This month I got annoyed with YouTube search in a very specific way: it’s fine when you know exactly what you want (a song, a podcast, a specific video), but it totally falls apart when you’re trying to mix ideas and discover something new. My example was “woodworking and electronics” — YouTube gave me ads and Shorts, even though channels like “I Like To Make Stuff” literally exist for that exact overlap. So I did what I always do when something annoys me: I built my own version. In this video I walk through how I added a custom search algorithm to MyTube (my DIY YouTube from the previous recommendation-system build). The flow is: you type a query X, I send it to an LLM (I’m using Groq’s Llama right now) to generate relevant channels, then I pull videos per channel, and then I use another LLM pass to pick the best 10 videos based on titles/descriptions while avoiding the same creator dominating the list. The “build-in-public” reality check is the YouTube API quota pain—so I talk about caching, and why I’m exploring YouTube RSS feeds as a way around heavy API usage. Also: I share a mistake I made with mixing test and prod data, and why clean data matters even for “just a side project.”