Read the code, its callers and the existing helpers. Reuse the harness, fixtures and fakes the repository already has before writing new ones.
Choose the shape:
- Pure rule: a table of inputs and expected outputs, co-located with the code.
- Property over many inputs: a small exhaustive generator (every single-digit change of a valid code, every transposition), or a property-based test with shrinking (test-properties).
- Boundary: one test per boundary proving invalid input is rejected before any service runs.
- Typed failures: every failure kind gets a test proving it maps to the right response code.
- Flow: an in-memory system through the real router or dispatcher, with fakes that speak the real protocol and fail by name.
Write assertions that a wrong result would break. Prefer the idioms below.
Run the targeted tests, confirm they were discovered, and prove at least one fails for the right reason: against the old code, or by breaking the behavior temporarily.

Idioms that keep tests honest

Collect offenders, expect empty. expect(variants.filter(isWrong)).toEqual([]) reports every offender at once and avoids conditional assertions.
Idempotence. Parsing or normalizing an already-normalized value changes nothing.
Never-accepts tables. Blank, wrong check digit, look-alike characters (letter O for zero), wrong types (number, null, object): each must be refused.
Exhaustiveness at compile time. Type the failure table as a mapped type over every kind in the failure union, so a new kind does not compile until it has a test and a message. Pair it with a switch ending in an exhaustive-never check.
Fail where it happens. An unwrap helper that throws with the refusal attached beats three assertions that fail later with undefined.
Invariant helpers. A harness method such as "debits equal credits" called after flows.

Harness and data

Deterministic time and ids. A test clock that advances a fixed step per read and has a settable business date; sequential id generators. No wall clock, no random ids.
Seed once, clone per test. When seeding dominates test time, build the seeded store once per file, serialize it, and restore a copy per test.
Real-protocol fakes for external services, with named failure modes (not sent, lost response, refused, internal, auth) instead of per-test mocks.
Scale data through the real services with a fixed seed, never through raw inserts that skip invariants.
Golden files only with an explicit regeneration switch and a rule that every changed line is reviewed. Regenerate only when the meaning changed, never to make a test pass.
Query-plan assertions belong to test-integration: they need the production engine.

Anti-patterns

Mocking the unit under test, or a mock that restates the implementation.
Tests for trivial getters, plain assignments, or round trips through storage code you do not own (an ORM save then load).
expect inside if, or assertions with custom messages instead of offender lists.
Snapshotting a large output nobody reads.
A retry, or a sleep, to stabilize a unit test.

Done when

Each contract in scope has a test that fails on a meaningful violation, and one was shown failing.
Runner output is recorded with the command and the discovered test count.
Gaps (no fake for a service, an untestable boundary) are listed with the next action.

Continue with test-properties for wide input spaces, test-integration when the risk lives in a real dependency, test-component or test-e2e when it lives in the UI, and test-policy for repository rules.

test-unit

Unit tests

Process