Writing Effective Tests
<activation>When This Skill Activates
- Writing or modifying test files (
*.test.*,*.spec.*,*_test.go,test_*.py) - Setting up test infrastructure or test utilities
- Debugging flaky, brittle, or slow tests
- Deciding between unit, integration, and E2E tests
- Choosing when and how to use mocks, stubs, or fakes
Natural Language Triggers
- "add tests", "test coverage", "testing strategy", "write tests for", "how should I test"
Overview
Test behaviors through public APIs. Verify state, not interactions.
Core principle: If refactoring breaks your tests but not your users, your tests are wrong.
Relationship to TDD: The shipyard:shipyard-tdd skill covers WHEN to write tests (test-first, red-green-refactor). This skill covers HOW to write tests that are effective, maintainable, and trustworthy.
The Iron Law
TEST BEHAVIORS, NOT IMPLEMENTATIONS
A test should break only when the system's observable behavior changes -- never because of refactoring, renaming internals, or restructuring code.
No exceptions:
- Don't test private methods
- Don't assert on internal state
- Don't verify method call sequences
- Don't couple tests to data structures users never see
Test Structure (AAA)
Every test follows Arrange-Act-Assert:
Arrange -- Set up preconditions and inputs
Act -- Execute the behavior under test
Assert -- Verify the expected outcome
Separate the three sections with blank lines. One Act per test. One logical assertion per test.
<examples> <example type="good" title="Clear AAA structure, one behavior per test"> ```python def test_expired_subscription_denies_access(): # Arrange user = create_user(subscription_end=yesterday())# Act
result = check_access(user, resource="premium-content")
# Assert
assert result.denied is True
assert result.reason == "subscription expired"
Clear name, tests one behavior, obvious structure.
</example>
<example type="bad" title="Multiple behaviors, internal state tested">
```python
def test_subscription():
user = create_user(subscription_end=yesterday())
assert check_access(user, "premium-content").denied
user.subscription_end = tomorrow()
assert check_access(user, "premium-content").allowed
assert user.access_log == [("denied", "premium-content"), ("allowed", "premium-content")]
Two behaviors in one test, tests internal log, vague name. </example>
</examples>Keep Tests DAMP, Not DRY
Prefer Descriptive And Meaningful Phrases over eliminating duplication. Duplicating setup across tests is fine if it makes each test self-contained and readable. Extract shared setup into helpers only when it improves clarity, not to reduce line count.
No Logic in Tests
Tests are straight-line code. No loops, conditionals, ternaries, or string concatenation. If you need logic, the test is too complex -- split it or simplify the design.
What to Test
Test These
- Behaviors: What the system does from a user's perspective
- Edge cases: Empty inputs, boundaries, overflow, zero, null
- Error paths: Invalid input, missing dependencies, timeouts
- State transitions: Before and after an operation
Skip These
- Trivial getters/setters with no logic
- Generated code (protobuf, ORM migrations)
- Configuration files
- Third-party library internals
Test Via Public APIs
Invoke the system the same way its callers do. If the only way to test something is through a private method, the design needs to change -- extract it into a collaborator with its own public interface.
Naming Tests
Name tests after the behavior, not the method.
| Good | Bad |
|---|---|
rejects_empty_email_with_validation_error | test_validate |
returns_cached_result_when_within_ttl | test_cache |
retries_three_times_before_failing | test_retry_logic |
grants_access_when_subscription_active | test_check_access_method |
Patterns that work:
[action]_[condition]_[expected_result]should [expected behavior] when [condition]- Plain descriptive sentence
A test name containing "and" usually means two tests.
Choosing Test Level
digraph test_level {
rankdir=TB;
start [label="What are you testing?", shape=diamond];
pure [label="Pure logic?\nNo I/O, no side effects", shape=diamond];
boundary [label="System boundary?\nDB, API, file, queue", shape=diamond];
journey [label="Critical user journey?\nLogin, checkout, signup", shape=diamond];
unit [label="UNIT TEST\nFast, isolated, many", shape=box, style=filled, fillcolor="#ccffcc"];
integration [label="INTEGRATION TEST\nBoundary focused, fewer", shape=box, style=filled, fillcolor="#ffffcc"];
e2e [label="E2E TEST\nFull stack, minimal", shape=box, style=filled, fillcolor="#ffcccc"];
start -> pure;
pure -> unit [label="yes"];
pure -> boundary [label="no"];
boundary -> integration [label="yes"];
boundary -> journey [label="no"];
journey -> e2e [label="yes"];
journey -> unit [label="no\n(rethink)"];
}
The pyramid: Many unit tests. Fewer integration tests. Minimal E2E tests.
Push tests down. If a behavior can be tested at a lower level, test it there. Lower = faster, more stable, cheaper to maintain. Only go higher when the lower level can't verify the behavior (serialization, wiring, full user flow).
Duplicate coverage? If a unit test and integration test verify the same behavior, keep the unit test. Remove the integration test unless it adds confidence about wiring.
Test Isolation and Doubles
When to Use Test Doubles
| Use doubles for | Use real objects for |
|---|---|
| Network calls (HTTP, gRPC) | In-process collaborators |
| Database queries (in unit tests) | Value objects and data structures |
| Filesystem I/O | Pure functions |
| Non-deterministic sources (time, random) | Deterministic transformations |
| Expensive operations (ML inference, rendering) | Lightweight dependencies |
Types of Doubles
- Stub: Returns canned responses. Use when you need a dependency to return specific data.
- Fake: Working implementation with shortcuts (in-memory DB, local file store). Preferred over mocks for complex interactions.
- Mock: Records calls and asserts on them. Use sparingly -- only when the interaction IS the behavior (e.g., "sends an email").
The Mocking Rule
Verify STATE over INTERACTIONS.
Mock only when the side effect IS the behavior under test.
<examples>
<example type="good" title="Testing resulting state">
```typescript
// Testing that withdrawal updates balance (state)
test('withdrawal reduces account balance', () => {
const account = new Account(100);
account.withdraw(30);
expect(account.balance).toBe(70); });
Tests resulting state -- survives internal refactoring.
</example>
<example type="bad" title="Testing internal interaction">
```typescript
// Testing that withdrawal calls the right methods (interaction)
test('withdrawal calls debit', () => {
const ledger = mock(Ledger);
const account = new Account(100, ledger);
account.withdraw(30);
expect(ledger.debit).toHaveBeenCalledWith(30);
});
Tests internal interaction -- breaks if refactored. </example>
</examples>Unchanging Tests
A well-written test changes only when requirements change. It survives:
- Refactoring -- internal restructuring, no behavior change
- Bug fixes -- new tests added, existing tests unchanged
- New features -- unrelated tests unaffected
If your tests break during refactoring, they're testing implementation details. Fix the tests to test behavior instead.
Clear Failure Messages
A failure message should let you diagnose the problem without reading the test code.
<examples> <example type="good" title="Descriptive failure message"> ``` FAIL: re