Prompt engineering: the skill that won't stay still

When this all started, getting a model to do what you wanted felt like an arcane art, in fact it was a job role in itself! You'd coax it, trick it, wrap your instructions in elaborate rituals. Please return JSON. Only return JSON. Do not return anything other than JSON. And still, half the time, you'd get a friendly preamble before the curly braces, or it would decide to explain its reasoning, or it would suddenly hallucinate a field that didn't exist.

Consistency was the battle. You'd get something working perfectly, run it ten more times, and watch it drift. The same prompt, the same input, different outputs. You'd add more instructions, more guardrails, more "I really mean it this time" language. Sometimes it helped. Sometimes it made things worse.

That was prompt engineering in 2023. A discipline built on frustration and folklore.

What changed

The models got better. Not just smarter in the general sense, but more controllable. More willing to follow instructions. More consistent in their outputs.

The biggest shift was structured outputs. Instead of begging the model to return JSON and hoping for the best, you can now define a schema. Here's the shape I want. These are the fields. These are the types. The model doesn't have a choice about whether to comply. It's constrained by design.

This changed everything. Suddenly the elaborate prompt gymnastics weren't necessary. You didn't need to threaten or plead or repeat yourself. You defined what you wanted, and you got it.

The same thing happened with instruction following more broadly. The gap between "what I asked for" and "what I got" narrowed. Models became better at understanding intent, better at staying on task, better at not randomly deciding to do something different.

The dark art started feeling more like a craft.

The knack

There's still a knack to it. Good prompting isn't just about the words you use. It's about understanding how the model thinks, what it needs to do a good job, what trips it up.

Context matters enormously. The more relevant information you give, the better the output. Not just what you want, but why you want it. Not just the task, but the background. The model isn't reading your mind. It's working with what you've given it.

Specificity beats vagueness. "Write something good" will get you something generic. "Write a two-paragraph summary for a technical audience who already understands the basics" gives the model something to aim at.

Examples work better than explanations. Instead of describing the format you want, show it. One good example is worth a hundred words of instruction.

Structure helps. Breaking complex tasks into steps, using clear sections, giving the model a framework to follow. Not because it can't figure things out, but because clarity in means clarity out.

Prompts writing prompts

Here's the thing that still surprises people: you can get prompts to write prompts.

If you're struggling to articulate what you want, describe the problem to the model and ask it to help you write a better prompt. It's seen more prompts than you have. It knows what works.

If you've got a prompt that's nearly there but not quite, paste it in and ask for improvements. What am I missing? How could this be clearer? What edge cases haven't I covered?

This feels like cheating, but it's just using the tool well. The model is good at language. Prompts are language. Let it help.

We do this constantly now. First drafts written by us, refined by the model, tested, adjusted, refined again. The collaboration applies to prompts themselves, not just the outputs they produce.

The skill that won't stay still

The tricky thing about prompt engineering is that it keeps changing.

Techniques that were essential a year ago are often unnecessary now. The elaborate workarounds we developed for getting consistent outputs are less relevant when structured outputs exist. The careful instruction layering we used to do is less critical when models follow instructions better.

At the same time, new capabilities create new patterns to learn. Extended thinking, tool use, multi-turn orchestration. Each new feature is a new surface to explore.

The skill isn't learning a fixed set of tricks. It's staying curious about what works today, being willing to throw out what worked yesterday, and constantly testing assumptions against reality.

People ask whether prompt engineering will become obsolete. Whether models will get so good that it doesn't matter how you ask. Maybe. But we're not there yet, and even if we get there, the underlying skill is really just clear communication. Knowing what you want, articulating it precisely, giving the context needed to do a good job.

That skill won't become obsolete. It'll just change shape again.