AI Coding Assistants Are the New Compilers
I’ve been using AI coding assistants a lot lately. I can confirm that Claude Code is amazing and probably worth the hefty price of the Claude Max subscription. But my main insight is that AI coding assistants are fundamentally changing how we build software. It’s not just about writing code faster. We're witnessing a shift where the AI acts less like a smart autocomplete and more like a modern compiler, turning our ideas directly into applications.
This isn't a new trend, but an acceleration. GitHub Copilot was a revolution for speeding up small, single-file tasks. Then came tools like v0.dev, which could scaffold applications made up of several files. They are great for getting a simple project off the ground but struggle as complexity grows, often adding messy code that creates technical debt. The latest generation of AI tools, like Claude Code, goes a step further. They can navigate medium-sized codebases and even contribute to large, well-structured projects. This leap is what got me thinking: the very nature of our work is changing.
English is the New Programming Language
In essence, these new AI tools translate natural language into a working application This makes a simple English sentence a command that a machine can execute. It’s like a universal programming language. I can describe a web app, a mobile app, or a data analysis script, and the AI can build it.
I don't need to be an expert in TypeScript to bootstrap a web app, or in Swift to bootstrap an iOS app. I can just prompt what I want, see the result, and iterate with more prompts. The barrier to entry for creation is dropping dramatically.
This new universal programming language doesn’t need versioning, documentation, or linter. For the most part, we already know it. And it’s permissive : if I need to develop something that I can’t precisely describe, I can use an approximate description, and refine from the result.
And even if English isn’t my mother language, I can use another language like French and it works just as well.
LLMs are the New Compilers
The TypeScript or Swift code generated by AI tools is like an intermediate representation of my original intent. Most of the time, I don't even look at the generated code, just like I never inspect the assembly code from a C compiler. Why would I? I only dive in when a feature doesn't work exactly as I want. And even then, my first instinct isn't to debug the code myself. It's to write a better prompt and let the AI fix its own work.
I’m starting to treat the AI as a compiler for my thoughts. I write instructions in plain English (the "source code"), and the AI turns them into executable JavaScript, Go, Python, or SQL (the "machine code"). This code doesn’t need to be well-written because most of the time, no human will ever look at it. It’s really only intended for the computer.
This is why these new tools don't need to be IDE plugins. An IDE is a tool for writing code. Claude Code is a tool for writing applications. All I need is a text box.
Business Experts are the New Developers
A direct consequence of this shift is that you don't always need a programmer to build an application. The most important skill is becoming the ability to clearly describe a business need and its expected outcome. This is traditionally the role of a business analyst.
We already saw business analysts gain autonomy with no-code tools. AI coding assistants are the next level, empowering them to build new features and entire apps, often without a developer in the loop.
Slow is the New Fast
Using these AI tools involves a lot of waiting. The model spends time reading the existing code, thinking, and checking its work. In my experience, I spend about half my time just waiting for it to finish.
It feels a lot like the old days of compiling large projects. You kick off the build, grab a coffee, and hope for the best.
But even with the waiting, my productivity has skyrocketed. I finish tasks faster, maybe because I'm less bogged down by the details of the code. And while the AI is "compiling" one feature, I can spin up another instance to work on a different task in parallel. It’s a great way to multiply my output, though it does demand more context switching.
Developers Are Still Needed
If you're a developer, this may frighten you. But in my experience, Claude Code still needs a qualified programmer in the driver's seat. The developer's role is just shifting from writing to directing.
Sometimes, the AI needs a human to verify its work. For instance, I was building a feature to transcribe audio recorded in the browser. The AI tried to test it with a blank audio file and correctly realized it needed me to provide a real recording.
Other times, the AI gets stuck. It might try and fail to fix a bug several times because it's working with outdated knowledge of an API. This happens about 10% of the time, and it requires a human to step in, diagnose the problem, and write the fix.
Most importantly, a developer’s expert eye is crucial for maintaining quality. AI can write messy, unstructured code that quickly becomes technical debt. I often have to tell Claude to refactor a file it just wrote. It does a great job, but it needs the instruction. An expert is still needed to manage the overall architecture and ensure the project stays healthy over time.
This directing role also involves codifying my expertise to prevent rework. Claude Code supports memory (you can add permanent instructions or context for a particular project), so the right way to fix a recurrent problem is to use Claude.md. It’s like encoding my own patterns and preferences, teaching the AI to handle those tasks autonomously in the future.
Maybe this will change in the future, but for now, you still need a developer to maintain a large application over time.
The Big Hurdles Remaining
This new world isn't perfect. We still need to solve major issues around reliability, privacy, security, price, and the environmental cost of these massive models. The time I earn by leveraging Claude Code translates into thousands of tons of CO2 emissions, and aggravate global warming. This is the main obstacle in my conversion to these new tools.
And for the "AI as a compiler" analogy to hold true, the biggest problem is reproducibility. A traditional compiler produces the exact same output from the same source code every single time. An AI does not. You can give it the same prompt twice and get two different results. This lack of determinism is a major hurdle for building reliable, production-grade systems.
What's Next for Developers?
If AI is the new compiler and business analysts can act as programmers, what does that mean for us, the developers?
Our value is shifting up the stack. We're moving from being bricklayers to architects. Our job is less about the syntax of a for-loop and more about designing robust systems, making smart architectural decisions, and guiding AI to produce high-quality, maintainable results. The craft of software development isn't disappearing; it's evolving.
AI coding tools are poised to trigger a Cambrian explosion of digital services. It's becoming significantly easier to build and maintain such services, and the potential productivity gains from digitalization across every sector of our economy are immense.
Which effect will dominate? Will AI reduce the need for developers because building apps becomes easier, or will it increase the demand for developers because we'll simply want more apps? It's hard to say for sure, but one thing is certain: we won't do the same job in five years as we do today.
