Prompt engineering is the practice of writing instructions for AI models that produce consistent, useful outputs in production, not just once in a test. It's part writing, part debugging, part documentation.
It's not purely an engineering job. Product managers are often good at it because they're used to writing requirements that others can act on without ambiguity. Same skill, different audience.
What it involves
A good prompt tells the model its role, gives it the right context, constrains the output format, and handles edge cases. You don't say "summarize this feedback." You specify what kind of summary, how long, what to emphasize, what to leave out, and what to do if the input is empty.
The difference between a vague prompt and a specific one is the difference between an output you have to rewrite and one you can ship directly.
It's a PM skill
When building AI features, someone has to own the prompt layer. If nobody does, engineers write something that technically works and PMs wonder why the output is always slightly off. That's a prompt problem, not a model problem.
Write prompts the same way you'd write acceptance criteria: precisely, with examples, with defined failure states. Prompt changes are also instant deploys, so iteration is fast.