What is Prompt engineering?

Start with success criteria

Anthropic’s guide to prompt engineering says to begin with a clear definition of success for your task, a way to test against it, and a first draft prompt to improve (Claude documentation). Without a test, you cannot tell whether a change made the prompt better or worse.

Core techniques

1. Be clear and direct about the task and the audience.
2. Give examples of the output you want.
3. Separate instructions from the material with labels or tags.
4. Give the model a role, such as "You are an accounts assistant".
5. Ask it to think through the steps before answering hard problems.
6. Break a big job into a chain of smaller prompts.

Where it is used

In a business, prompt engineering sits behind customer reply templates, product listing generators, report summaries and AI agents. A prompt that works 9 times out of 10 in a chat window may fail when it runs 500 times a day inside an automation. Engineering the prompt is what closes that gap.