Tooling

There’s no HDHomeRun app in Samsung’s store, so I built one with Claude Code during a football game. The Apple TV app exists, but it lists channels like files in a folder that you click through, and my Apple TV isn’t even set up right now. I’d been watching through VLC network streams on my phone or laptop. I wanted the TV itself to behave like a TV: open on a channel, flip with Ch+ and Ch-, see what’s on. As far as I could find, that didn’t exist. (SiliconDust has a sideload-only test build on their forum that I never tried.) ...

My resume lives in a git repo. Every time I push, a GitHub Action compiles it to PDF and drops it on a stable URL. No local toolchain needed after the initial setup, no manually uploading PDFs anywhere. Why LaTeX Word docs drift. PDFs exported from Google Docs look fine until they don’t. LaTeX gives me exact control over layout and is just a text file — diffable, versionable, no binary blobs. ...

derekwelty.com was a Bootstrap 5 single-page site built on the Start Bootstrap Freelancer template. It worked, but it hadn’t been touched in years and updating it required wading through jQuery, a Gulp pipeline, and a bunch of hand-rolled CSS. Time to replace it. The new stack: Hugo with the Congo theme, deployed to GitHub Pages via GitHub Actions. Why Hugo I already run this blog on Hugo + PaperMod via Netlify. The toolchain was familiar, the build times are fast, and Congo gave me a clean profile layout that actually looked like a personal site and not a blog template with the posts removed. ...

ChatGPT worked fine for one-off questions, but it was isolated from the systems I actually use: calendar, email, Strava, GitHub, Obsidian, Spotify, and the half-finished project notes scattered across repos. I wanted a small assistant I could run myself, with persistent memory, scheduled checks, and enough tool access to be useful without turning into a full product. So I built Molty. I didn’t type its code or its instruction files; Molty did, and 81 of the first 83 commits in its workspace repo are under its name. But it wasn’t a one-shot prompt: I was in the details the whole way, directing each change. ...

I kept missing shows I would have gone to. The data was public, but checking it required a weird amount of manual attention: venue calendars, Spotify history, ticket prices, neighborhoods, and a final “would I actually leave the apartment for this?” filter. So I had Molty, my Telegram assistant, build a weekly concert digest. It wrote all of the code on March 29: the Spotify login, a scraper for each venue, and the scoring. My part was taste: where I go, what I’d pay, how far I’d walk, and telling it when a pick was wrong. Every Monday morning it checks SF venues, matches artists against my Spotify library, and sends a ranked list of shows. ...