I rewrote our app with vibe coding. I still can't read the code.

Six weeks ago, neither of us could reliably get a build running. The new Slax Note shipped this week. I still can’t read the code.

I’m a PM. My teammate is a designer. Our boss told the two of us to rewrite Slax Note — our voice-first notes app — from scratch. Using AI.

This is what happened, in order. Disbelief, rage, the night I couldn’t feel anything, and the Claude Code agent we ended up calling Rocky.

”Wait, what did I just do?”

Two coworkers chatting at their desks. Engineer: "What did you just do?" PM: "I don't know, but it actually ran successfully."

One afternoon I was sitting next to one of our actual engineers. Something I’d just typed into Rocky had built. Compiled. Run. The feature worked.

He looked over. “What did you just do?”

“No idea,” I said. “But it ran.”

That feeling — being happy for no reason you can explain — became the recurring emotion of the next month. I’d ship something, look at it, and have no clue how it got there. Confusing. Also kind of wonderful.

It started with panic

Boss tells two non-developers they'll be in charge of building the app. The designer thinks: "I trust AI… but I don't even trust myself yet."

Rewind to early March. My feed was nothing but AI coding takes. Vibe coding. Cursor. Claude Code. “Anyone can ship now.” My designer teammate was getting the same firehose. We’d send each other links and quietly panic.

The unspoken question: what’s left for people like us?

Then our boss walked over and made it concrete. “You two. Rewrite Slax Note. Use AI.”

Part of me wanted to refuse. The other part had already opened a new VS Code window. Panic gets you to the keyboard pretty fast. (It’s worse at the question of when to leave it. We’ll get there.)

Most of the bugs were me

Four panels: environment errors, opening the wrong folder, debugging a white screen with AI for two hours, and the senior dev finds the bug in seconds.

For the first week, every time something didn’t work, my reaction was: the AI is broken.

Then I started keeping score. Roughly half my problems had nothing to do with the model:

I configured my environment wrong.
My VPN dropped mid-request.
I told Rocky to edit a file in the wrong folder. Twice.
I let Rocky debug a white screen with me for two hours, only for a senior engineer to glance at my laptop and say, “You set a breakpoint here. That’s why.”

Half a day, gone. The model had done its job. I just didn’t know how to drive the things around it — the editor, the terminal, where files lived. Embarrassing, but kind of a relief: that part I could learn.

I yelled at it for a week

Left: rage phase, yelling at the AI. Right: providing context, giving clues, showing patience. Bottom: collaborating like teammates, then realizing — "Why did I forget how to talk when I was talking to AI?"

I got back from a work trip and my designer teammate had built most of the app. Then we hit the part nobody warns you about: bugs.

The bugs were the worst part. I handled them badly at first. Paste an error, get a wrong answer, yell at the screen. “You’re useless. Are you even capable of this?” Pure rage. Tokens spent. Nothing solved.

Eventually I noticed something embarrassing. I was giving Rocky worse instructions than I’d ever give a human teammate.

So I switched:

Give context. What the feature is for. What I already tried. What the logs say.
Give clues. “Check this endpoint. It returns 500 when the user is logged out.” Not “fix it.”
Be patient. If it’s stuck, feed it the docs. It hasn’t read the codebase. You have.

First time I did it properly, the bug got fixed in one round. I sat there and thought: why did I forget how to talk, just because the teammate was an AI?

I still don’t have a clean answer to that.

I couldn’t put it down

There’s no comic for this section. By the time I noticed the problem, I wasn’t laughing.

I’d stopped going home on time. I’d stopped putting the laptop down at dinner. I’d be in bed at 11:30pm thinking, “if I just kick off one more run, it’ll be done by morning.” Every few minutes: accept this diff? Reject? Re-prompt? Roll back?

A feature used to be ten decisions in a day. Now it was a hundred in an hour. I’d lose the thread of a sentence I was speaking out loud.

One Tuesday around midnight I was in bed with my laptop, watching a green checkmark scroll across the terminal as Rocky finished a run. I was holding my breath for it. For a checkmark. From a robot. My partner had said something to me twice and I hadn’t heard either time.

I closed the laptop and said out loud: wasn’t AI supposed to give me my time back? How did it lock me to this screen instead?

The honest answer was that I never had to wait for anything anymore. Old me would’ve made coffee while a build ran. New me had nothing to wait for. So I just kept going.

Three things I changed

In the end, I changed three habits.

Hours. After dinner, laptop closed. Rocky will be there tomorrow. The bug hasn’t escaped.
Scope. Half of what I’d been building, nobody had asked for. Vibe coding makes adding features cheap, which means adding the wrong features is cheap too. I started killing tickets instead of finishing them.
Don’t turn Rocky loose. You can’t just hand him a task and walk away. You have to scope it, brief him, and come back to check. Slower per step. Faster overall, in my experience.

I still can’t read most of the code in our repo. The features I shipped are in there somewhere, and real users use them. Six weeks ago I couldn’t have said either of those things.

Last night Rocky asked me to confirm a refactor at 10pm. I told him no, it’s late, we’ll do it tomorrow. He said okay. I closed the laptop.

That’s the part I didn’t see coming.

Slax Note is the app we rewrote. It’s a voice-first notes app — you talk, it writes, you pick what shape you want it in.

Part 2 is the technical write-up — the stack, the pipeline, what broke. Coming soon on the Slax blog.