A colleague once showed me a prompt he was unusually proud of. It was almost two pages long—headings, rules, role descriptions, forbidden phrases and a stern warning that the AI must “think carefully before responding.”
He called it his master prompt.
It worked for three weeks. Then the team changed the kind of report they were producing, somebody pasted in messy source material, and the whole thing began returning polished nonsense. They spent more time repairing the prompt than writing the report.
That story comes to mind whenever someone announces that prompt engineering is dead.
The old version may be dying—the one built around secret wording and giant LinkedIn templates. Good. I won’t miss it. The useful part has moved closer to ordinary problem solving, which is less glamorous and more valuable.
Magic words were always a strange business
For a while, prompts were treated like cheat codes.
People sold collections containing phrases such as “act as a world-class expert” and “use first-principles thinking.” The implication was that the model had a hidden genius mode and you merely needed the correct incantation.
I tried some. Occasionally the output improved. Often it just became longer.
Here’s the thing: saying “be brilliant” doesn’t tell the AI what brilliant means for your task. A sharp note to a busy finance director is different from a detailed explanation for a new employee. The model can’t know that your manager hates tables or that yesterday’s meeting changed the plan.
Fancy wording was never the main skill. Supplying the missing reality was.
Better models didn’t remove unclear thinking
AI systems have become easier to talk to. You can write a normal request and often receive something usable. You don’t need to assign the model twelve professional identities before asking it to summarise a PDF.
That’s progress.
It also creates a new mistake. People assume that because the model understands casual language, it understands an unfinished thought.
Someone types, “Write a launch email for our new feature,” and dislikes the generic result. But which customers are receiving it? What changed? Is the feature already available? Should the email drive upgrades, encourage trials or reduce support questions?
The AI didn’t fail to follow the prompt. The prompt contained a hole.
Most weak outputs begin there—not with bad grammar, but with a person who hasn’t decided what they want.
Context beats cleverness on an ordinary Tuesday
I once watched a team use AI to prepare a weekly project update. Their prompt was enormous, full of tone instructions and formatting rules. Yet it kept missing the decisions senior managers cared about.
The fix was embarrassingly plain.
They added the milestone, last week’s commitments, current blockers, decisions needed and the audience. The prompt became shorter. The output became better.
This is where prompt engineering is heading. Less word decoration, more context design.
A useful request might include a few Slack messages, a customer complaint, the latest product note and a clear definition of the decision being made. None of that looks impressive in a screenshot. It does help the model produce something connected to the real situation.
Anyway, templates are fine. A template should remind you what information to provide. It shouldn’t replace thinking about the current job.
The first answer can be mediocre
Another odd habit is treating prompting like a one-shot examination.
You submit the perfect instruction. The AI returns the perfect result. Everyone goes home early.
Real work rarely behaves that way. You read the draft and notice the opening is too cautious. A key risk is missing. One paragraph sounds as if it was written by a management consultant trapped inside a brochure.
So you respond.
“Keep the recommendation, but make the reasoning more direct.”
“Use the customer’s exact concern from the notes.”
“Cut the first section. The audience knows the background.”
That is still prompt engineering, though the phrase feels too grand. It’s closer to editing, briefing and reviewing. The skill is not getting everything right in one message. It’s noticing what is wrong and steering the work without starting over.
People who are good with AI tend to be good at this loop. They inspect rather than admire.
One heroic prompt is usually the wrong idea
The biggest change has little to do with prettier instructions. Useful AI work is being divided into steps.
Instead of asking for a complete market report in one breath, you might first extract facts, then identify patterns, challenge the strongest conclusions, and only after that draft the report. Each step has a smaller job and a result you can check.
This matters in coding too. Asking an AI to “build the app” often creates a confident pile of files. Asking it to inspect a GitHub issue, propose an approach, identify affected modules and make one reviewed change is slower for five minutes and faster for the next two hours.
The same applies to support, recruiting and finance. Good systems bring in the right information, restrict what the AI may do, check the output and pass useful results forward.
The prompt becomes one part of the machinery. Not the machinery itself.
The boring part is checking
If a prompt produces a nice answer once, that proves very little.
Try it on an incomplete document. Try it when the customer is angry. Try it with two policies that contradict each other. See whether it invents a number when the source is silent.
This work is boring, and therefore important.
A business does not need someone who can make an AI look impressive during a demo. It needs someone who can tell whether the output is correct enough to send, useful enough to save time and safe enough to place inside a real process.
That person may be called an analyst, designer, engineer or operations lead. Probably not “prompt engineer” forever. Job titles become less dramatic once the work is understood.
Clear instructions will survive the label
I don’t think memorising prompt formulas is a serious career plan. Models will improve, interfaces will become more conversational, and software will quietly supply context that users once had to paste manually.
But defining the goal, selecting the right information, breaking work into sensible steps and judging the result—those aren’t temporary tricks. They are the job.
The irony is that prompt engineering is becoming less about prompts just as AI is becoming more useful.
Which is probably healthy. Nobody needs another folder containing 500 “ultimate” prompts.
A short request, the right documents and someone paying attention after the answer appears will usually do more.
