# Prompt engineering is not the hard part

> Clever prompts get the attention. The things that actually decide whether an AI feature works in production are less glamorous — context, evaluation, guardrails, and cost.

*Source: https://www.lazlosoftwaresolution.com/resources/prompt-engineering-is-not-the-hard-part*

Prompt engineering gets a lot of attention, and a good prompt does matter. But if you are building an AI feature that real people depend on, the prompt is one of the smaller problems. The things that decide whether the feature works — and keeps working — are less glamorous and much more like ordinary engineering.

## Context beats phrasing

A model can only be as good as the information it is given. Most of the difference between a useful answer and a useless one is not how the question was phrased — it is whether the right context was retrieved and included. Time spent making retrieval good pays off far more than time spent polishing wording.

## If you cannot measure it, you cannot improve it

Changing a prompt and eyeballing a few outputs feels like progress and is not. Without a set of real cases with known good answers, every change is a guess, and you will not notice when one quietly makes things worse. Evaluation is the unglamorous work that turns tinkering into engineering.

## Assume the output can be wrong or hostile

A model's output is untrusted input. It can be wrong, and if users can influence the prompt, it can be manipulated. Anything the model produces that flows into another system — a query, a command, a message sent on the user's behalf — needs the same validation you would apply to any input from outside. Never wire a model's raw output straight into an action that matters.

 - Validate and constrain what the model can trigger.

 - Keep a human in the loop for anything costly or irreversible.

 - Log inputs and outputs so you can see what actually happened.

## Cost and latency are design decisions

Per-call cost is small until traffic multiplies it, and a slow model call is a slow feature. Which calls can be cached, which can use a smaller model, and which need the largest one — these are choices to make deliberately, and they are easier to make before launch than after the first invoice.

> The prompt is the visible part. The engineering around it — context, evaluation, guardrails, cost — is what makes the feature dependable.

## The reframe

Treat an AI feature like any other part of the system: know your inputs, validate your outputs, measure whether it works, and account for cost and failure. Do that and a decent prompt is plenty. Skip it and no prompt, however clever, will save you.