BoosterTutor

BoosterTutor is a coach for Magic: the Gathering Limited. You hand it a draft or a sealed pool and it tells you which picks cost you, and why. Not a score out of ten. An actual explanation, the kind you’d get from someone sitting next to you.

The whole thing rests on one rule I wrote into CONTEXT.md on day one: Python computes every fact and every verdict, and the LLM only explains them. Ask a language model to grade a pick and it will give you a confident number it can’t back up. Hand it a number and ask it to explain, and it stays useful.

Grading a pick

Every card gets a fitness score, which is what it’s worth to your specific pool at that specific moment, measured in percentage points of win rate. The underlying data comes from 17lands. Putting everything on one scale means you can say “that card was worth 2pp less to your deck” and have it mean something.

Picks are graded on the gap to the best card in the pack rather than on rank. Most picks are close. If you flag every pick that wasn’t optimal, the three that actually mattered get buried under thirty that didn’t. So the bands are fixed: solid within 2pp, questionable within 4pp, mistake past that. If a pack had nothing at stake, meaning one playable, nothing scorable, or everything already solid, it comes back indifferent instead of graded, because no choice you could have made would have cost you anything.

The hard part is working out what a pool is trying to become. Ten two-colour archetypes get scored as competing hypotheses, each with a float for how much of your pool it could actually play. A card’s value then depends as much on what it commits you to as on how strong it is.

Building it

The API is FastAPI and SQLAlchemy on Postgres, with Alembic for migrations, uv for packaging, and mypy, ruff and pytest running over it. The front end is SvelteKit and Tailwind, tested with Vitest in a real browser via Playwright.

The stack isn’t really the point though. docs/decisions/ is. There are sixteen ADRs in there tracking how the scoring changed over time: the fitness function, the synergy term, the decay coefficient, and one full rewrite of the classification system after I realised the original was asking two questions at once. “What did this pick cost you?” and “is this card any good?” were sharing a single label, and the results read like noise.