test-architect — tests that would catch the bug

When to use this skill

Trigger when the user wants tests for specific code. Strong signals:

• "write tests for <file or function>"
• "cover this with tests"
• "add unit tests"
• "I need tests for this"
• A function pasted with no further context

Do not trigger for: e2e/browser flows (use browser-qa), test infrastructure setup, or when the user only wants you to fix a failing test (just fix it).

The output contract

Tests that:

1. Run — npm test, pytest, go test, etc., all green on the new file.
2. Match the codebase — same framework, same test file location, same naming convention as the existing tests.
3. Cover the contract, not the implementation — happy path, failure modes, boundary values, empty/null inputs.
4. Fail when behavior breaks — every assertion would detect a real regression, not just confirm the code ran.
5. Are independent — no test depends on the order or state of another.

Workflow

1 — Detect the test stack

Inspect the repo first:

• Which framework? Look at package.json (jest, vitest, mocha), pyproject.toml / pytest.ini, go.mod, Cargo.toml.
• Where do tests live? __tests__/, test/, *.test.ts next to source, *_test.go?
• What conventions? describe/it vs test(), AAA vs given/when/then, fixture style, mocking library.

Match what's there. Don't introduce a new framework just because you prefer it.

2 — Read the function like an adversary

Before writing assertions, list the failure modes you can think of:

For a function with this signature:

function parseDuration(input: string): number  // returns ms

list:

• "30s" → 30000
• "5m" → 300000
• "1h" → 3600000
• "" → throws? returns NaN? returns 0?
• "abc" → throws?
• "5" → no unit — throws or assumes ms?
• "5x" → unknown unit — throws?
• "-5s" → negative — allowed?
• null, undefined → TypeScript prevents at compile, but runtime?
• Float input: "1.5s" → 1500?

That list is the test plan. Now write one test per item.

3 — Structure the test file

For each function under test, group:

describe('parseDuration', () => {
  describe('happy path', () => { ... })
  describe('boundary values', () => { ... })
  describe('invalid input', () => { ... })
})

Name tests with the input → output, not the implementation:

✅ 'parses "30s" as 30000' ✅ 'throws on empty string' ❌ 'works' ❌ 'returns the result' ❌ 'test 1'

4 — Write real assertions

Each test asserts a specific observable. Never:

expect(result).toBeTruthy()      // ❌ what does truthy mean?
expect(result).toBeDefined()     // ❌ a string '' is defined too
expect(true).toBe(true)          // ❌ this is not a test

Always:

expect(result).toBe(30000)
expect(result).toEqual({ id: 1, name: 'Ada' })
expect(() => parseDuration('')).toThrow(/empty/)
expect(result.items).toHaveLength(3)

For async: await expect(promise).rejects.toThrow(SpecificError).

5 — Stub at the right layer

• Pure functions: no mocks needed.
• Functions that call other modules: stub at the import boundary, not deep inside.
• HTTP: use msw (browser/node), nock, or responses (Python). Don't stub fetch globally — that breaks other tests.
• DB: prefer an in-memory or test-container instance over mocking the ORM. Mocks of complex ORMs lie.
• Time: use the framework's fake timers; never await new Promise(r => setTimeout(r, ...)) in tests.

6 — Run, then expand

Run the file. If green, add coverage for one more failure mode you initially skipped. If red, the test caught a real bug — flag it to the user before "fixing" the test.

Patterns and anti-patterns

✅ Do:

• One concept per test. If you need && in the name, split it.
• Use table-driven tests for many similar cases:

  test.each([
    ['30s', 30000], ['5m', 300000], ['1h', 3600000],
  ])('parses %s as %i', (input, expected) => {
    expect(parseDuration(input)).toBe(expected)
  })

• Test error messages too — they're part of the contract for humans.
• Keep test setup local to the test or the describe block. Avoid module-level mutation.

❌ Don't:

• Don't test private methods directly. Test the public contract.
• Don't share state between tests. Reset before each.
• Don't test the framework. expect(typeof fn).toBe('function') adds nothing.
• Don't snapshot-test arbitrarily. Snapshots are for stable serialized output; otherwise they just rubber-stamp regressions.

Example invocation

User: "Write tests for src/utils/slugify.ts." (file exports slugify(s: string, options?: { maxLen?: number }): string)

1. Detect: Vitest, __tests__/ adjacent.
2. List failure modes: empty string, all spaces, unicode ("café" → "cafe"?), trailing dashes, very long input (maxLen kicks in), already-slugified input (idempotent?), multi-byte emoji.
3. Write __tests__/slugify.test.ts with three describes: happy path (8 cases), boundary (4 cases), invalid (2 cases).
4. Add table-driven test for unicode normalization.
5. Run: npm test slugify — 14/14 green.
6. Spot during writing: slugify(" ") returns "-", which the user probably didn't intend. Flag it; ask whether to write the test for current behavior or the expected behavior.

See also

• code-auditor — to find the untested edge cases before you start writing
• browser-qa — when the behavior under test is a user flow, not a function
• refactor-master — if you need to refactor to make the code testable