Yesterday I opened my project and just stared at it.
Not in a dramatic way. More like: I opened Zed, looked at the file tree, and immediately wanted to close it.
On the left it looked like a real backend already: auth.go, tickets.go, invitations.go, users.go, migrations, models, storage, configs. The kind of tree that makes you think: okay, something serious is happening here.
Then you open the files.
Half of the handlers are stubs.
Some things are wired. Some things are half-wired. Some names already feel too permanent. And in my head there is this stupid list growing by itself:
mobile API later, admin panel, roles, tenants, deployment, maybe Kubernetes, maybe not, what if we need events, what if the permissions model becomes ugly, what if I build this wrong from the start.
The feature still doesn’t exist.
Nothing works end to end.
But the project already feels heavy.
That specific moment kills me more than debugging does.
Debugging at least means something exists.
I let Codex write the first real piece
Today I didn’t try to be heroic.
I gave Codex a small task: implement auth and invitations.
Not the whole project. Not the perfect architecture. Not “build the startup”. Just one piece I could actually run.
Before that I gave it rules. Handlers stay thin. Business logic does not live in HTTP handlers. Use the existing storage.go pattern. Don’t invent a second way to talk to the database just because it is convenient right now.
It generated a large commit. Around 1200 lines.
Auth, invites, some Podman setup, a bunch of files.
And this is exactly where AI coding can become dangerous, because the temptation is strong: big diff, code compiles, endpoint responds, nice, accept everything.
I didn’t.
I read the diff.
Pretty quickly I found raw SQL inside a handler.
The endpoint worked. That’s the annoying part. It wasn’t broken code. It was worse: working code in the wrong place.
I already had a storage layer. I don’t want a handler to suddenly become aware of database details just because the agent took the shortest path.
So I moved it.
Cleaned it up.
Made it fit the project instead of letting the project fit the generated patch.
That was the first moment where I thought: okay, this is not the same thing as being an AI operator.
An operator would stop at “it works”.
I didn’t want to.
Then I opened Apidog. Register worked. The response had a user, tenant, institution, access token. On the right side Apidog complained that it expected 201 but got 200.
Stupid little detail.
But I felt good.
Not because Codex wrote 1200 lines.
Because I sent a request and something real answered.
That is the part I missed.
gofro was a hint, but I misunderstood it
After one coursework project, I built gofro — a CLI that generates Go projects. I already wrote about it here: I got tired of setting up Go projects from scratch, so I built a scaffolding CLI.
The reason was simple: I was tired of starting every Go backend with the same boring ritual.
Configs. Dockerfile. Compose. Postgres. Redis. Migrations. Graceful shutdown. Database connection. Environment variables. All of that before touching the actual feature.
At some point I thought: maybe scaffolding is the thing I hate.
So I built a scaffolder.
And to be honest, gofro itself was mostly vibe-coded with Claude Code. I’m not going to pretend I hand-crafted every line in a cave with Vim and discipline. I wanted to remove a boring part of my workflow, and Claude Code was very good for that at the time.
It helped.
But it didn’t fix the whole problem.
Because you can generate a nice skeleton in seconds and still sit there thinking:
okay, now what?
Folders exist.
Compose exists.
Migrations exist.
But there is still nothing to call.
No endpoint that proves the project is alive.
That was the part I missed with gofro. I thought I hated setting up projects. I did. But what I hated even more was spending hours before the first visible result.
Scaffolding is not the same as a working flow.
Annoyingly obvious now. Was not obvious to me then.
Coursework was the opposite feeling
The funny thing is that my coursework had already shown me what kind of work actually motivates me.
I worked on it with a friend. And I remember being genuinely happy after pushes.
Not “good, task done”. More like I wrote a few endpoints, ran the whole flow, found a bug, fixed it, opened an MR/PR, and felt like I had built something.
The strongest memory is geocoding.
Coordinates finally processed correctly. A route got built. Data started going through WebSocket.
I was sitting there happy like an idiot.
Not because it was some genius engineering. It was just alive.
Code → run → result.
Very simple.
And in the same project there was the other part: configs, migrations, wiring, setup, all the stuff around the feature.
That part felt like mud.
You work for hours and the only thing you can say afterwards is: “I prepared the foundation.”
Cool.
I hate that sentence.
I want to say: “you can run this now.”
I also overdid system design
Another problem: at some point I went too deep into system design.
