The web app makes 176 requests and downloads 130 megabytes.

It's also <div> soup and largely inaccessible.

These and other issues could be fixed fairly quickly with a little care (if anyone cares).

!!!

Yeah, it turns out those speaker avatar images are 1MB+ PNGs! And there are 170 of them.

What a fantastic cautionary tale about vibe-coding on a mobile phone (where performance analysis tools aren't easily available.)

I just fixed that with Codex - thanks for the tip: https://chatgpt.com/s/cd_6879631d99c48191b1ab7f84dfab8dea

As far as accessibility goes... yeah, the lack of semantic markup is pretty shocking! I'll remember to prompt for that next time I try anything like this.

I just tried it in VoiceOver on iOS and the page was at least navigable - the buttons for the days work - but yeah, I'd be very ashamed to ship something like this if it wasn't a 20 minute demo (and I'm a bit ashamed even given that.)

I'm running "Make open-sauce-2025.html accessible to screenreaders" in Codex now to see what happens.

OK, Codex seemed to do a decent job of fixing up the accessibility, I just shipped its changes: https://github.com/simonw/tools/issues/36

it's concerning these vibe coding tools must be coerced into semantic markup

could that be solved by prefixing every prompt with a reminder?

> it's concerning these vibe coding tools must be coerced into semantic markup

These things are becoming more and more humanlike every day

Definitely. If I had a Claude.md or agents.md or whatever in that repo saying "make mobile friendly sites that use semantic HTML and are accessible for screenreaders" I bet I'd get much better results from them.

would have been fun to see how conference wifi handled that :)

Am I sad for being sad when a post by @simonw gets less than a 100 points?

And this is a pretty good one too. But I guess the mods might want to keep things a bit balanced which is fair enough.

It might be a vibe coding fatigue. Or simonw fatigue. I for one lose any curiosity when I see a vibe coded project/post because I've used LLMs so much that I don't find the generated results impressive or even interesting at all.

They all still sucks big time and this won't change with a better prompt or model, there's no breakthrough and compared to 6 months ago they did not get that much better.

I think both vibe coding fatigue and simonw fatigue are understandable at this point - though I do think this is one of my best posts in a while, so I was hoping it would hit the homepage.

The thing that changed in the past six months is that tool calling in a loop got really good, and a bunch of new tools - like OpenAI Codex and Claude Code - showed up that take full advantage of that.

I don't think my "build a scraper from my phone with a single prompt" example from this post would have worked six months ago - the tooling wasn't there yet.

> I don't think my "build a scraper from my phone with a single prompt" example from this post would have worked six months ago

Even if that is true (which I disagree), isn't it telling that only yourself are impressed by it?

What LLM tool would you have used six months ago to write and execute a custom web scraper from a mobile phone?

Not sure about any tool but I use whatsapp + whatsapp web to "write and execute a custom web scraper from a mobile phone" since ChatGPT 4. Currently migrating from whatsapp to bittorrent or nostr. Maybe I should have turned that into an app.