Glossary
What is prompt engineering
In short
Prompt engineering is the work on the task text for an AI model: you write it, test it on examples and edit it until the answers are consistently accurate.
Prompts · prompt engineering
MMOOZZGG.com editorsChecked: 1 min read
For a child
A teacher wants the whole class to understand the homework. She explains it in some words, sees that half the children got it wrong, and explains it differently: she adds an example and draws a diagram on the board. After a couple of tries, almost everyone does it right. People do the same with a neural network: they choose the words of the task until it understands the task as intended.
In plain words
The same task, written in different ways, gets different answers. Prompt engineering turns choosing the wording from guesswork into a method. First you decide which answer counts as good. Then you write a first version of the task, run it on a dozen typical cases and see where the model made mistakes. Then you edit: make the role and format clear, add samples, split a hard task into steps, and ask the model to reason first and answer after. This is how a reply template for complaints reaches a form an employee can send without rewriting.
For business
For a business, this is the first and cheap way to raise the quality of AI work: the model stays the same, and only the instructions change. Sometimes this is enough to solve the task without fine-tuning or new infrastructure. The costs are an engineer’s hours and model runs on the test set. The risk is that the wording gets tuned to five lucky cases, and quality drops on the real flow. So collect a set of real requests in advance, fix the metric and check every edit against it. Repeat the check when you change the model or its version. The metric usually goes into the pilot plan before the first line of the instruction is written.
For a specialist
Prompt engineering is iterative design of the model’s input against a set quality metric. The Anthropic guide starts with three conditions: criteria of success, a way to test them empirically and a first draft of the task. The basic techniques are a clear instruction and a role (role prompting), examples inside the request (few-shot), tags that mark the parts of the input, long documents placed before the question, step-by-step reasoning (chain-of-thought), prompt chaining and a required output format. A systematic survey by Schulhoff and co-authors (2024) offers a vocabulary of 33 terms and a taxonomy of 58 techniques for text models. Automatic approaches tune the wording by search or with the help of another model (meta-prompting). The result is evaluated on a held-out set, not on the examples used for editing; otherwise the instruction overfits to them.
Example
A support team gets answers from the model that are too long and full of extra caveats. The team rewrites the task: it sets the tone, a limit of five sentences and three samples. On a set of one hundred past requests, the share of answers that need no edits grows, and the template goes into use.
Related terms
Questions and answers
Sources
Where companies use it
Not sure whether your company needs this?
AI solutions map