VibeJournal

Misc. thoughts after poking around with Claude and Claw

A blank canvas reflects its artist. It’s an open blue sky of possibilities, with no seed to point you in a direction. The artist has to fill that potent emptiness. Vibe coding promises a blue sky of opportunity, but it’s almost too blue. I have a full-time white collar worker available. What did I do with it? So far, all I’ve made is a library bot that recommends and checks out books for my kids at the local library, and a smash bros data analyzer. I think that reflects on who I am at the moment.

I’m still pretty fresh to all this AI agent stuff, but I feel like I’ve gained some insight from poking around openclaw. One revelation was the importance of context, and managing that information. Context mattered far more than model power or token size. Many openclaw “skills” were actually just a markdown file (a fancy text document, like wiki pages) with all text and no code. I was surprised; I thought it would be like most code packages or libraries, full of subfolders, but this was just flat and simple.

The inverse is also true; if AI agents are missing context, then they struggle. But this is no problem! It can do its own research. For the Melee data analyzer, it had to learn how to work with the niche melee emulation library, Slippi. Claude searched the web, found top github projects that interacted with Slippi, and read a bunch of their code to understand their codebase. Then, Claude searched for social media or strategy guide content on melee to better understand the technical jargon and natural language of the smash community. It knows the frame data better than I do.

I’m starting to notice more simple hobbyist websites cropping up. I suspect they’re vibe-coded. Kirby Air Riders character stats, Iterated prisoner’s dilemma contest. They’re creative and curious ideas brought to life. But there’s something that feels janky and incomplete about them. They’re created once and never updated, like they were a one-shot vibe coding project.

I’ve been getting use out of my library concierge bot. It automatically gets books for everyone in the family; picture books for the 4yo, animal adventure books for the 7yo, and veggie-heavy ethnic cookbooks. Though I’m not sure it was worth it; it took me probably 12 hours to set up, and saves me maybe 10m every week. It has to run for a year and a half to break-even. It’s likely to break during that payback window, which makes the expected value even worse. Is this thing actually any faster? I find it hard to judge. Sometimes it feels like it miraculously solves a problem, and other times I find it creating more issues for me to investigate.