Lee Hayward

./about Who's behind the builds

I'm a UK based vibe coder. AI drafts, I proof, I ship, and I hold code and copy to the same standard. I riff on domain ideas and build side projects like DomainRiff and LookMa from my own homelab.

./projects Featured builds

./stack How it gets built

./hardware The rigs behind the site

./homelab The machines that never sleep

./latest Latest on X

Lee Hayward@vibedcoder · 14 Sep 2026, 15:55When my agent gets the same bug wrong twice, I stop it and go back to the last good commit. No third attempt.

The spiral goes like this. It fixes the symptom, that breaks something next door, so it patches that, then patches the patch. Ten minutes later there are four fresh bugs and none of them are the one I started with.

Reverting costs me an afternoon of work. Letting it cook costs me a codebase I no longer trust, because I have stopped being able to tell which edits were deliberate.

Two strikes, then reset, and put what went wrong the first time into the next prompt.000019Lee Hayward@vibedcoder · 14 Sep 2026, 13:38My agent apologised to me this morning. "You're absolutely right, I should have caught that earlier."

I hadn't said a word. I was still typing the question. 😂003024Lee Hayward@vibedcoder · 14 Sep 2026, 11:05A blog post doing the rounds this week makes the case for never letting an LLM near your code. The author wants programs that are correct by construction, and he reckons a model separates him from the low level detail that actually matters. Fair enough, that's his call.

The bit I disagree with is the label. If you use AI you get filed under something lesser, he says the word engineer is being claimed by people who mostly manage, and everybody ends up arguing about titles instead of code.

I read every diff my agent produces. Line by line, with the file open beside it. That's where the understanding comes from, not from typing the boilerplate myself. I know exactly where the permission check lives and why it's there. I have never once typed an import by hand and felt the craft deepen.

What I want out of code isn't provable perfection. It's that it still works in month three.100017Lee Hayward@vibedcoder · 14 Sep 2026, 10:34Everyone is building agentic Kanban boards. I went the other way.

A plain todo list, a daily email, and an AI that does the parts I approve. My list is 7 lines. Here is why that beats a board for me.

A board is for work that moves between stages and people. Mine does not move. Every line is either a decision I have not made yet, or a thing only I can do. Columns would sit empty, and empty columns are just decoration I have to maintain.

Kanban answers "where is this work". I kept asking "what do I actually decide today". Different questions. A card sitting in Doing for three weeks tells me nothing except that I moved it there on a Tuesday.

So the list is one file. Every line carries a rating, who it is for, what it is, why it matters, and what happens if I say yes. It emails itself to me at 09:00 with a calendar file attached. Anything urgent pings my phone.

Every line that needs my yes or no gets two buttons: Do it and Ignore. Clicking is the answer. The agent picks it up within minutes, does its half, and hands me anything I have to run myself. Ignore means it never mentions it again.

The real change was the rule for adding an item. If I would ignore it and nothing breaks, it does not go on the list. Seven honest lines I actually answer beat forty lines I scroll past. The tool never made me faster. The filter did.

Screenshot of the list below. It is not a product and there is nothing to sign up to. It is one file, one cron job, and a box in my hallway. The buttons are live if you want to poke at them: hayward.click002056Lee Hayward@vibedcoder · 14 Sep 2026, 09:10SubmitHub, the site bands use to pitch songs to curators, now rejects any submission its AI detector scores 85 percent or higher as likely AI. The reason isn't ideology. Spotify started burying AI Persona profiles out of recommendations in August, so selling promotion to an AI act was selling a lottery ticket.

The numbers behind the cutoff are the interesting part. About 11 percent of tracks uploaded last month scored 85 percent or higher, and curators approved those 16 percent of the time against 31 percent for everything else. They were already passing on it before anyone wrote a policy.

Nearly one in three AI creators hide that they used it. That hiding is the bit I can't get past. If a tool wrote your track, say so and let the listener decide.

Would you rather platforms flag AI music, or leave listeners to work it out?

submithub.com001057