Three ways to describe what you want: a Prisma schema.prisma file, raw SQL DDL, or a live Postgres connection (introspected via information_schema + pg_catalog). All three normalize into the same internal model so the rest of the pipeline doesn't care how you got there.

Composite primary keys, deferred constraints, partial indexes, generated columns, enums — all preserved. We build the FK dependency graph and topo-sort it for insert order.

For each column, seedkit picks a generator from a layered strategy: column-name + type heuristics first ( email, phone, currency, jsonb shapes), then domain-aware LLM prompts when the column is open-ended (free text, names, addresses by locale).

A per-seed PRNG drives sampling, so --seed my-fixture on a Tuesday yields the exact same rows on a Friday. No model temperature roulette, no “regenerate to get something usable.”

Pass --prompt "b2b crm" or --scope healthcare and the generators bias accordingly. CRM gets account hierarchies, fintech gets currencies that don't make accountants cry, healthcare gets ICD-10 codes that resolve to real diagnoses.

Free text is locale-aware — German addresses look German, Japanese names look Japanese, prices follow the currency you'd expect.

The dependency graph is computed before generation begins. Composite keys are treated as units. Cycles (like a users.invited_by → users.id self-reference) are broken by deferring the constraint and inserting nullable first, then patching the references in a second pass.

Run --preview to see sample rows and the FK satisfaction report before touching a database.

The same generation engine runs in our browser playground — paste a schema, write a prompt, watch rows appear in a live table. Iterate on the shape without committing to a database. When it looks right, promote it to a live ephemeral DB or a Mock API in one click — same seed, same rows.

Useful for: scoping a new product, sanity-checking a schema before a migration, generating a demo dataset that PMs can poke without your help.