AI, explained

What is prompt engineering?

It is the practice of writing instructions precise enough that a model behaves the same way on the thousandth request as it did on the first.

The term attracted ridicule during the period when it was mostly folklore: magic phrases, elaborate personas, superstitions passed around without evidence. Most of that has been sanded away by better models. What remains is unglamorous and genuinely valuable, and it looks much more like specification writing than like incantation.

What a good prompt actually contains

• A defined task. Not "help with support tickets" but "classify this ticket into one of these six categories and return the category only".
• The context the model needs. Supplied deliberately, rather than assumed from training data.
• Output format. If something downstream will parse the response, specify the structure exactly, and prefer a schema-enforced output where the API supports it.
• Constraints and edge cases. What to do when the input is ambiguous, out of scope, or missing. An explicit path to "I cannot answer this" prevents a great deal of invention.
• Examples. Two or three well-chosen examples of correct behaviour usually outperform another paragraph of description.

What changed as models improved

Older models needed heavy scaffolding: step-by-step reasoning had to be requested explicitly, formatting had to be coaxed, small rewordings caused large swings. Current models follow instructions far better and reason without being told to. The result is that clever phrasing matters less and clear specification matters more. Prompts have generally become shorter, plainer, and more structured.

Where the real work sits now

In production the prompt is one component among several, and rarely the hardest. Retrieval quality determines whether the model even has the right facts. Tool design determines whether it can act correctly. Evaluation determines whether anyone knows if a change helped. Teams that spend a week rewording a prompt and no time building a test set are optimising the cheapest variable. The professional version of this work is: write a plain prompt, build a set of real cases with known correct answers, measure, change one thing, measure again.

A note on prompt injection

If your system reads untrusted content, web pages, emails, uploaded documents, then that content can contain instructions aimed at your model. Treat everything retrieved as data rather than commands, keep tool permissions narrow, and require approval for consequential actions. This is a security property of the architecture, not something a cleverer prompt can fix.

Prompting is now a skill any competent team can hold internally within a few weeks. The parts that stay hard are retrieval, evaluation, and safe tool design, which is where an outside AI audit tends to earn its cost.

Frequently asked questions

Related