AI Coding: Why It Feels So F***ing Hard (At First)

You think you’re a good coder. So now that there’s AI... surely, you should be A-MA-ZING, right?

Right? ...right?

And then you actually try it. Maybe on a side project, coding an hour here, two hours there after work.

At the beginning? Pure magic.

It’s just so fast — generating functions, churning out boilerplate. You're euphoric.

But then... Then, it starts to hurt. Badly.

If that’s you, well, you’re not alone{style="color: red;"} . I’ve been through that phase, and honestly, I'm still learning, still getting roughed up.

The Great Skill Inversion: Welcome to the Reading Gauntlet

My first big realization? The whole game has changed.

It's a meta-game shift.

The micro skills that were useful before are now less useful. And vice versa.

For example, let's talk about typing speed?

All those years spent mastering Vim, optimizing myself to type code at lightning speed?

This suddenly feels… less useful.

You know what is useful now?

Reading shit tons of code. Fast.

Because these machines, they write. And write. And write.

And suddenly you're drowning in piles and piles of code you didn't create, often structureless, sometimes pure garbage. You have to wade through it, understand it, verify it. The burden shifts from creation to curation and validation.

And it's exhausting.

Thousands of lines of broken code per hour. Variations of variations of the same garbage pattern.

It’s exhausting. So exhausting that eventually, maybe you stop reading closely.

You start just… accepting changes. Clicking 'yes' like a zombie, hoping for the best.

Losing Your Mind (and Your Codebase)

That zombie state might get you a functional script or a small app technically beyond your level. But on projects you care about, projects you need to maintain? It falls apart.

What really happens is your brain unplugs. The deep mental model you instinctively build, that intricate map of the codebase you hold in your head – the one that lets you navigate complexity and debug effectively? It vanishes. Code becomes fuzzy, intangible.

You open the same file you looked at two minutes ago, and the AI has ripped it apart and stitched it back together differently. It’s like trying to navigate an ever-changing maze. How can you reason about something that won't stay still?

Why would you even try?

AI's Alien Logic and Tooling Hell

You attempt a refactor with Claude Sonnet MAX inside Cursor. It chews through prompts, takes its sweet time, and costs you a euro… for something that isn't demonstrably better, maybe even adds hidden complexity.

The thing is that when you try to operate on the full codebase, it tries to read slices of the file instead of the full text.

You try Claude 3.7 on the website, maybe with that fancy GitHub integration. It is amazing pulling context like magic without need to copy-paste files.

But, 3.7 has a little defect. It feels like collaborating with a utterly mad mad mad scientist – throwing complex, smart-sounding overengineered solutions at you that might just explode later and fry your brain right now.

You find yourself pleading with it: "Hey Claudy, this seems overly complex, can you make it simpler?"{style="color: red;"} (...wait 2 minutes) "Simpler?" (...wait another minute) "Even simpler?!" ... "PLEASE"

So, you heard Gemini 2.5 is supposed to be better at code. But in cursor in agent mode, it's just stuck and doesn't do anything.

So maybe copy-paste everything into Gemini 2.5 chat website directly instead. But now you have to use something like 16xprompt if you don't want to spend 5 minutes copy pasting files, because no fancy github integration here.

And yes, Gemini 2.5 can handle longer context and seems smarter, but it will make some other mistakes... You still need to debug, in a different way, again.

"No free lunch" they say.

FUCK. That sucks, you know?

Coding used to feel like flow , but I’m not in flow with the machine. The machine get's me out of flow. 🥲

The Future is Coming (Ready or Not)

Yeah, these machines will get better. We see the benchmarks, the predictions (like METR's time horizon suggesting agents capable of handling an afternoon's worth of work in a near future). That future is barreling down on us.

So, I need to adapt.

You thought you were a good coder before? You think you’ll just be a good coder after AI, without changing anything?

You haven't really tried yet, have you? You haven't truly wrestled with coding with AI, day in, day out.

Are you afraid to try?

Because I was.

The Tech Debt Time Bomb

But let’s be brutally honest. AI isn't a magic pill.

What happens when somebody uses AI naively on your team?

Who cleans up the mess when they check in mountains of unvetted, poorly understood, AI-generated spaghetti?

Not them. If they lacked the context or skill to guide the AI properly in the first place, they certainly can't sort out the resulting tangle.

It’s going to be you. The experienced dev who can see the flaws, who understands the architecture, who holds the memory of the system. That AI-generated technical debt? It lands squarely on your shoulders. Unavoidable.

And when you're vibe-coding in zombie mode, who's going to be the one to clean it up? Well, you, my dear. A very tired and demoralized version of yourself.

Get Back to Getting Good (The New Way)

Honestly, sometimes I look at the code that's been generated and my first instinct is to rm -rf the whole thing and start over manually. How long before an AI-assisted codebase becomes an unmaintainable nightmare?

It's a new world.

The tools have changed. The meta game has shifted. It’s scary, it’s messy, but ignoring it isn’t the answer. It’s time to be open-minded, to try shit out, to figure out the new rules.

Those durable skills – critical thinking, system design, rigorous testing, the ability to evaluate code quality, functional programming, blablablabla Those might be even more important now.

Maybe? Maybe not.

The exact playbook is still being written, but relying only on the old skills feels like a losing bet.

I liked typescript because it used to catch errors for me when I code. I could just fix type errors and the code would work. It was bliss.

(I have ADHD so something that catches inattention mistakes is a blessing, honestly.)

But when AI codes, it creates code that typechecks but still freaking bugs 🥲, and this shit is not ADHD friendly.

It's like an generator of vicious inattention mistakes that have become resistant to typechecker antibiotics.

So I need to evolve, to adapt.

I need more coding hygiene, more tests, more checks? A better git workflow, should I just encapsulate the AI + zombie mode version of myself in pull requests? What's the flow?

It's a new sport season, a great equalizer. We were good before, so we might have a head start, but cards have been shuffled. Time to get back to getting good.

End of the rant, I'm going to bed. Have a good day my friend.

Rémi

PS: Anxious about what AI means for your coding experience? I'm sure I can help you with that.

Let's grab a coffee in virtual here https://cal.com/remi-connesson/free-call and I'll answer any questions you have.

PS2: Excited? Let's chat and brainstorm this new world together my friend.