Testing Discipline

Overview

A good test pins the contract, uses concrete examples, names the scenario in domain language, and uses test data safe to show a customer. Run the checklist in order when writing or reviewing a test.

When to invoke

Invoke when you're about to:

• Write a new test (unit, integration, end-to-end, characterization)
• Name or rename a test method, test class, or test file
• Design a fixture, mock, stub, or shared test helper
• Choose test data — names, IDs, sample strings, sample numbers
• Decide what to assert in a test (exact value? structural property? both?)
• Schedule a long-running suite (soak, perf, cross-platform matrix)
• Hand off requirements between programmers and testers, or work on acceptance tests
• Review existing tests for quality, coverage gaps, or test smells
• Evaluate whether a test suite adequately covers a contract

Non-triggers — do NOT invoke for

• Running the existing test suite (npm test, pytest)
• Fixing a one-line broken assertion where the intent is unchanged
• Adding a docstring or comment to an existing test
• Reordering imports in a test file
• Updating a snapshot file when the underlying intentional change is already understood

If you're not sure whether a change counts as "writing a test," invoke anyway — the checklist is short and skipping it produces tests that pass forever while protecting nothing.

Test quality checklist

Run every step in order when writing a new test. When reviewing, use the same checks to evaluate quality.

1. Tests are part of the change, not optional polish. Software's only practical pre-deployment validation is execution under realistic conditions. A change without tests is an incomplete change. (Ford, 97/83.)
2. Assert the required behavior, not an incidental of the current implementation. Ask: what does the contract actually promise? For a 3-way comparator the contract is "negative if less, positive if greater, zero if equal" — not "exactly -1 or +1." Asserting ±1 nails an incidental and will go red the day someone returns -2. Tests that mirror code structure end up asserting that the code does what the code does. (Henney, 97/80.)
3. Be precise and accurate. Use concrete examples. "Result is sorted and same length" passes for [3,3,3,3,3,3] against an input of [3,1,4,1,5,9]. The full postcondition is sorted and a permutation of the input — but expressing that as a generic checker is often more code than the function under test. Prefer a concrete example pair: input [3,1,4,1,5,9], expected [1,1,3,4,5,9]. "Adding to an empty collection" is not "now non-empty" — it's "now contains exactly one item, and that item is X." (Henney, 97/81.)
4. Write the test for the next person who has to read it. A good test is documentation. Make three parts visible in this order: the context/preconditions, the call into the system, the expected result. Hide trivia behind named helpers (Extract Method) so the reader sees the scenario, not the scaffolding. Give the test a name describing the scenario and the entry point — Stack_pop_on_empty_throws reads better than test17. Then test the test: introduce a deliberate bug into the code under test on a private branch and verify the failure message tells you what went wrong. (Meszaros, 97/95.)
5. Choose test data that is safe in front of a customer. Placeholder names, fake company names, sample log strings, mock error dialogs, and seeded database rows have a habit of surfacing in screenshots, demos, leaked source, and production logs. Do not use real people's names you mean to mock, do not use band names or song titles as a private joke, do not write "don't click that again, you moron" as a placeholder dialog, do not use four-letter words as fake stock tickers. Use boring, obviously-fake, professional data. (Begbie, 97/25.)
6. Accept acceptance tests as requirements input. When acceptance tests exist for a behavior (from QA, from the customer, from a spec), use them as input to the work — they surface ambiguities cheaper than defect tickets. (Hufnagel, 97/60; Gregory, 97/92.)

Red Flags

These thoughts mean STOP — restart the checklist:

What "done" looks like

A single well-written test is done when all of the following are true:

• The assertion targets the contract, not an incidental of the current implementation.
• The expected result is concrete enough that a reader can check it by eye in under thirty seconds.
• The test name describes the scenario and the entry point in domain language.
• Context, action, and expected result are visible as three readable sections; scaffolding is behind named helpers.
• Test data (names, IDs, strings, log lines, error messages) is safe to appear in a customer screenshot.
• You verified the test fails for the right reason — by introducing a deliberate bug on a private branch and reading the failure message.

If any box is unchecked, the test is not done. Either finish, or delete it and start over.

Principles in this skill