I thought my weakness was not coding. I thought my weakness was designing things properly.
So I watched lectures, went to meetups, attended UWDC, listened to a talk about EDA and C4 on a real application. Books didn’t happen. I tried, got bored, lost.
But the mindset stuck.
Sometimes too much.
The project doesn’t have an MVP yet, and I’m already thinking about exposing the API for a future mobile app, admin panel options, tenant isolation, roles, deployment, Kubernetes, events, module boundaries.
It feels productive because technically you are thinking about real problems.
Just not real yet.
A friend told me once: don’t optimize what doesn’t exist.
I nodded. Very smart. Very reasonable.
Then I continued doing it anyway.
And it wasn’t even premature optimization.
It was premature architecture.
A nice respectable way to avoid building the ugly first version.
That’s probably the part I needed to admit.
I wasn’t blocked because I didn’t know enough architecture.
Sometimes I was blocked because I tried to apply architecture before there was anything worth architecting.
The AI panic was mixed with job panic
Vibe coding annoyed me for a long time.
Not the tools exactly. The culture around it.
That whole internet vibe of:
I don’t know programming, but Claude Code shipped my fifth SaaS
When you’re trying to enter the market as a junior, this doesn’t sound inspiring.
It sounds like someone removed the door before you reached it.
Add fewer junior vacancies, less remote work, endless posts about AI replacing entry-level developers, and suddenly the thought is not “the market is rough”.
The thought becomes: maybe I’m useless.
That is not a great mental place to evaluate tools from.
I started hating AI partly because it was standing next to that fear.
But using AI and becoming a person who understands nothing are not the same thing.
I had already used Claude Code a lot. gofro came from that period. Back then Claude in Zed’s Agent Panel felt close to pair programming: you could see what it was touching, what it was writing, where it was going. The limits were okay. Reviews didn’t destroy the whole window. Hallucinations felt less painful.
Later it got worse for me. Limits started to hurt more. Large reviews became expensive. Hallucinations became more noticeable. In my bubble people started mentioning Kimi, Codex, other tools.
I tried Kimi Code, hit a timeout, got annoyed, bought ChatGPT Plus, and tried Codex.
Codex is also inside Zed for me, but the feeling is different.
Claude felt like watching someone code next to me.
Codex feels more like sending work to another room and waiting for a batch of changes.
Less live feedback. Slower feeling.
But the limits for $19 felt almost illegal after what I was used to.
So now I’m less interested in “which agent is morally pure” or whatever.
The real question is: does this tool help me reach a working state faster without making me lose ownership of the code?
For this auth flow, yes.
“It works” is not enough
The Codex commit made me calmer about AI coding, but not more trusting.
Actually the opposite.
If AI writes code, I need to review harder.
Auth? Check password hashing, token expiration, middleware, errors.
Invites? Check if an invite can be accepted twice. Check tenant boundaries. Check roles. Check institution links.
Migrations? Check constraints, indexes, nullable fields.
SQL in handlers? No.
A 200 OK response is not proof that the code belongs in the project.
It only proves that one path responded once.
That’s not enough.
My current rule is simple, I guess:
I’m fine with AI writing the first version.
I’m not fine with not understanding the final one.
If code stays in the repository, I need to be able to explain it.
Why this endpoint exists. Why this logic is not in a handler. Why this table is related that way. Why this error is returned. Why this can be deployed without embarrassing me.
If I can’t explain it, it’s not really mine.
Even if my name is on the commit.
Not a manifesto
I still want a job.
Pet projects, startups, freelance — all useful, sure. But a real job gives you other people’s code, reviews, processes, responsibility, commercial experience. Hard to fully replace that alone.
I also still don’t want to become a prompt operator.
But I don’t want to keep proving that I’m a real programmer by manually suffering through every repetitive stub either.
If AI gets me faster to the point where I can send a request and see a real response, fine.
After that, the actual work starts.
Review the diff.
Fix the weird parts.
Move SQL out of handlers.
Write tests.
Clean up.
Refactor.
Deploy.
Own what remains.
Maybe vibe coding is not the thing I hated.
Maybe I hated the part before the project becomes alive.
And maybe the real failure mode is not “AI wrote the first draft”.
Maybe it’s keeping code you don’t understand just because it worked once.
Anyway.
That’s the current working theory.
It took being sick of a file tree and reviewing 1200 Codex-generated lines to get there.
Top comments (0)