Comment on Using AI to write better code more slowly by Dmitry Jum

This is exactly how I work, thank you for this article. I've been working solo for myself lately and it's hard t compare myself to, although I'm trying to watch and read of what's happening out there as much as I can.
But I see they yell from out of every corner: AI generates slop! Senior engineers will have to fix all that! Or: "I build 10 features in parallel in 1 hour and ship an MVP in a weekend that makes me 50k MRR"

I generally discuss a feature idea, then I break it in a couple of chats to discuss an architecture and user flow or UI. Then another chat plans out a list of tasks based on the conclusion of the previous 2-3 conversations. We come up with a style guide and UI design prompts and I create mockups with Stitch and GPT images.
Then I go task by task with Codex or Claude, or both. And when Codex builds a task, I review it and leave a bunch of comments and ask how things work if I don't understand them when I read it and ask why. I have Codex Github reviewer review. I ask Claude Code review the changes as well and if I have a Coderabbit subscription, I have that service review the code as well. And then I have Codex or Claude comment or address that dozen of comments (or more) that have accumulated from 2-3 AI reviewers and me. It may take a few iterations to achieve the final clean product out of the task we've been working on. But by the end of it multiple edge cases are revealed and covered and I actually understand and agree with 90% of AI written code and can describe it and advocate for it to another person who reads it or asks me about.

My imposter syndrome bashes me for still being slow, but I've released a couple of application products to a couple of clients lately that have't had a single bug (granted they aren't that big, but still) for months since their release (and not yet) with my approach.

I've been hearing that junior and mid level devs have AI build and AI review the code and they ship without even looking or reading it and deliver exponentially more and faster than me. But if that had been the right approach, I probably wouldn't have been correcting and leading the state of the art models to different patterns or solutions, or correcting architectural flaws, that would only reveal themselves at scale down the road.

I love working with AI, the development is still tedious, slow and difficult, but I'm building bigger and more complex, more reliable and better looking things on my own than ever before.