There is a quiet shift happening in software development. For decades, the most valuable skill for a developer was the ability to write code efficiently and correctly. Entire careers were built around mastering frameworks, memorizing patterns, and becoming faster at turning requirements into working software. But the rise of AI coding agents is changing that equation. Today, machines are becoming exceptionally good at the one thing developers spent years mastering: churning out code. Tools like OpenAI’s ChatGPT and AI assistants integrated into editors like Visual Studio Code can already generate large portions of working software with minimal effort. churning out code OpenAI’s ChatGPT Visual Studio Code This does not mean developers are becoming obsolete. Far from it. What is actually happening is that the center of gravity in development work is shifting. The critical skill is no longer the mechanical act of writing code. Instead, it is the ability to clearly define what the code should be so that AI can produce it correctly. center of gravity in development work is shifting clearly define what the code should be Put differently, the modern developer is increasingly becoming a manager of AI coding agents. manager of AI coding agents And that requires mastering something many developers never formally learned: writing effective specifications. writing effective specifications Watch Video Here https://youtu.be/7ddLUYieGow?embedable=true https://youtu.be/7ddLUYieGow?embedable=true The New Core Skill: Understanding Specification Modes One of the biggest misunderstandings about AI-assisted development is the idea that it is simply about writing better prompts. Prompts matter, but the deeper skill lies in understanding the different modes of specification that guide AI coding agents. different modes of specification These modes exist on a spectrum. At one end is what many developers now call vibe coding. This is the most informal style of interacting with AI. A developer describes what they want in broad terms and lets the AI generate code freely. It works remarkably well for prototypes and quick experiments. The speed is impressive, and the barrier to entry is extremely low. vibe coding However, this approach begins to break down when the complexity of the system increases. Architecture becomes inconsistent, naming conventions drift, and hidden assumptions creep into the generated code. For anything beyond small projects, the lack of structure becomes a real problem. At the opposite end of the spectrum is spec-driven development. In this mode, the system is described in detail before any code is generated. The specification might include architecture decisions, data models, API contracts, workflow definitions, validation rules, and infrastructure details. AI then uses that specification to generate the implementation. spec-driven development This approach can produce very reliable results, but it comes with significant overhead. Writing a full specification can sometimes take longer than the coding itself. Between these two extremes lies a middle ground that is becoming increasingly useful in the age of AI coding agents: structured instruction specifications. Instead of writing a full formal spec, the developer creates a focused document that describes the architecture, folder structure, entities, and implementation constraints. This document often lives in a file such as instructions.md, which an AI coding agent can follow during code generation. structured instruction specifications instructions.md Understanding when to use each of these modes is rapidly becoming one of the most important skills in modern software development. Looking Back at the Classic Developer Interview To understand how much things have changed, it helps to look back at a typical developer interview from roughly 2010 to 2018. During that period, a second-round interview for a .NET developer often included a practical coding exercise. A very common challenge was building a small CRUD application backed by two database tables. The front end would typically be implemented using ASP.NET MVC, while Entity Framework would handle the data access layer. Entity Framework The purpose of this exercise was simple. If a candidate could design the models, connect the database, build the controllers, and produce a working application within a limited amount of time, they demonstrated a solid grasp of backend development fundamentals. In that era, writing the code was the test. If we ran that same interview exercise today, the situation would be very different. A candidate could simply open ChatGPT, describe the requirements, and generate a convincing solution within minutes. In fact, someone with very little development experience could produce code that looks reasonably professional. ChatGPT But appearances can be deceiving. Real-world applications require far more than basic CRUD functionality. Enterprise systems demand careful architecture, consistent design decisions, and the ability to debug and refine systems when the initial implementation is imperfect. This is precisely where the developer’s role becomes essential. A Modern Version of the Exercise Out of curiosity, I decided to revisit that classic interview-style challenge—but with a modern workflow. Instead of writing the application manually, I let an AI coding agent generate the implementation. My role was not to type the code. My role was to design the instructions that the AI would follow. design the instructions that the AI would follow To do this, I first used ChatGPT to help draft a structured instructions.md file. This document described the architecture, entities, and implementation constraints that the generated application should follow. The instructions were then fed into an AI coding agent running inside Visual Studio Code. ChatGPT instructions.md Visual Studio Code Importantly, this approach was not pure vibe coding, but it was also not full spec-driven development. It lived squarely in the middle of the specification spectrum. Writing the instructions turned out to be the most time-consuming part of the process. Even with AI assistance, refining the specification took roughly an hour. That is significantly longer than simply prompting an AI and hoping for the best. However, when the AI coding agent began generating the implementation, the benefits became immediately clear. The resulting code followed the expected architecture, respected the structural constraints, and required very little rework. There were no major architectural mistakes and no need to perform large-scale rewrites. In total, the project took roughly three hours. About one hour was spent crafting the specification, and the remaining time went toward resolving small issues and polishing the application. The interesting part is that the time investment shifted. Instead of spending hours writing code, most of the effort went into thinking carefully about the instructions. thinking carefully about the instructions The Real Skill Developers Need Now The biggest lesson from this experiment is that modern development is no longer just about coding. It is about orchestrating AI effectively. orchestrating AI effectively Developers now need to understand the entire spectrum of AI-assisted workflows. There are moments when vibe coding is perfectly acceptable, particularly when exploring ideas or building quick prototypes. There are other situations where detailed structured instructions are necessary to keep a system coherent. And in large or complex systems, full spec-driven development may still be the safest approach. The key is knowing when each mode is appropriate. when each mode is appropriate As AI coding agents continue to improve, the developers who thrive will not necessarily be the ones who type code the fastest. Instead, they will be the ones who can define problems clearly, design effective specifications, and guide AI systems toward the correct solution. In the AI era, writing code is becoming easier every day. Writing the right instructions is becoming the real craft. right instructions

This story contains AI-generated text. The author has used AI either for research, to generate outlines, or write the text itself. 

Why “Write Faster Code” Is the Wrong Career Strategy Now

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

AI Code Review Is Great at Nitpicks, Terrible at Systems

How Artificial Neural Networks Can Code Smarter Than GUI Programmer

How GPT Pilot Codes 95% of Your App

The Modern AI Stack to Increase Developer Productivity

When AI Joins the Dev Team

How I Made My Own Screenshot-To-Code Project

AI Code Review Is Great at Nitpicks, Terrible at Systems

How Artificial Neural Networks Can Code Smarter Than GUI Programmer

How GPT Pilot Codes 95% of Your App

The Modern AI Stack to Increase Developer Productivity

When AI Joins the Dev Team

How I Made My Own Screenshot-To-Code Project

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps