Test-Driven Development Workflow

Write the test before the code, watch it fail for the right reason, then make it pass. This skill owns the loop, the coverage gate, and the mocking boundary. coding-standards is the hub above it and the language testing skills hold the per-stack mechanics.

When to activate

• Writing a new feature, an endpoint, or a component.
• Fixing a bug, where the test that reproduces it comes first.
• Refactoring, where the existing tests are the safety net that has to stay green.
• The coverage gate is failing, or a pull request is blocked on it.
• Deciding what to mock, or whether a test belongs at the unit, integration, or end-to-end layer.

When not to activate

• Browser journeys through a real UI, and the flaky-test policy. Use e2e-testing, which is canonical for both.
• Capability sweeps against a deployed stack with an API client. Use e2e-runbooks.
• Regression tests aimed specifically at AI-introduced defects and sandbox-path drift. Use ai-regression-testing.
• Per-language mechanics: pytest fixtures and parametrisation, Go table-driven tests and fuzzing, JUnit slice tests. Use

python-patterns, golang-patterns, or springboot-patterns.

• Load, soak, or capacity testing. Use performance-optimization for the measurement discipline instead.

The loop

Write the user journey, turn it into test cases, watch them fail, implement the minimum, watch them pass, then refactor under green.

State the journey in one sentence before writing any test, because a test suite that cannot be traced back to a journey usually tests the implementation instead.

As a signed-in customer, I want to search the catalogue by keyword,
so that I can find a product without knowing its exact name.

Turn the journey into cases that name the behaviour and the condition.

describe('catalogue search', () => {
  it('returns matching products for a keyword', async () => {})
  it('returns an empty list for a query that matches nothing', async () => {})
  it('falls back to substring matching when the search index is unavailable', async () => {})
  it('orders results by relevance score, highest first', async () => {})
})

The RED gate

Running the new test and seeing it fail is mandatory before any production code changes. A test that was written but never executed proves nothing, and a failure caused by a typo proves nothing either.

npm test

A valid RED state is one of two things. Runtime RED: the target compiles, the new test actually executes, and the result is a failure. Compile-time RED: the new test references code that does not exist yet, and the compile failure is itself the signal.

In both cases the failure must be caused by the missing implementation or the bug under repair, not by unrelated syntax errors, broken setup, missing dependencies, or a pre-existing regression elsewhere.

Fail: the failure is real but proves nothing about the behaviour under test.

FAIL  src/search.test.ts
  Cannot find module '@/lib/testing/helpers'

Pass: the failure is the behaviour the test was written to demand.

FAIL  src/search.test.ts > returns an empty list for a query that matches nothing
  expected [] but received undefined

Do not touch production code until a valid RED state is confirmed. Then implement the minimum that turns it green, rerun the same target, and only refactor once green.

npm test

Refactoring under green means removing duplication, improving names, and simplifying structure, with the suite rerun after each step.

Coverage gate

Around 90 percent line coverage of real logic, across unit, integration and end-to-end tests combined. Branch coverage sits at 70 to 80 percent. Critical logic reaches 100 percent where that genuinely adds value rather than as a ritual. No exclusion pattern may be added to dodge a meaningful test, and every edge case, error path, and boundary condition is covered.

Enforce it as a CI gate so a pull request that drops coverage below the threshold is blocked.

Fail: the config key is plural, so Jest ignores the whole block and the gate never fires.

{
  "jest": {
    "coverageThresholds": {
      "global": { "branches": 70, "functions": 90, "lines": 90, "statements": 90 }
    }
  }
}

Pass: the key Jest actually reads.

{
  "jest": {
    "coverageThreshold": {
      "global": { "branches": 70, "functions": 90, "lines": 90, "statements": 90 }
    }
  }
}

Verify locally before pushing.

npm run test:coverage

Test layers

Three layers, with the bulk of the suite at the bottom. The proportions and the tooling per layer are in references/advanced-testing.md.

• Unit. Individual functions, pure logic, component behaviour, helpers. No input or output.
• Integration. Endpoints, data access, service interactions, external boundaries, run against real engines through

Testcontainers.

• End-to-end. Critical user journeys through the real UI. e2e-testing owns this layer entirely, including the

Playwright patterns, locator rules, and CI wiring. Write the journeys there and reference them from here.

Test naming

Descriptive names in the project's existing style win. A reader who has never seen the code should learn the behaviour and the condition from the name alone.

Match what the project already does. Where a project has no convention, use the natural-language form from coding-standards, which is the canonical rule: a sentence describing the behaviour and the condition that triggers it.

Fail: names that describe nothing.

test('works', () => {})
test('test search 2', () => {})
test('happyPath', () => {})

Pass: natural-language form, the default.

test('returns an empty list when no product matches the query', () => {})
test('throws when the API key is missing', () => {})

Pass: the <method>_<scenario>_<expectedResult> form, which is the same rule spelled differently and is idiomatic in JUnit codebases. Use it when the project already uses it, not as a second convention alongside the first.

