I don’t write most of my prompts by hand anymore, and it has nothing to do with prompting mattering less.

A lot of what I now build starts as research, a plan, and a pile of requirements I’ve already worked out before writing any prompt.

Typing all of that back out into a “well-crafted prompt” by hand means redoing work I’ve already done, in a worse format.

So I don’t. I hand the AI the research and plan, and it writes the prompt.

That can sound a little strange, if you were told, like I was, that learning to prompt well is THE AI skill. I’d argue it isn’t.

“Write a better prompt” stopped being my answer to every problem because most of the time the prompt was never actually the problem 💁‍♀️

When something I’m doing with AI isn’t working, I no longer reach for better wording first. Instead, I ask what’s broken.

Is it the instruction? What the model has access to? Whether it can take the action at all?

Or is it just inconsistent, and no amount of rewording is ever going to fix that?

Prompting answers exactly one of those. Figuring out which one you’re looking at is most of the actual skill.

Note 📌
Whether you’re a developer, marketer, or managing a team that’s adopting AI tools, if someone told you to “learn prompt engineering” as THE AI career move, this is for you too. Prompting is one of the easier AI skills to pick up on the surface, which is exactly why it gets handed out as advice so often. What most advice usually skips is that making it hold up for real, repeated work is a different thing.

What AI Job Postings Show About Prompt Engineering in 2026

I wanted something more solid than a feeling, so I looked and found that across 315 AI and machine learning job postings in September, prompt engineering showed up in about 5% of them, while agents showed up in close to 40%.

Considering how much “learn to prompt” is still floating around as career advice, that’s not a gap you’d expect.

Tip 💬
That gap doesn’t mean prompting stopped mattering. It means almost nobody is hiring a dedicated “prompt writer” anymore. The work got folded into the job itself, so the developer who owns an AI feature is expected to own its prompts too, not treat it as a separate line on a resume. I can attest to this as a developer who manages prompts (usually within workflows).

Keep in mind that this is a keyword count from a specific, startup-leaning sample, not a full picture of the labor market.

What it does show is that the vocabulary moved away from “prompt” as the headline skill toward “agent” and “eval” as the things people are actually building.

Though prompting didn’t die, it stopped being the whole job.

Why a Better Prompt Doesn’t Always Fix the Problem

A prompt is an instruction. Usually, a clearer prompt is the fix when a model is misunderstanding the task or ignoring a constraint you gave it. Definitely worth learning to write well, no argument there.

But an instruction can’t hand the model a fact it never had (aka missing context).

For example, if an AI assistant invents a company policy, that’s not a wording problem. It’s working without the real policy, and no amount of rephrasing the question gets it something it was never given.

You could write the most elegant prompt in the world and it still wouldn’t know what it doesn’t know.

It’s the same story with permissions. If you need an agent to update a record somewhere or send something on your behalf, the thing standing in your way is what it’s allowed to touch (aka it’s tool access), not how nicely you asked or how precisely you described the task.

And then there’s the one I run into the most, where something works fine on one input, and quietly breaks on the next. No prompt catches that.

You need a way to actually check, which is what evals are.

An eval is really just a test, an evaluation. You feed the AI real inputs, check the output against what you expected, and score it.

Think of it like a tweak that fixes one case but can quietly break three others, and you won’t know until someone else finds it first 😬

So far, we’ve outlined four different problems here. None of them get solved by writing a better sentence.

How to Diagnose What’s Actually Broken: Prompting, Context, Access, or Evals

Four problems is a lot to hold in your head mid-task, so I don’t actually sort through them by thinking harder. I ask the AI.

Say the output’s wrong, and I can’t immediately tell why. Instead of guessing at a better prompt, I’ll ask directly: what information would you need to answer this correctly? (Usually after analyzing what we’ve done.)

If it names something it clearly doesn’t have, like a document, a number, a piece of context I never gave it, that’s not a prompting problem anymore. That’s missing context, and now I know what to go get instead of what to rewrite.

The same move works for the other three. Ask whether it understood the task the way you meant it, and a prompting problem usually surfaces on its own.

Ask what it would need to actually complete the action instead of just describing it, and a permissions problem shows up fast.

