Seed -- Database Seeding & Test Data

Activate When

• /godmode:seed, "seed the database", "generate test data"
• "create fixtures", factory patterns (FactoryBot, fishery)
• Fake data generation (Faker.js, Faker Python)
• Data anonymization for production snapshots

Workflow

Step 1: Detect Environment

# Detect ORM and factory libraries
grep -r "prisma\|typeorm\|sequelize\|django\|sqlalchemy" \
  package.json pyproject.toml Gemfile 2>/dev/null
ls prisma/seed.ts db/seeds.rb scripts/seed* 2>/dev/null
grep -r "@faker-js\|faker\|factory_bot\|Bogus" \
  package.json pyproject.toml 2>/dev/null

SEEDING ENVIRONMENT:
Language: <TS | Python | Go | Ruby | Java>
ORM: <Prisma | Drizzle | SQLAlchemy | ActiveRecord>
Database: <PostgreSQL | MySQL | SQLite | MongoDB>
Existing seeds: <path or "none detected">
Factory lib: <fishery | factory_boy | none>
Faker lib: <@faker-js/faker | Faker Python | none>

Step 2: Factory Pattern

Library selection:
  TypeScript: fishery
  Python: factory_boy
  Ruby: FactoryBot
  Go: table-driven builders + gofakeit
  C#: Bogus
  Java: Instancio

Key patterns:
  Base factory with faker defaults + sequence IDs
  Traits for variants (admin, inactive, published)
  build() for in-memory, create() for persisted
  buildList(N) for batches
  Override any field at call site

Step 3: Seed Script Architecture

IDEMPOTENT SEED RULES:
1. ALWAYS upsert (insert or update), never plain insert
2. Stable identifiers (slug, email) not auto-increment
3. Dependency order (users before posts)
4. Wrap each group in transaction
5. Log: created vs skipped vs updated
6. Re-runnable: running twice = same result

SEED STRUCTURE:
  faker.seed(42) for determinism
  Phase 1: reference data (roles, categories via upsert)
  Phase 2: core entities (fixed + random users)
  Phase 3: dependent entities (posts, comments)
  Check count vs target, only create delta

Step 4: Relationship Handling

Seed in topological order: reference -> independent -> first-level deps -> second-level deps -> M:N joins -> derived data.

Use Pareto distribution: 20% of parents get 80% of children for realistic associations.

Step 5: Large Dataset Seeding

NEVER insert one row at a time.
Batch size: 500-5000 rows per INSERT.
IF 100K+ rows: streaming or raw COPY (PostgreSQL)
IF bulk loading:
  SET session_replication_role = 'replica'
  (disable FK checks during load)
  After: REINDEX, ANALYZE

Performance thresholds:
  50 rows: <1s (dev seed)
  5K rows: <10s (staging seed)
  100K rows: <60s (demo seed via COPY)

Step 6: Deterministic Seeds

• faker.seed(42) = same data every run
• Use sequences instead of random IDs
• Different seeds per environment
• Per-test: reset seed before each suite

Step 7: Data Anonymization

IF production snapshot needed:
  Replace emails: user_N@example.com
  Replace names, phones, addresses: faker
  Remove: credit cards, API keys, tokens, passwords
  Add noise to financial amounts (+-10%)
  Shift dates by random offset
  NEVER copy production without anonymization

Step 8: Cleanup Strategies

Truncate cascade: full reset
Delete by marker: DELETE WHERE source = 'seed'
Transaction rollback: BEGIN->seed->test->ROLLBACK
Database drop/create: CI pipelines
Snapshot restore: pg_restore from dump

ENVIRONMENT-AWARE:
  Dev: 50 users, deterministic, minimal
  Staging: 5K users, randomized, weekly reset
  Demo: 200 curated, deterministic
  Production: NEVER seed fake data

Step 9: Commit

Commit: "seed: add <desc> seeding infrastructure"

Key Behaviors

1. Keep seeds idempotent. Upsert, never blind insert.
2. Factories for tests, scripts for environments.
3. Seed in dependency order. Parents first.
4. Deterministic for development. faker.seed(42).
5. Batch insert always. 100x faster than loops.
6. Tag seeded data. @example.com for cleanup.
7. Never seed production with fake data.
8. Never ask to continue. Loop autonomously.

HARD RULES

1. NEVER seed production with development data.
2. ALWAYS use fixed faker seed for determinism.
3. NEVER use auto-increment IDs as stable refs.
4. ALWAYS use batch inserts. No single-row loops.
5. NEVER include real user data in seed files.
6. ALWAYS make seeds idempotent with upsert.
7. NEVER seed without a transaction.
8. ALWAYS provide --reset for clean-slate option.

Auto-Detection

grep -r "prisma\|typeorm\|sequelize\|django" \
  package.json pyproject.toml Gemfile 2>/dev/null
ls prisma/seed.ts db/seeds.rb scripts/seed* 2>/dev/null

TSV Logging

Log to .godmode/seed-results.tsv: step\tentity\torm\tmethod\tstatus\tdetails

Output Format

Print: Seed: {N} entities, {rows} rows. Idempotent: {yes|no}. Env guard: {active}. Status: {DONE|PARTIAL}.

Keep/Discard Discipline

KEEP if: idempotent (re-run safe) AND batch-inserted
  AND dependency order correct
DISCARD if: unique constraint on re-run
  OR FK errors OR single-row loops

Stop Conditions

STOP when:
  - All entities seeded idempotently
  - Env guard active, --reset works
  - User requests stop