Describe it. Get the whole course.
The Builder exists because “generate 20 flashcards” is not how anyone actually learns a subject. A real topic is a course: ordered sections, foundations before edge cases, sized by what the material needs — up to 2,000 cards in one deck.
The brief.
One or two sentences about what you're learning and why — “Georgian basics, I live in Tbilisi”, “NCLEX pharmacology, exam in June”. The Builder triages: does this need web research for freshness? How big should it honestly be? Is anything ambiguous enough to ask about? It asks at most a few questions, and only when the answer genuinely changes the outline.
One checkpoint: the plan.
You get back a premise — one line stating what the deck assumes about you and where it's taking you — plus a course-ordered outline of sections and a recommended size. This is the single moment the Builder waits for you: adjust the size, answer its questions, or just say what's off in plain words (“more conversational phrases, less grammar”) and it re-plans. Approve, and you're done supervising.
If you're studying a language: the deck teaches toward the language you're learning, in the language you learn in. The premise states both, so a wrong guess is visible before a single card exists.
The build runs on the server — walk away.
Sections are written one by one, about 40 cards per call, each batch passing the same card-quality guard that polices agents on the MCP. Research-flagged sections are grounded in real web sources first. You can close the app, put the phone in your pocket, lose signal — the build continues server-side, and every surface (home row, deck screen, the Create tab) shows live progress when you come back. Cards appear as sections finish, so you can start studying section one while section twelve is still being written.
If something stalls — a model hiccup, a rate limit — simply opening the plan resumes it. Failed sections can be rebuilt individually; you never lose what's already built.
Sized by the material, bounded by the budget.
The Builder recommends the size the material deserves — a tipping mechanism is 15 cards, a language foundation is several hundred — and it tells you why. Hard bounds exist so one runaway build can't eat everything you have left to build: per deck, and per month. They're published on Limits & pricing and shown live in your Account — the same numbers the server enforces.