The bar is ENGINEERING_STANDARDS.md's TDD section (referenced, not restated): every new public function, branch, and error path needs a test; ≥90% lines on new/changed modules is the default target; a test that can't run in the project's real runner doesn't count.
A failing test against existing code is a finding, not noise. If a test you write to cover a "should never happen" case actually fails, you found a real bug or an already-broken regression — report it with repro evidence and do not rewrite the test to match the broken behavior. Fix the test's scope only when the spec was wrong, never to paper over a real defect without telling the user.
Every test you write or touch follows the Google practices in Step 5 — this is not optional flavor text, it's the acceptance bar for the tests this skill produces.
Smallest correct test level. Default to unit. Escalate to integration or E2E only per the pyramid rule in Step 4 — don't reach for Playwright to cover a pure function.
Ground every gap in evidence: file:line of the untested logic, the specific input class missing coverage, and (for regressions) the commit that introduced the bug.

Step 1: Scope

If $ARGUMENTS gives paths/globs, scope to those. If it says --full, scope to the whole repo. Otherwise default to the current branch's diff: git diff --name-only <base>...HEAD plus git status --porcelain (same base-branch detection as /qa-full Step 1).
Detect the test framework and runner from the project (package.json scripts, pytest.ini/pyproject.toml, go test, bats, *.csproj, Gemfile). If ambiguous, ask rather than guess — writing tests against the wrong runner is worse than asking one question.
Record scope, runner, and command at the top of your working notes.

If the scope is empty, say so and stop — there's nothing to cover.

Step 2: Find undertested complex logic

For each in-scope file, flag functions/methods that are "complex" by any of:

Multiple branches — 2+ if/switch/ternary paths, especially ones that return different values or take different side-effecting actions.
Loops with accumulation or early exit — aggregation, search, dedup, pagination.
State machines / status transitions — anything with an enum of states and rules about which transitions are legal.
Calculations with domain rules — money/currency, dates/timezones, permissions/authorization, rate limiting, parsing/serialization, validation.
Recursive functions.
Anything already flagged by /db-optimize, /perf-profile, or a prior /code-review as high-risk — risk-driven prioritization: spend the budget here first, not on trivial getters.

For each, grep the test tree for a reference to that function/class name. No reference (or only a shallow "doesn't throw" smoke test) ⇒ gap.

Step 3: Enumerate edge and corner cases per gap

For each gap from Step 2, work out the input-partition classes relevant to that logic — don't apply a rote checklist blindly, reason about what this function actually branches on. Common classes to check against:

null / undefined / missing-key vs. present-but-empty ("", [], {}, 0)
boundary values: zero, negative, max/min of the type, off-by-one at a loop/array edge
single-element and duplicate-element collections
malformed / unexpected-shape input (extra fields, wrong type, truncated)
concurrent or repeated calls (double-submit, idempotency)
unicode/encoding edge cases in string handling
timezone/DST boundaries and leap years in date logic
floating-point precision in money/percentage math
combinations of two+ conditions that individually pass but together hit an unhandled branch (the actual "corner" case, not just one axis at a time)

Only list classes that are reachable through the function's real callers — don't invent tests for input shapes the type system or a validator already rules out upstream.

Step 4: Regression coverage

git log the in-scope files for commits whose message matches fix|bug|regression|hotfix|revert (case-insensitive).
For each match, check whether that commit (or a nearby one) added or modified a test. No test touched ⇒ regression gap: the bug could reappear silently.
For each regression gap, write a test that reproduces the original failure mode (read the pre-fix diff to know what broke) and pins the fixed behavior — the classic "this test would have caught it" test.

Step 5: Decide the test level — pyramid, not instinct

Unit (default). The logic's correctness can be verified with real objects/pure functions and no I/O. This is nearly everything from Steps 2–4.
Integration. Escalate only when the real risk lives at a boundary a unit test would fake away — an actual SQL query's behavior, a real filesystem/queue interaction, a multi-module contract. Prefer a real dependency or an in-process fake/hermetic server over mocking the boundary (see rules/real-fake-mock-hierarchy.md); reach for a mock only when no hermetic option exists.
E2E. Reserve for a small number of critical golden-path user journeys where the risk is the wiring between already-unit-tested pieces, not any single piece's logic. Use /playwright for browser E2E — don't hand-roll a driver here.
State which level you chose per gap and why in one line — this is the artifact a reviewer checks when they ask "why is this an integration test."

