I vibe coded this to make some custom trivia for a family gathering. It would benefit from another pass at the UX for playing, and some question design changes, but that’s just the limits of my attention not the vibe coding.

trivia.ianbicking.org

For some non-code Claude Code projects I’ve been making a little local issue system, which works kind of similarly to what you describe. Mostly it’s skills and commands to describe the layout, not much code is required

Subagents really burn the tokens, I wonder if it avoids them on that plan

Confiding in AI is deeply private. You can just delete the conversation. Undo the whole thing. It can forget in a way that people never forget. The only thing close is like you said, talking to anonymous randos on the internet.

So if a kid confides in someone about suicidal ideation there's no undoing it, they can't experiment or test the waters to see what the reaction will be. They can do that inside their own head. We all experiment inside our heads, try out a thousand things in our imagination that never leave our head

First, I don't think we should be so down on ourselves. There's openness around so many topics that wasn't there ten, twenty years ago.

But hard things are going to be hard. We can (individually, collectively) be chill about homosexuality, but we can't be chill about suicidal ideation

And I think you're right: bad criticism is a better vehicle for this.

In my last encounter I felt like there was a very strong desire to manufacture consensus, defining out anyone who disagreed. So I was defending guidelines written by academics from other academics, but they latched onto me and not the original source because they could (gleefully) dismiss me.

When you play the Wordle board game with your daughter noting the green/yellow letters, you may be playing this version

Do you have a sense of how much "contest culture" is a personal perception, vs a property of the environment?

So I guess I understand why some people think these tools are so bad, the AI happy stack is so much better.

I don’t think it’s just training! Tailwind is great for locality (without it I got so many zero-impact CSS changes), typescript for organization, and Next.js is… fine.

I just wouldn’t rely on an LLM’s native sense of surprise. Even for things well within the training set I don’t think LLMs are great at discernment, especially plucking out something unusual. But there are systematic ways to approach this without relying on surprise

This feels like another “I tried a little bit and it didn’t work, so it can’t work” critique of LLMs.

You still have to build systems. Come up with rubrics. Find information that defies categorization in that rubric. LLMs need an investigative process just like researchers

I had come upon a map of where they ban corporal punishment of children, which notably includes most of Latin America, and made me think of this… brilliantmaps.com/corporal-pun...

I’m giving Claude Code small bugs with only a functional description, in environments where it can’t reproduce the error itself or verify the fix, and it’s very frequently fixing them on that alone. Bugs I know I couldn’t fix without manual testing and iteration. I find it surprising!

Interestingly while I assume education levels are different between white and Hispanic parents, apparently use of spanking isn’t.

From www.pewresearch.org/social-trend...

It's like they were all trying very hard to get me to hate academics and academia, which I do not care to do.

It did get a little weird when I was advocating that research has value outside academia, even historical research, and a historian was like "no it doesn't"

Selfishly, it's a more pleasant for me to interpret it this way