aleshh
Joined 3 years ago
Joined 3 years ago
I also love to read reviews, so I aggregate them from a bunch of websites, and I end up with a sort of Rotten Tomatoes for new music? (An offline worker crawls the review sites and an LLM extracts the sentiment and a quote.)
The challenge with these things is how do you get started. I've tried to create starter stacks and I've got a way to import from Spotify or Apple Music playlists.
You can explore without an account. It's required for searching and saving albums. Free for now, I'm hoping to add a paid tier at some point. Letterboxd was a big inspiration.
Tech stack: Next.js on Vercel, Supabase. iOS app is almost finished, it's mostly a SwiftUI shell around server-rendered pages. This has been an adventure in seeing how far I can push LLM coding while still trying to be a software engineer. Pushing back on its wild schemes, trying to keep the DB schema sensible, and collaborating with it on UX ideas.
Curious to hear what people think. Anything missing that would make you use something like this?
I'm not sure I like this at all, but hopefully it means the ChatGPT app won't leak memory and need to be restarted every few hours!
Do you think that would help? Something else?
I've taken some steps to mitigate this. I made a design brief (https://gist.github.com/aleshh/7435682311b6cb944bf18ecc3751f...) for LLMs to follow to kind of steer them in a particular direction. And I do spend time tweaking design with them.
But I agree it's not enough. The problem is, there are as I see it really two options here
(1) Guide the LLM and just spend more time on the visual stuff. The trouble is those refactors take as much time as adding major new features, sometimes more, because the LLM doesn't get it right and you gotta go around and around. So it takes a big time investment to go from something that's pretty good to something that's merely somewhat better.
(2) The other option is to do it all by hand. I sometimes do some of this, but the further along you get the harder it is, because the LLMS write lots of CSS, and (at least the last time I tried) refactoring their stuff is time-consuming and kind of a prerequisite to tweaking things.
Probably what I should do is expand on the design spec, and add it to AGENTS.md (instead of just throwing it in on the initial prompt) and generally try to get good on this aspect.
I should note that on my "real" project (https://flipper.fm), I spend a lot more time steering the design, creating a component library, Storybook, creating CSS classes for the LLM to use, etc etc. Again, this thing is a toy.
Sorry for the long answer, but you're really hitting on something real here.
- All the notes on fingering charts can be clicked and they play a sound
- Chords get a little playback button
- Compose mode gets its own playback controls
The sound is very basic, I'll see if I can fix that next.