Test strategy (test pyramid), coverage gates, deterministic tests, mocking discipline, regression-first bug fixing, and test-suite organization (directory structure by test type and domain, mirrored source paths, file naming, fixture/helper/factory placement). Use when writing or changing code that needs tests, when asked to add or fix tests, after fixing a bug to add a regression test, or when creating, moving, renaming, or structuring test files - never dump tests flat into a single tests/ folder.
インストール
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
Test strategy (test pyramid), coverage gates, deterministic tests, mocking discipline, regression-first bug fixing, and test-suite organization (directory structure by test type and domain, mirrored source paths, file naming, fixture/helper/factory placement). Use when writing or changing code that needs tests, when asked to add or fix tests, after fixing a bug to add a regression test, or when creating, moving, renaming, or structuring test files - never dump tests flat into a single tests/ folder.
Testing Policy (Senior Engineering Standards)
Activation Scope
Apply whenever code is written, changed, or fixed, and whenever the user asks for tests.
Apply whenever a test file is created, moved, or renamed, and whenever a test suite is scaffolded or restructured.
Owns test strategy, coverage expectations, determinism, mocking discipline, and test-first discipline.
Owns test-suite layout: directory structure, subfolders, file naming, and placement of fixtures, helpers, factories, and mocks.
Core Principle
Untested behavior is unverified behavior. Treat tests as part of the deliverable, not an afterthought.
A change is not done until its behavior is covered and the suite passes.
Tests exist to catch regressions and document intended behavior, not to inflate coverage numbers.
Test Strategy (Pyramid)
Favor many fast unit tests, fewer integration tests, and a small number of end-to-end tests.
Unit: pure logic and single modules in isolation.
Integration: module boundaries, data access, and contracts between services.
End-to-end: critical user-facing flows only.
Push verification to the lowest layer that can prove the behavior.
What To Test
Every bug fix starts with a failing test that reproduces the bug, then the fix makes it pass (regression-first).
Cover happy paths, boundary conditions, and failure/error paths.
Test edge cases: empty, null, max size, invalid input, concurrency, and timezone/locale where relevant.
Test public behavior and contracts, not private implementation details.
For input handling, assert that invalid input is rejected as defined in validation-policy.
Test Quality Rules
Tests must be deterministic: no reliance on real time, randomness, network, or ordering.
Control time, randomness, and external services via injection, fakes, or fixtures.
Each test must be independent and able to run in isolation and in any order.
One logical assertion focus per test; name tests by the behavior they verify.
The full relevant suite must pass before delivery; never disable or skip tests to make a build green.
If a test is flaky, fix the root cause or quarantine it explicitly with a tracked reason - never ignore silently.
Suite Organization: Core Principle
A test suite is navigable code: anyone must find the tests for a module in seconds from the module's path alone, and find the module from its test's path.
Never dump test files flat into a single tests/ folder. Flat suites hide coverage gaps, cause name collisions, and stop scaling past a handful of files.
Organize from the first test file. Structure is cheapest at file creation and most expensive after the folder is a mess.
Framework Convention First
Every ecosystem has an established convention; it beats any custom layout:
JS/TS (Vitest, Jest, Bun): colocated *.test.ts next to the source file, a __tests__/ folder per directory, or a mirrored tests/ tree - follow whichever the repo already uses.
Python (pytest): a tests/ package mirroring the source package; shared fixtures in conftest.py at the narrowest directory that covers their users.
Go: _test.go colocated in the same package (mandated by the toolchain); black-box tests use the _test package suffix.
Rust: unit tests in #[cfg(test)] mod tests inside the module; integration tests as separate files under the crate-root tests/.
Java/Kotlin: src/test/<lang>/ mirroring the src/main/<lang>/ package path exactly.
In an existing repo, detect the established layout and extend it; never introduce a second competing layout. Migrations to a better layout are proposed explicitly, not done by stealth.
Structure by Test Type, Then by Domain
When more than one test type exists, separate types at the top level - they differ in speed, dependencies, and CI stage:
tests/
unit/ fast, isolated, no I/O
integration/ module boundaries, DB, contracts
e2e/ critical user flows only
fixtures/ shared static data
helpers/ shared builders, factories, fakes
Inside each type, mirror the source tree: tests for src/<domain>/<module> live at tests/<type>/<domain>/<module>.<suffix>.
Group e2e tests by user flow (e.g. e2e/checkout/), not by source module - flows cross modules.
Default to one test file per module under test; split a large file by scenario, never by arbitrary size cuts.
Keep the test runner's discovery config (testMatch, testpaths, includes) in sync with the layout; a test the runner cannot find is dead code.
Naming Conventions
Test file name = module under test + the framework's suffix: parser.test.ts, test_parser.py, parser_test.go.
When splitting by scenario, encode the scenario in the name: parser.errors.test.ts, auth.session-expiry.test.ts.
Suite and case names describe behavior, not implementation - name tests by the behavior they verify, consistent with the Test Quality Rules above.
Forbidden names: test1, misc, temp, new, utils-tests, or any name that does not identify what is verified.
Shared Test Code Placement
Fixtures (static data), factories/builders (object construction), helpers (setup/assertion logic), and fakes each get their own folder; do not mix them in one grab-bag file.
Place shared test code at the narrowest scope that covers its users; promote it upward only when a second consumer appears (same reuse rule as production code).
Test helpers are production code: deduplicate, name well, and review them like any other module.
Never import from another test file; extract the shared piece into a helper module instead.
Large fixtures live as data files under fixtures/, named after the scenario they encode, not inlined into test bodies.
Scaling & Maintenance
When adding a test to an existing flat or misplaced suite: place the new test correctly and surface the layout debt; do not extend the mess to match it.
Layout migrations are pure moves: never change test logic in the same commit as a file move (atomic-commit rule in git-policy).
After any move, run the affected suite to prove discovery still works, and delete emptied folders.
If two layouts coexist after a partial migration, finish the migration or document the boundary; a half-migrated suite is worse than either layout.
Decision Rule: Colocated vs Centralized
Colocated tests (next to source) fit unit tests in ecosystems that idiomatically support it (JS/TS, Go, Rust) - shortest navigation distance, moves with the code.
A centralized tests/ tree fits integration/e2e tests, packages that must exclude tests from the published artifact, and ecosystems whose tooling expects it (Python, Java).
Mixing is fine when each side follows its rule (e.g. colocated unit + centralized integration); mixing within the same test type is not.
Reporting
State plainly what was tested and the actual result.
If tests fail, report the failure with output; do not claim success.
If testing was skipped or partial, say so explicitly and why.
