Prompt engineering probably isn’t the skill you think it is

I had a really interesting conversation with my brother last night about prompt engineering and optimisation.

There is obviously a lot of noise around this at the moment. Social media is full of people sharing the perfect prompt, selling enormous prompt libraries and creating increasingly elaborate frameworks for asking an AI to do something.

Some of it is useful.

Some of it feels like the entirely predictable process of taking “clearly explain what you need” and turning it into a product.

I use Codex a lot. Most of the time I talk to it in much the same way I would talk to a human co-developer I know.

I explain what I’m trying to achieve, provide the relevant context, point out anything unusual and make it clear what I don’t want it to break. If the task is complicated, the explanation is normally complicated. If the task is simple, the prompt can be a sentence.

I’m rarely sitting there trying to construct the perfect prompt.

I’m trying to clearly communicate a piece of work.

That distinction feels important, because prompt engineering is increasingly presented as a technical discipline in its own right. Learn the correct structure, remember the right phrases and better answers will apparently follow.

There are certainly techniques that help. Context matters. Examples help. Clear constraints help. Explaining the intended audience usually helps.

But those things improve the response for the same reason they improve a brief given to another person.

You need to understand what you are asking for.

You need to know what outcome you actually want, which details are relevant and where the ambiguity might be. You need some sense of what success looks like and what the consequences are if the answer is wrong.

That is true whether you are asking an AI, briefing a developer, delegating something to an employee or asking a consultant for advice.

“Make the website better” is not a useful request.

Better in what way?

More enquiries? Faster page loads? Clearer messaging? Better accessibility? A complete visual redesign? Fewer people ringing the office because they cannot find the opening hours?

Until you understand the problem properly, there is only so much prompt optimisation can do for you.

The other half of this is understanding the answer.

AI is exceptionally good at producing responses that look complete. They are normally well structured, confidently written and plausible enough that accepting them is often easier than properly interrogating them.

That is incredibly useful when the answer is good.

It is slightly more problematic when the answer is subtly and convincingly wrong.

When Codex suggests a change, I still need to understand what it has changed. I need to know whether its assumptions are sensible, whether it has solved the actual problem and whether the solution has introduced another problem somewhere else.

I do not necessarily need to have known the answer before asking. Exploring unfamiliar approaches is a huge part of the value.

But I need enough understanding to evaluate what comes back.

The same applies when using AI for marketing, business planning, financial analysis, recruitment, customer service or almost anything else.

A response does not become trustworthy simply because it arrived quickly and was formatted nicely.

This is the part of the current AI conversation that concerns me more than prompt writing.

There is a huge amount of focus on helping people generate more output, with rather less attention given to whether they are equipped to judge it.

AI can help someone produce far more code, copy, analysis or strategic advice than they could have produced alone.

That sounds like a straightforward productivity gain.

It probably is, provided they can distinguish the useful work from the plausible nonsense.

Without that ability, they may simply become much faster at creating problems.

None of this means you need to become an expert in everything before using AI. That would rather defeat the point.

It does mean expertise still matters.

Experience helps you ask better questions because you understand where the complexity is. It helps you notice what has been missed. It helps you test the result and ask a useful follow-up rather than simply pressing accept.

The skill is not learning to write the perfect prompt.

It is learning to fully understand the request you are making, and the response you are given.