AI models are getting easier to use.
They understand ordinary language better. They follow longer instructions more reliably. They can often infer what you meant even when your request is not perfectly structured.
So it is reasonable to ask:
Do prompts still matter?
Yes.
But perhaps not in the way people once thought they did.
The future of good prompting is probably not learning a collection of secret phrases, elaborate formulas, or perfectly engineered sentences.
It is learning how to direct a capable system toward the result you actually need.
That distinction matters.
A stronger AI model can compensate for imperfect wording.
It cannot automatically know your real purpose, the context you forgot to provide, the constraints you never mentioned, the source material you intended it to use, or the standard by which you will judge the result.
Better AI reduces the importance of prompt tricks.
It does not eliminate the importance of human direction.
The old idea of prompting was often too narrow
For several years, prompt engineering was often presented almost like a technical code.
Use this phrase. Put the instructions in this order. Tell the AI to “act as an expert.” Add a long list of rules. Use a particular template.
Sometimes those techniques genuinely helped, especially with earlier models.
But they also encouraged a mistaken idea:
If you can find the right words, the AI will somehow know exactly what you need.
That was never the strongest way to think about prompting.
A prompt is not a spell. It is an assignment.
And a better AI model does not make good assignments irrelevant.
Better models can handle worse wording
Modern models are often much better at understanding casual requests than earlier systems were.
You may no longer need a 500-word template to get a useful answer.
“Compare these two proposals and tell me which one better addresses the customer’s actual problem.”
That is progress.
But notice what still exists inside that short prompt:
- a task;
- two things to compare;
- a decision criterion;
- and an intended outcome.
The prompt got shorter. The direction did not disappear.
What matters less: ritualized wording
If a model is capable enough, small wording differences often matter less than they once did.
You probably do not need to obsess over whether you wrote “Act as a senior marketing strategist” instead of “Evaluate this marketing plan.”
You do not need to collect hundreds of supposedly perfect prompts.
A stronger system can often work from a clearer, shorter description of the outcome.
That lets us spend less time engineering sentences and more time thinking about the work.
What still matters: the aim
Suppose you ask:
“Help me with this report.”
A capable AI can certainly do something with that.
But what does success mean? Do you want it summarized, edited, fact-checked, reorganized, shortened, converted into an executive briefing, or evaluated for weaknesses?
The model may make a reasonable guess. But a reasonable guess is still a guess.
That is why the first part of AIDSRVA remains Aim.
Before worrying about prompt wording, know what you are trying to accomplish.
Context still matters
Imagine giving an excellent employee this assignment: “Write the email.” They still need to know who is receiving it, what happened, what decision was made, what must not be disclosed, what tone is appropriate, and what action the recipient should take.
A more capable model can infer more from less information. But inference is not a substitute for context when the missing information changes the answer.
The practical question is not “How much information can I stuff into the prompt?” It is “What information would change the result?”
That is the Inform stage.
Direction still matters
Better models can often infer an obvious task. But whenever the action matters, say what you want done.
Compare “Here are my meeting notes” with “Turn these notes into a decision summary. Separate decisions already made from unresolved questions. Do not invent owners or deadlines.”
The second request is better not because it contains magical syntax. It is better because the human supplied clearer direction.
Shape still matters
An AI may know the subject well and still produce the wrong kind of answer.
Maybe you need three options rather than twenty, a table rather than paragraphs, language suitable for a beginner, an executive summary rather than a technical explanation, PowerShell rather than Bash, recommendations separated from verified facts, or a draft rather than a final decision.
The Shape stage tells the system what useful output looks like.
Review becomes more important as AI becomes more capable
There is a paradox here.
As AI improves, people may feel less need to review it. The answers look better. The reasoning sounds stronger. The output needs less editing.
But capability is not authority.
The better the output looks, the easier it may become to forget that consequential claims still need verification and consequential actions still require human approval.
A more capable model can reduce the amount of correction you need to do. It cannot assume responsibility for what you decide to use.
Prompting is becoming broader than the prompt box
Modern AI systems may use files, tools, persistent instructions, project context, connected data, agents, and multistep workflows.
The effective context may include your request, previous conversation, uploaded documents, project instructions, available tools, retrieved information, examples, system rules, and application state.
That means the important skill is increasingly broader than writing one clever prompt.
What does the AI need to know, what should it do, what boundaries apply, and how will I judge the result?
That is much closer to direction and orchestration than to finding perfect wording.
You can often use shorter prompts—but you still need clearer thinking
“Review this proposal. Identify the three biggest risks, explain why each matters, and tell me what evidence would change your assessment.”
That is not complicated, but it contains strong human thinking: the task is review, the quantity is three, the focus is risk, reasoning is required, uncertainty matters, and evidence can alter the conclusion.
The intelligence of the model allows the wording to become simpler. It does not make the human thinking behind the request less important.
AIDSRVA becomes more useful, not less
Aim. Know what you need.
Inform. Supply the context that changes the answer.
Direct. Tell the system what work you want performed.
Shape. Define what a useful result looks like.
Review. Examine what the AI produced.
Verify. Check what matters.
Approve. Decide whether the result should actually be used.
The better the AI becomes, the less those stages need to feel like a rigid prompt formula. They become a way of thinking.