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 · 24 Sep 2026, 20:49Someone going quiet is the easiest thing in the world to miss, because quiet looks exactly like busy. Everyone is busy. Nobody thinks twice about a mate who has not replied in three weeks, and that is how a bad stretch stays invisible.

The check that works is not a text asking if they are alright and waiting for a yes. It is something easier to answer than a question. What I am doing at the weekend and whether they fancy it. A call where I do most of the talking first. No agenda, no need to reply today. That last one gets more honest answers out of people than any question I have ever asked.

And if you are the one who has gone quiet, it is not a character flaw and it is not you falling behind. Everybody assumes the rest of the room is coping better than they are, and it is almost never true. You do not need the whole story worked out before you tell somebody the first bit of it.

It's okay not to be okay. A message is enough if a call is too much today, and if it is heavier than a mate can hold, that is what your GP and the helplines are there for. Nobody minds being asked.00105Lee Hayward@vibedcoder · 24 Sep 2026, 17:08Watched my agent spend ten minutes writing documentation for a function it deleted three steps later. Thorough, accurate, and describing something that no longer exists. I have left it in. 😂002024Lee Hayward@vibedcoder · 24 Sep 2026, 15:30A model that makes typed decisions instead of writing text, at $0.042 per million input tokens. I was the customer for this until I priced it against what my agent already runs on.

Jev does not write code, hold a conversation or call a tool. TypeSafe's own docs say it is not a replacement for the model behind your coding agent, so the 444x cheaper line on their home page was never a claim about your agent bill. It is about a different kind of call.

So I costed it on the job I would actually give it: a decision inside a prompt that is already cached. Their own worked example ranks a list of agent skills, about 40k tokens a run. On Jev that is around $0.0017 a run. The same list, sitting in the cached part of my prompt, comes in near $0.0002. Cheaper than a fresh call, several times dearer than the call it would replace.

And I cannot rerun their numbers. Their own note says the benchmark workflows were made by people on their model capabilities team and that the 444x sits at the higher end of real world gains. No architecture, no weights, no paper.

None of that makes it useless. If you have a high volume of typed decisions in your own app and no cached prompt behind them, it may well pay. Mine is not that shape, and the number in the pitch is not the number on my bill.412084Lee Hayward@vibedcoder · 24 Sep 2026, 12:58The best review step I have costs nothing and takes a minute. It is a second session that never saw the code being written.

Ask an agent about the change it just made and you get agreement. Its own reasoning is still sitting in the context, so it defends what it wrote, and it is very good at sounding sure while it does it.

So the diff goes into a fresh session with the requirement and nothing else. No history, no explanation of the choices I made, no apology for the mess. Then I ask three things: what would this break, which line do you not believe, and what is the smallest input that gets it wrong.

Same model, different hat, and the answers are worth reading. A reviewer who watched the code being written is not a reviewer. That goes for me too.

Do you re-run a change past a clean session, or trust the one that wrote it?001121Lee Hayward@vibedcoder · 24 Sep 2026, 10:06An agent scouted an easy docs bug in an open source project, wrote the patch, and a human put his name on it and submitted it as bounty work. The maintainer merged it and then said the quiet part: what made it acceptable was that a human had actually tested it.

The problem is not the code, and the patch was a legitimate fix. The cost of not noticing lands on somebody who never asked for the work. An agent can open a pull request all day. A maintainer still has to read it, and the reading is the expensive part.

Someone in the thread called for the account to be banned, which is not only about the AI. Drive-by pull requests existed long before agents, from credit seekers and bounty hunters, and maintainers were tired of them already. An agent just makes a drive-by cheap enough to run at volume.

Every argument I have had about AI workflows comes back to the same gate, and it is not the quality of the output. It is whether somebody who understands the change is accountable for it. That is the difference between a contribution and spam, and it does not care who wrote the diff.

Human attention is the scarce resource. It always was.

github.com1020148