Test Writer

Add tests that catch real regressions. Coverage is a tool, not a goal — a test that only exercises the happy path with the same data the implementation was written against catches nothing.

Use this skill when

• The user asks to write tests for an existing function, class, or module.
• The user wants to add coverage for a feature or bug fix.
• The user describes a behavior they want pinned down so it does not regress.
• The user is writing new code TDD-style and wants the tests first.

Do NOT use this skill when

• The user is debugging a known failure. Use debugging — once the root cause is clear, come back here to write the regression test.
• The user has not yet implemented the behavior and the design is unsettled. Use a coding skill to pin the design first.
• The user wants tests for code they are about to delete.

Procedure

1. Identify the unit under test. Be explicit: "the parseDate function in src/date.ts", "the /users/:id GET handler", "the checkout flow from add-to-cart to receipt". Vague targets produce vague tests.
2. Pick the right test type:
   • Unit: pure function, single class, mockable boundary. Fastest to run, cheapest to maintain. Default to unit tests when the logic is mostly in-process.
   • Integration: code that crosses a boundary you cannot reasonably mock (database, queue, file system). Use a real dependency in a controlled environment (testcontainers, in-memory SQLite if the queries match).
   • End-to-end: whole-system behaviors the user sees (full HTTP request, full UI click path). Expensive — reserve for the handful of flows that earn the cost. Each level has a place. Pushing logic into unit-testable functions is usually a better answer than writing more e2e tests.
3. Enumerate cases before writing assertions. For each behavior list:
   • Happy path — typical input, typical output.
   • Boundaries — empty, single, max, just-over-max.
   • Edge cases — null, undefined, negative numbers, unicode, duplicates, ordering, concurrency.
   • Error paths — invalid input, dependency failures, timeouts.
   • Idempotency / side effects — does running twice change the result? Should it?
4. Write tests that test the contract, not the implementation. Assert on observable behavior (return values, emitted events, written rows) — not on which internal method was called. Tests that over-specify implementation break on every refactor and teach future readers nothing.
5. Pick fixtures and factories deliberately. Prefer minimal, purpose-built fixtures over shared "kitchen sink" objects. A test that fails for unclear reasons usually started with a fixture that carries more state than the test needs.
6. Choose assertions that fail descriptively. expect(result).toEqual(expected) beats expect(result.length > 0).toBe(true) — when the second one fails, you learn nothing. Reach for snapshot tests sparingly: they catch unexpected changes but make intentional changes noisy.
7. Run the test in two states to prove it actually checks something: first with the implementation broken (assert the test fails), then with the implementation correct (assert the test passes). A test that passes against both implementations is decorative.
8. Keep test runtime in mind. A suite that takes 20 minutes to run gets skipped. Move slow setup behind beforeAll, parallelize what is parallel-safe, prefer in-memory fakes for unit-level tests of code that talks to a slow boundary.
9. Be deliberate about what to mock. Mock external services you do not own; do not mock the code under test or its immediate collaborators if you can use the real thing cheaply. Over-mocked tests pass without exercising any real code path.
10. For each test, write a name that describes the behavior, not the function. parses an ISO date in UTC reads better than test_parseDate_1. The test runner output becomes a readable spec of the system.
11. Group related tests with describe blocks that name the unit under test. Avoid deeply nested describe trees — they look organized but make failures hard to locate.

Examples

In scope: "Add unit tests for parseDate in src/date.ts."

→ Read parseDate. Enumerate inputs: valid ISO string, valid local date, empty string, malformed string, dates at DST boundaries, leap year, year 0, future year. Write one assertion per case. Include a "round trip" test: format(parseDate(s)) === s for the canonical formats.

In scope: "Pin down the behavior of the checkout flow."

→ This is integration- or e2e-shaped. Write an integration test that posts a cart, applies a coupon, posts payment, and asserts that an order row exists, an email was queued, and the cart was cleared. Skip the UI rendering details — those belong in a separate UI test.

Out of scope: "Why is this test flaky?"

→ Flakes are a debugging task. Use debugging to diagnose; once the cause is clear, come back here to harden the test (or delete it if the test was wrong).

In scope: "We just fixed a bug — add a regression test."

→ The diagnosis already produced the smallest input that reproduces. Write a test that pins exactly that input and assertion. Name it after the bug ("returns 400 when the limit is negative"), and keep the scope tight so the test does not break for unrelated reasons.

Out of scope: "Write tests for code I'm about to delete."

→ Don't.

Out of scope: "Generate 100% coverage for this module."

→ Coverage targets without behavior targets produce implementation-coupled tests. Ask which behaviors are load-bearing, write tests for those, and let coverage land where it does.

Self-check before responding

• Is the unit under test stated precisely?
• Did I pick the smallest test type that meaningfully exercises the behavior?
• Did I enumerate happy-path, boundary, edge-case, and error-path cases before writing assertions?
• Do the assertions describe observable behavior, not internal implementation details?
• Did I confirm the test fails when the implementation is broken?
• Are fixtures minimal and purpose-built, not shared kitchen-sink objects?
• Will the test still pass on a refactor that preserves behavior?
• Will the test fail descriptively, so a future reader can diagnose from the failure message alone?
• Is the total runtime cost worth the regression protection?