How AI Coding Changed Completely in 18 Months — Is Prompt Engineering Dead?

 

In late November 2024, I used AI for the first time.

Eighteen months later, the changes I've witnessed aren't just about better tools. The entire way of working has transformed — and the pace of that transformation is accelerating. At first, a new paradigm emerged roughly every six months. Then every three months. Now, something new drops almost every week.

AI has moved past its infancy and entered a full-blown transition period. Standing in the middle of it, I thought it was worth looking back at these 18 months.

---

The Past — Copy-Paste and Prompt Engineering

The Chat Window Era

To be precise, it started with asking ChatGPT about code. I'd paste a function or an error message, get a response, copy it, and move it to my editor. The novelty of asking AI instead of searching Google was refreshing, and I was genuinely impressed by the accuracy. But fundamentally, the workflow wasn't all that different from a search engine. Ask a question, read the answer, manually apply it. Repeat.

At the time, I thought even this was revolutionary enough.

The Discovery of Prompt Engineering

After a few months, the limitations became apparent. The same question could yield vastly different answers depending on how you asked it. "Fix this code" would get you a sloppy patch, but assigning a role, explaining context, and specifying the output format produced results on an entirely different level.

In early 2025, I dove into prompt engineering.

System prompt design, chain-of-thought, few-shot learning, role assignment. I sometimes spent as much time crafting a prompt as writing the code itself. And it worked. The same model could swing between junior-level and senior-level responses depending on the prompt. I was convinced: if you want to truly leverage AI, prompt engineering is essential.

That conviction didn't last long.

The Decline of Prompt Engineering

By mid-2025, CLI-based AI tools like Claude Code and OpenAI Codex appeared.

These tools didn't work through a chat window. They read your entire codebase directly from the terminal, modified files, and ran tests. They operated with full project context. Without me crafting elaborate prompts, the tool itself understood the context and took appropriate action.

I could feel prompt engineering's importance dropping rapidly.

It was hard to accept at first. A skill I'd spent months learning was losing value this fast? But I soon realized why. Reverse prompt engineering had emerged — ask AI to "create the optimal prompt for this task," and it would generate something more precise than anything I could design manually. The skill of writing good prompts was being replaced by the skill of asking for good prompts.

The fundamental principles of prompt engineering still matter, of course. Understanding how AI interprets input makes communication more efficient. But the act of manually fine-tuning elaborate prompts is becoming increasingly pointless.

Looking back, everything up to this point was AI's infancy. Tools appeared, people learned how to use them, new techniques were discovered. The cycle of change ran on roughly six-month intervals. A new paradigm would show up about every half year.

---

The Present — In the Thick of Transition

The End of Copy-Paste

Looking back, the most significant shift was in the workflow itself.

Early AI usage looked like this: type a request in the chat window, AI shows code, copy that code, paste it in your editor, run it, if there's an error copy the error message back into the chat. Repeat. It was an endless loop of context switching, and half the work was spent on copying and pasting.

Now it's completely different. CLI tools read the entire codebase. They find and modify the files that need changes on their own. They test the modified code. If tests fail, they analyze the cause and fix it again. They plan, execute, verify results, and adjust the plan when needed. The entire sequence that a developer goes through — AI now performs autonomously.

My role has shifted to providing direction and reviewing results. Instead of writing code line by line, I describe what to build and review what AI produces. The developer's role is moving from author to architect and supervisor.

The Rise of Agent Teams

And now, 2026. AI utilization has evolved yet another step.

Beyond a single agent writing code, multiple agents now form teams and collaborate. A strategy agent designs the direction, a production agent executes, and a quality agent reviews. Each agent specializes in its own domain, and an orchestrator coordinates them all. It's the structure of a software development team, reproduced in AI.

What's even more fascinating is the recent trend of building AI companies. Create a CEO agent, and it establishes the company's vision and recruits sub-agents like a CTO and CPO. These agents then form their own business teams. It mirrors exactly how humans build organizations. When I first heard about this, I thought it was hype. But after building and running an agent team for my own blog, I can feel that this direction is more than a passing trend.

The Accelerating Cycle of Change

There's a reason this moment feels like a transition period.

Just a year ago, the AI ecosystem's cycle of change ran on roughly six-month intervals. ChatGPT arrived, and about six months later prompt engineering gained prominence. Another six months, and CLI tools appeared. But starting in late 2025, that cycle compressed to three months. Agent coding, multi-agent systems, MCP protocols, reverse prompt engineering — new concepts poured out every quarter.

And now, in 2026, something new drops almost every single week. This week it's a new model, next week a new framework, the week after that an approach that upends the existing paradigm. The interval between changes has compressed from six months to three months to one week.

This isn't the pattern of an infancy. During infancy, big changes happen occasionally. What's happening now is a relentless stream of smaller changes continuously reshaping the entire landscape. The next technology arrives before the current one matures, and the one after that arrives before it can take hold. This is the textbook shape of a transition period.

---

The Future — What Lies Beyond the Transition?

Every Transition Period Ends

Every transition eventually resolves. The internet went through this. During the dot-com bubble of the late 1990s, new services and business models appeared weekly. Most vanished; the survivors shaped the internet we know. Smartphones went through it too — dozens of mobile operating systems proliferated at first, but eventually converged to two or three platforms.

AI will converge as well. Most of the tools and frameworks flooding in every week will disappear, and the survivors will become standards. When that happens, today's chaos will settle, and AI will become infrastructure as unremarkable as electricity or the internet.

What Should We Do Until Then?

The problem is that nobody knows how long the transition will last. It could be one year, or it could be five. In the meantime, only one thing is certain: the way I work today will soon be obsolete.

Here's the timeline:

• November 2024 — Asking questions in a chat window, copy-pasting code (change cycle: ~6 months)
• Early 2025 — Learning prompt engineering, designing elaborate prompts
• Mid 2025 — CLI tools arrive, prompt engineering importance drops (change cycle: ~3 months)
• Late 2025 — Reverse prompt engineering, AI designs the prompts
• 2026 — Agent team composition, AI organization building (change cycle: ~1 week)
• The Future — ?

When I was diligently learning prompt engineering, I thought it would be the core competency of the AI era. Less than six months later, CLI tools had largely replaced that need. Agent teams are the frontier now, but at this rate of acceleration, another paradigm could emerge within months.

What matters in AI proficiency isn't any specific skill — it's the speed at which you adapt to change. More important than writing great prompts is the ability to quickly evaluate new tools and integrate them into your workflow when they arrive.

A transition period is unsettling. What you learned yesterday could be useless today. But because the landscape hasn't solidified, it's also an opportunity. An unchosen playing field means anyone who adapts fast enough can take the lead.

The way I work today will be outdated soon. The flexibility to accept that might just be the most realistic strategy for surviving AI's transition period.