AI + Code: New Era

The conductor era: Why the future programmer won't write code

I was interested in AI before the whole boom, even before I started programming, so topics related to it are particularly close to me.

When the first ChatGPT appeared, I was genuinely impressed - this wasn't the old Cleverbot, but a tool that could actually be useful and revolutionary. True intelligence won't emerge from transformers, I'm certain of that, but as an advanced tool it opens a lot of doors. I never thought AI would be capable of displacing developers though.

A free business recipe (for now)

I've been testing Claude Code on side projects - stuff I build "for the drawer" - and I have to admit, it's an excellent tool. It significantly speeds up the work. It's not yet ready for truly complex projects, or commercial ones where far more context is needed, but I can say this:

A mid-level developer could already start going to companies and offering internal process automation - not even for a regular salary, but for a cut of the costs saved.

Right now, a decent developer using this tool well can build finished products that meaningfully reduce company costs, sometimes significantly.

Write this down as a hustle method. If you're a good salesperson, this is genuinely a solid business idea.

Less writing

I think developers will gradually stop being musicians and start being conductors, it will significantly multiply the output of a single developer. And I genuinely feel for juniors, who in my opinion are already somewhat redundant. The first available AI model can complete a junior's tasks in a matter of moments.

Why AI will never fully replace developers

It's not really about complexity. At some point AI will get to a level where it can handle even the most intricate projects. But one question will always remain: responsibility. Who is accountable for that code? A developer. One instead of a whole team - but there will always be room for that person.

Context problem, time limits

There are already tools like the Ralph Loop that build applications by chaining Claude Code tasks, which work autonomously across a long session - while the human is out for a walk. Someone ran 27 separate tasks over 4.5 hours and came back to a complete, working application.

Code written by a machine, used by humans

This week, Boris - one of the creators of Claude Code at Anthropic - released a new tool called Cork. Think of it as an application-layer version of Claude Code, a more packaged experience for broader use. Around the same time, someone asked Boris on X what percentage of the code in that project had been written by Claude.

The answer: all of it.

I'm not particularly surprised. You don't have to literally write code anymore - but you do need to know how to write it, and what it should look like. "All of it" is a slick marketing shortcut. AI did write every line, sure - but that's a bit like saying a voice-to-text app wrote my application because I dictated it instead of typing. Technically true.

Claude code is not just for coding

One of the biggest misconceptions is that the name "Claude Code" implies it's a developer-only tool. It isn't.

It's very good at writing, debugging, and architecting software. But it also handles files, APIs, documentation, and structured workflows pretty well. Marketing teams can use it. Business analysts can use it. Non-technical founders can use it to build internal tools they'd previously have needed a developer for.

Small internal utilities that previously needed to be scoped, estimated, and handed to a dev for a sprint - now buildable by someone with minimal technical background and a clear goal. A transcription tool, a reporting script, an internal dashboard. Without hiring, without waiting, etc.

Teams that figure this out early are going to have a real advantage.

The trust problem

Claude Code and tools like it solve a lot of problems, but one always remains: hallucinations. Claude can write an application in 5 hours that you can actually use - but are you sure everything in it works the way it should?

Some people solve this by throwing another model at it. Use Gemini as a verifier.

The workflow looks something like this: Claude Code gives you a solution or architectural suggestion. You take that output, paste it into Gemini (or another capable model), and prompt it as if you received advice from a developer colleague - "a friend wrote this for me, but I'm not 100% sure they got it right. Can you review it, spot any errors or gaps?"

If both models agree, your confidence goes up. If they diverge, you know to dig deeper. It's not bulletproof - nothing with current AI tools is - but it's a useful sanity check.

AI and the job market

The developer job market is splitting.

One segment - likely a significant portion of junior and mid-level roles focused on routine implementation work - is genuinely at risk. If your primary value is translating requirements into code, that's exactly what AI does well.

The other segment - architects, engineers who understand systems deeply, people who can direct AI effectively and catch its mistakes - will likely see their leverage increase dramatically. One senior developer with strong AI orchestration skills can now produce what previously required a small team.

"Directing AI effectively" requires genuine depth. You can't direct an AI through a complex codebase if you don't understand what it's doing. The conductor still needs to know music theory.

The most dangerous position right now is a junior developer who decides AI tools make learning fundamentals unnecessary. You need the foundations - not to write boilerplate manually, but to evaluate whether the AI's output is correct, where its reasoning breaks down, and when to override it.

What to actually focus on

I don't think the takeaway here is to be scared. Learn to conduct.

The skills that actually matter: understanding systems well enough to direct AI at the right level of abstraction, knowing when the output is wrong even when it looks right, breaking goals into tasks an agent can execute reliably, and cross-checking with multiple models instead of trusting one blindly.

The developer role isn't going away. It's changing faster than most people expected.

I'm actively experimenting with Claude Code on real projects - this portfolio site included.