@Test
void calculateTotal_withEmptyCart_returnsZero() { }

@Test
void processPayment_whenCardDeclined_throwsPaymentException() { }

The two forms are alternatives, never a mix inside one codebase.

Test structure

Every test sets up state, performs one action, then asserts the observable outcome. Which words label the three phases is the project's choice: Arrange / Act / Assert and Given / When / Then are the common spellings. Read the existing tests and match them.

One behaviour per test. Several assertions are fine when they all verify that one behaviour, and each test builds its own data so the order tests run in never matters.

Fail: the second test depends on the first.

test('creates a user', () => { createUser('alice') })
test('updates the same user', () => { updateUser('alice', { name: 'Alicia' }) })

Pass: each test owns its setup.

test('creates a user', () => {
  const user = createTestUser()
  expect(findUser(user.id)).toBeDefined()
})

test('updates a user', () => {
  const user = createTestUser()
  updateUser(user.id, { name: 'Alicia' })
  expect(findUser(user.id).name).toBe('Alicia')
})

Test what the user observes

Assert the behaviour the caller can see, not the internal state that produced it. A test coupled to internals fails on every refactor and passes through real defects.

Fail:

expect(component.state.count).toBe(5)

Pass:

expect(screen.getByText('Count: 5')).toBeInTheDocument()

Never weaken an assertion to make a test pass. Fix the code, or fix the setup that was wrong.

Mocking boundaries

Mock only what you own, and only in unit tests. Do not mock the database when you can exercise it: integration tests run against a real engine through Testcontainers, never a shared staging instance. For third-party HTTP clients, ORMs, and cloud SDKs, prefer a real instance, the library's own official test double, or an HTTP-level fake such as WireMock or MSW. The detail behind both rules is in references/advanced-testing.md.

The mocks below are unit-test doubles for a boundary the test does not own, not a substitute for exercising the real datastore.

A repository, mocked at the interface the application defines.

jest.mock('@/lib/repositories/product-repository', () => ({
  productRepository: {
    findByKeyword: jest.fn(async () => [{ id: 1, name: 'Test Product' }]),
    findById: jest.fn(async () => ({ id: 1, name: 'Test Product' })),
  },
}))

An external HTTP client, mocked at the project's own wrapper rather than at the vendor SDK.

jest.mock('@/lib/clients/search-client', () => ({
  searchClient: {
    query: jest.fn(async () => [{ id: 'test-product', score: 0.95 }]),
    health: jest.fn(async () => ({ reachable: true })),
  },
}))

Note what both examples have in common: the seam is an interface the project defines, so the mock stays valid when the vendor library changes underneath it.

Flaky tests

e2e-testing holds the canonical flaky-test policy: the fix-or-quarantine SLA, the maximum quarantine period, what happens when it expires, the prohibited sleep patterns, and when a framework retry is allowed. It applies to every layer, not only browser tests. Do not restate it here, and do not adopt a second policy alongside it.

Continuous testing

Run the suite in watch mode while developing.

npm test -- --watch

Gate the commit on tests and lint.

npm test && npm run lint

In CI, run with coverage and upload the report.

- name: Run tests
  run: npm test -- --coverage
- name: Upload coverage
  uses: codecov/codecov-action@v7

Reference map

Related skills

• coding-standards holds the canonical test-naming and test-structure rules this skill applies.
• e2e-testing owns the browser layer and the flaky-test policy.
• e2e-runbooks owns capability verification against a deployed stack.
• ai-regression-testing owns bug-driven regression tests and sandbox-path parity.
• python-patterns, golang-patterns, springboot-patterns hold the per-language mechanics.
• performance-optimization owns the measurement discipline when a test proves something is slow.

Checklist

• [ ] A user journey was stated before the first test was written.
• [ ] The test was executed and produced a valid RED for the intended reason.
• [ ] No production code changed before that RED was confirmed.
• [ ] The implementation is the minimum that turns the test green.
• [ ] Refactoring happened only under green, with the suite rerun after each step.
• [ ] Around 90 percent line coverage of real logic, branch coverage 70 to 80 percent, enforced in CI.
• [ ] The coverage config uses the key the tool actually reads.
• [ ] Test names describe behaviour and condition, in one convention across the codebase.
• [ ] Every test builds its own data and passes in any order.
• [ ] Assertions target observable behaviour, and none was weakened to get green.
• [ ] Mocks sit at interfaces the project owns, and integration tests use Testcontainers.
• [ ] The full suite ran, not just the one test that was being fixed.
• [ ] No test is skipped or disabled without a tracked quarantine entry.