing email about our new project templates feature." Result: three generic paragraphs, an exclamation point in every sentence, a subject line like "Exciting News!", no sense of who is sending it or who is reading it. Nothing is wrong exactly, but nothing is usable without a rewrite.
Add role and task
Improved: "You are a product marketer at a B2B software company. Write a launch email announcing our new project templates feature to existing customers." What changed: the model now has a professional frame (B2B, existing customers, not cold prospects) instead of writing generic ad copy. What improved: the tone settled down, the email assumes the reader already knows the product, and it stopped over-selling basics they already know.
Add constraints and format
Improved further: "...Keep it under 120 words. One clear call to action: try the feature this week. No exclamation points, no 'exciting news' framing. Return a subject line, then the email body as plain paragraphs, no bullet list." What changed: length, tone limits, and a single desired action are now explicit, and the output shape is named instead of left to the model's default. What improved: the draft arrived at the length and structure needed for an actual send, not a first draft that still needed trimming and restructuring.
Add an example
Final addition: "Match this tone: 'Your dashboard just got faster. No setup, no new tab, it is already there.' Direct, short sentences, no fluff." What changed: instead of describing the tone in the abstract ("direct", "no fluff"), the prompt shows a sentence in that tone. What improved: the model's draft matched the target voice on the first attempt, the kind of close, specific match that takes several rounds of description to reach without an example.
Question 1 of 4
Why is prompt engineering described as necessary but not sufficient?