Key Takeaways

• Prompt engineering is shifting from a standalone craft to a background skill, because stronger chat models can generate prompts for specialist tools.
• Vibe coding still benefits from good inputs: clear requirements and a few explicit guardrails beat “best practices” hand-waving.
• Even when the AI writes the prompt, responsibility doesn’t go away: define what “done” means and verify outputs (tests, review, sanity checks).

Anyone working with AI professionally knows what prompt engineering is. Even occasional AI use (OpenAI’s ChatGPT, Google Search with Gemini) runs into prompt engineering every time a question is asked, even if it doesn’t get much attention. In essence, prompt engineering is the art of asking the right questions to get the best possible answers.

In fact, there’s a new discipline that’s been introduced recently—a quick search for prompt engineering courses returns an impressive list. It’s easy to feel late to the game without that kind of training, but maybe that’s the wrong worry. Here’s the take: prompt engineering is already becoming optional in a lot of real work—because one AI can write the prompt for another AI. The “prompt engineer” is getting automated.

Vibe coding is where prompts hurt

In practice, prompts matter most in one major use case: vibe coding. The “ask AI chat personal questions” category can wait. If the term “vibe coding” is unfamiliar: it’s when a special AI coding assistant is told what the software should do, and it writes the code.

Some people use “vibe coding” in a narrower sense—almost “no tests, just vibes.” Used more broadly here: any back-and-forth where intent is described and the AI produces code. Without getting into the argument about “Is using pro coding agents by pro devs considered vibe coding? "let’s call any such communication between the operator and the AI “vibe coding.”

So, in vibe coding, getting decent-quality software—not just a simplified proof of concept—still requires more or less detailed instructions (those instructions are called a prompt). So here comes the art of prompt engineering.

But even without prompt-engineering skills, and maybe without programming experience, vibe coding can still produce decent software. Didn’t AI get invented to solve all the problems?

Let the ‘big’ model write the prompt

With access to a powerful AI model—which is available to just about everyone these days with a paid subscription—that “super AI” can create the right prompt for its little brother: a coding agent.

By “pro AI chat,” this means a general-purpose chat model that’s good at reasoning and asking clarifying questions (examples: OpenAI GPT 5.2 Pro, Google Gemini 3 Pro). By “coding agent,” this means a tool that can actually write/edit source code files in a repo (and run tests and builds) based on instructions (examples: OpenAI Codex, Microsoft Copilot Agent).

With pro AI chats, the workflow can run in dialog mode: the problem to solve and very high-level requirements (like “follow best security practices”) can be described in normal human language—no prompt-engineering degree required.

A real example of what this looks like:

• Requester (to a general chat model): “Generate a prompt for a coding agent to build a small internal web app that ingests spreadsheets, validates them, and stores results. Use role-based access and don’t leak secrets. Ask any missing questions first.”
• AI: “Who are the users? What auth system? What’s the data retention policy? Any compliance constraints? Preferred tech stack?”
• Requester: Quick answers in plain English.
• AI (output): A structured, detailed prompt ready to paste into the coding agent, including tech choices, acceptance criteria, and a checklist.

That flow is basically “prompt engineering,” but done by the AI “supermodel” with an IQ higher than the coding agent’s model.

One caveat: “follow best security practices” is not a magic spell. When security matters, it still helps to name a few non-negotiables (auth model, secrets handling, dependency policy, logging rules, threat model, etc.). The good news is that those non-negotiables can be stated in normal language, and the “super AI” can translate them into a crisp prompt.

The useful preface is “Generate a prompt for a coding agent to …” Specifying the agent type and the model it uses can produce an even more precise prompt. That’s it. Strong chat models will usually ask a few questions to clarify requirements further and will get it done. Then the result can be copied to the coding agent, and it will take it from there.

This straightforward process works like a Swiss watch, as they say in Russian—meaning “works like clockwork” (or, loosely, “works like a charm”)—with one exception: it’s much cheaper than Swiss watches.

What happens next

As a “bonus track,” a reasonable guess for the next 12 months: the “lower-level” AI (a coding agent, in this case) will be powerful enough to retain enough information about the operator—preferences, work environment, recurring constraints—to infer requirements on its own. It might also ask a couple of questions—just to clarify—before executing the task. There will be less need to craft prompts using special techniques.

Prompts won’t disappear forever. The skill of prompt engineering is getting squeezed for everyday users and everyday tasks. Sorry, prompt engineers.

With that said, even if the AI writes the prompt, two responsibilities remain: (1) stating what’s actually wanted (requirements), and (2) checking what comes back (tests, review, sanity checks). The “prompt engineer” role shrinks, but the “responsible adult” role doesn’t. As noted in RSAC blog The Dark Side of AI Dependency, over-reliance on these tools can erode fundamental problem-solving skills, making the “responsible adult” role even more vital.