And if you can’t tell whether an answer that looked right actually is right, that’s not something a better question fixes at all. That’s the moment you need an eval, a way to check instead of a way to ask.

Though I’m using a prompt to identify the problem, keep in mind that it’s a different kind of prompt.

Instead of asking the AI to do the task (an execution prompt), it’s asking to tell you what’s missing before you waste another round guessing. Think of it like a utilitarian prompt 🛠️

Related: What Is an AI Harness? You’re Probably Already Using One

Why Prompt Engineering Advice Breaks Down on Multi-Step AI Projects

I started building something with real scope, multiple phases, and features that depended on each other only to discover that everything I’d learned about prompting (i.e., be specific, give examples, state your constraints) wasn’t enough by itself anymore.

Notice I’m saying it’s not enough. I’m not saying it’s not relevant.

A prompt for something complex like that (aka most things that go beyond demos) comes out of research I’ve already done, a system design I’ve already worked through, requirements I’ve already written down somewhere, usually across a few different docs by the time I’m ready to build.

Typing all of that back into prompt form manually means redoing finished work, rather badly, in a worse format, and losing half the specificity in the translation.

So I stopped doing that part by hand. I feed the AI the compiled research and the plan, and it builds the prompt from that.

This means I’m now prompting the thing that writes my prompts. Literally, meta on meta, though these days that’s not unique.

The first time I caught myself doing it, I stopped and wondered if I’d made the whole process needlessly complicated. But I hadn’t.

The scope had just gotten bigger than a single hand-written prompt could reasonably hold. It warranted execution finesse, however that took form.

Ironically, I thought I’d get better at writing prompts out of all this. What I actually got better at was catching myself before I wrote one.

The lesson was pausing long enough to ask whether the thing in front of me was really a wording problem, or just looked like one because wording was the tool I already had in my hand.

Explore: A Smart Free Chrome Extension That Upgrades AI Prompts

Why Saved Prompt Templates Stop Working on the Same Task

Say you’re using AI for the same work task every week, something ordinary, not an app or serious build. The first time, you write a decent prompt and get a decent result. You save it.

The second time, you reuse it, and it’s fine.

The fifth time, the inputs have drifted just slightly from what they were originally, and that same saved prompt starts missing things it used to catch.

You didn’t get worse at prompting, and the prompt itself didn’t break. What changed is everything around the task:

• what you’re starting from this time
• what’s different about this particular input
• what you actually need out of it now

A template treats a task like it’s always the same task. As in it will produce the same result, and it’s affected by, and based on, the same underlying principles.

The problem is that it most likely isn’t.

However similar a task may be, things that affect it differ just slightly, and that’s the part that a saved prompt can’t see on its own.

It’s also the part nobody mentions when they hand you “just write a good prompt” as the whole skill.

Hey! I’m not against prompt templates, they have they’re own use. In fact, I support that very feature on my free Prompt Optimizer Chrome extension because I was using templates so much while browsing. The key here is to avoid making prompting your sole AI power skill and identity.

The Mistake on the Other Side: Learning Evals or Agents Too Early

Now, this might come as a surprise, but if I had to warn you against one thing after all this, it wouldn’t be sticking with prompting too long.

In fact, I would warn you against grabbing for evals or agent orchestration before you actually need them, just because those are the words popping up everywhere right now.

A two-step prompt doesn’t need an agent framework built around it.

This isn’t a developer-only trap either. The same question of “what’s actually failing here” works whether you’re debugging a tool call or figuring out why your AI-drafted email keeps missing the point on the fourth try.

Being non-technical doesn’t exempt you from asking it 😌

And don’t swing the other way and drop prompting entirely just because the job title stopped showing up in listings. The skill didn’t go anywhere; it only stopped being the whole job (it never was, let’s be real).

It’s a Wrap

I went from writing most of my prompts by hand to writing almost none of them, and somewhere in there I stopped feeling like that was a loss.

The work didn’t get easier. It moved up a level, into the research and planning that now builds the prompt itself, and that part was never going to fit in a template to begin with.

That’s probably the reason the original advice never quite worked for me the way it was supposed to. It was teaching the layer I’d already outgrown.

Now, do share: what’s the last AI task that quietly stopped working for you, where the fix probably wasn’t a better sentence?

Thanks for reading this one all the way through 👋