| name | write-testing |
| description | Add or improve test coverage. ALWAYS apply this skill when the user asks to add tests, increase coverage, write a regression test, or improve an existing test suite — including for a feature, fix, or audit "Needed" item. Do not test implementation details, skip the assertion-flip proof, or modify production code to make a test pass without spec authorisation. Skip this skill if the task is implementing or fixing code itself — tests are part of those deliverables, not a separate phase. |
Skill: write-testing
Purpose
Tests that pass when commented out, tests that test internals, and tests that bundle unrelated assertions are net-negative — they take maintenance and don't catch bugs. This skill is the discipline that keeps tests honest: behaviour-focused, one-reason-to-fail, robust against refactors that preserve behaviour.
Tests are specifications by other means. A good test fails for one reason, exercises the behaviour the user-or-caller cares about, and survives a reasonable refactor of the implementation. A test that's testing the implementation is a maintenance burden disguised as a safety net.
Project context (the AGENTS.md contract)
Resolves project commands via the consuming repo's AGENTS.md — Commands > Test, Commands > Validation. If AGENTS.md is missing or an entry is undefined, ask the user which command to run before declaring a new test "passing" — do not guess.
Core rules
1. Test behaviour, not implementation
Exercise the public surface. Do not reach into private methods, module-private state, or other internals. A test that breaks when the implementation refactors but the behaviour is preserved is a bad test.
2. One test, one reason to fail
Each test asserts one specific behaviour. Bundling multiple unrelated assertions into one test means a failure tells the developer "something broke" without saying what. Split.
3. Flip the assertion to prove the test means something
After writing each test:
- Flip its assertion (or comment out the production code path it exercises).
- Run the test → must fail.
- Restore → must pass.
Paste a representative sample of the failing-then-passing transition into the self-review. A test that passes when the assertion is flipped tests nothing.
4. Place tests per the project's conventions
Don't guess test placement. If the project has a testing-layout convention (file naming, directory placement, runner config), follow it. If multiple conventions exist, pick the one closest to the code under test and document the choice.
5. Coverage numbers are a smell, not a target
A covered line that's poorly tested is worse than an uncovered one — it gives a false sense of security. Targeting "85% coverage" produces shallow tests. Focus on what behaviour deserves a test, not on the percentage.
6. Failure messages must be useful
When a test fails, the failure message should tell the developer what behaviour broke. Use descriptive test names, descriptive assertion messages where the runner supports them. "AssertionError: expected true" tells the developer nothing.
7. Tests must be deterministic
Avoid ordering dependencies, timing dependencies, shared mutable state, network calls without sandboxing, randomness without seeding. Flaky tests train developers to ignore failures — the worst outcome.
8. Do not modify production code to make tests easier
Unless the spec authorises it. The test exists because of the production code's behaviour; bending the production code to suit the test inverts the dependency. If the code is genuinely hard to test, that's a finding (promote to an audit), not licence to weaken the production code.
What does not belong
- In a test: assertions on internal state, multiple unrelated behaviours bundled, ordering dependencies, time-of-day dependencies.
- In a testing task: chasing coverage percentages, refactoring production code "while you're here" to make testing easier (without spec authorisation).
Anti-patterns
- Testing implementation details (private methods, internal state)
- Bundling assertions across unrelated behaviours
- Chasing coverage numbers for their own sake
- Tests that pass even when commented out
- Modifying production code to make tests easier (unless spec authorises)
- Flaky tests left flaky ("it usually passes")
Bundled resources
references/task-template.md — a fillable testing-task template with coverage-gap block, test-cases table (behaviour / inputs / expected outcome / reason this test exists), test-placement table, progress checklist that includes the assertion-flip step, and a self-review hard gate covering behaviour-over-implementation, failure-mode clarity, placement, and robustness.
Copy it into your project's task file location, substitute the {{...}} placeholders, and fill it in as you work.
