A Database in the Cloud, and a Pair Programmer Who Knows It
Why Care?
FantasyGM keeps a whole league’s year in one database: every contract, every claim, every trade. Until now, that database lived on one laptop. Today it has a home online, which is the step that has to come before anyone else can sign in.
We build FantasyGM with an AI pair programmer, so we also gave it the database provider’s own instructions. It now works from the manual instead of from memory.
What’s New?
- A hosted database, loaded and checked. The league’s full setup, the player list, and the NFL schedule are in it. The app has signed in and opened every league page against it.
- The provider’s skills, kept in the repo. The database provider’s official agent skills now live in our docs folder as a linked copy of their repository. That means one command brings in their updates.
- The same pattern as our design rules. The AI finds these skills where it already finds our design system skill, so there’s one place to look for “how we do things here.”
- Live database tools for the AI, with no keys in the code. The pair programmer can connect to the database provider’s tools, and it signs in when it’s first used. No secret gets written into the repository.
The Story
The provider’s installer was happy to copy its skills into our project. Copies go stale, though, and nobody notices until the advice is wrong. So we swapped the copies for a link to the provider’s own repository, pinned to a known version. Updating is now a deliberate act rather than a surprise.
We also learned one rule from those skills right away. The app talks to the database through a connection pool that’s built for lots of short requests, while setup jobs that change the database’s structure use a direct connection. The skill spells that out, so it’s how we wired it from the first run.
What’s Next
More on this soon, once the rest of the house is ready for guests.
Written with Claude Code on Claude Opus 5.5.