Step 6: Google testing best practices (apply to every test you write)

These are Google's Testing on the Toilet rules that close real gaps in most test suites. Each is a short standalone file under rules/ — a rule plus a concrete good/bad example — so you pull in the detail only for the rule that's actually in play, instead of re-reading all eleven every run. Skim the one-liners below; open a file when its situation applies to the test you're about to write. When you're already touching an existing test for a nearby reason, fix violations you see rather than adding a TODO.

Rule	Applies when
`real-fake-mock-hierarchy.md`	About to reach for a mock — check real/fake first
`dont-mock-types-you-dont-own.md`	The dependency is a third-party type, not your own
`verify-state-changes-only.md`	About to `verify()`/assert-called-with on any call
`state-over-interaction-testing.md`	Choosing what the assertion should check
`avoid-change-detector-tests.md`	The test's structure is starting to mirror the code's branches
`know-your-test-doubles.md`	Picking which kind of double to use
`limit-mocks-per-test.md`	The test already has 2+ mocks
`damp-not-dry.md`	Tempted to extract a shared fixture/setup helper
`test-behaviors-not-methods.md`	Naming a test or deciding how many tests one function needs
`inject-dont-hardcode-statics.md`	The gap traces back to a singleton/static/`Date.now()`
`descriptive-names-clean-data.md`	Naming a test or building its fixture data

Step 7: Write the tests

For each confirmed gap (Steps 2–4), at the chosen level (Step 5), following the Step 6 rules/ checklist:

Write the test. Run it. If it's testing genuinely new behavior it should fail first (red) against a stubbed-out expectation, or pass immediately because the behavior already exists and you're only closing a coverage gap — either is fine, but run it and read the actual result, don't assume.
If a new test fails against unmodified production code, stop and report it as a bug/regression finding (see Hard rules) instead of silently adjusting the assertion.
Apply the test file changes (this skill writes and commits to the working tree).

Step 8: Run the full suite and report

Run the project's real test command. Then report:

# Test coverage — <scope> @ <date>

Runner: <command>   Files scoped: N   New tests added: M

## Gaps closed
- <file:line> <function> — <input class(es) covered> — <unit|integration|e2e> — <test file:line>

## Regressions pinned
- <original bug commit> — <file:line> — reproduces: <what broke> — <test file:line>

## Bugs found while writing tests (not fixed here)
- <file:line> — <test that exposes it> — <why it's a real defect, not a bad test>

## Best-practice fixes applied to existing tests touched
- <file:line> — <which rules/ file> — <what changed>

## Suite result
<pass/fail summary from the real run>

Anti-patterns (do not do)

Writing a mock for a dependency you own instead of just using the real thing or a fake (rules/real-fake-mock-hierarchy.md).
Mocking a third-party type directly instead of wrapping it (rules/dont-mock-types-you-dont-own.md).
verify() on a getter/query call (rules/verify-state-changes-only.md).
A test that re-implements the production branching logic to compute its expected value — that's not a test, it's a mirror (rules/avoid-change-detector-tests.md).
Reaching for Playwright/E2E to cover a pure function that a unit test covers in milliseconds (Step 5).
"Fixing" a failing test by loosening its assertion to match a real bug, without telling the user (Hard rules).
Padding the report with trivial getter/setter coverage while a genuinely complex branch stays untested — risk-driven prioritization (Step 2).

Related commands

/tdd — the write-code-first-with-a-failing-test loop while implementing new code. /test-coverage runs after code exists, to close gaps in what already landed.
/qa-full Step 9 — the coverage check in the pre-ship pipeline; it audits for gaps, runs /test-coverage to close them, then re-runs the suite. New public surface still untested afterwards is a ship blocker.
/playwright — the E2E harness this skill hands off to when Step 5 decides a gap needs browser-level coverage.
/code-review — correctness review of the diff; /test-coverage is the complementary "does the diff have adequate tests" pass, and can be run right after a /code-review pass identifies risky logic.
/debug — when Step 7 surfaces a real bug via a new test, use /debug to root-cause and fix it (this skill reports the bug, it doesn't fix production code).

test-coverage

/test-coverage

Hard rules