I've been building a small dashboard on a Raspberry Pi 4B that sits on my desk as a second screen, currently showing local weather (plus a meta touch: it tracks my own Claude Code usage too). Flask serves the page, and two collectors run independently on systemd timers. The app never touches an external API directly, it just reads whatever the collectors last wrote, so a dropped Wi-Fi connection means stale data instead of a blank screen. Before writing any code, I used the WAT framework from our earlier lessons to build a claude.md file for the project: environment, tools, architecture, and a running list of gotchas I'd already hit once. That shared context is what made the AI collaboration work. What helped wasn't Claude writing code for me, it was Claude explaining why one architecture beats another before I touched a line, like why timestamps belong in UTC or why paths need to be anchored to the script instead of relative. Small steps, an explanation before each change, verify it actually worked. Closer to a patient TA than a code generator. Discussion question: Weather's just the start. Calendar, unread email count, Pi health stats, and stock positions are all on the list next. If you were adding a data source to this dashboard, what would you pick, and what's one design decision (like the "stale over blank" tradeoff) you'd have to make for it to fail gracefully?