QA: prove it works (verify by default, tests on request)

Proving a change works is mandatory: though the proof is rarely a test suite. Default to verify (drive the real thing, watch what it does). Tests / TDD are a project choice: offer them, don't impose them. After a build, the honest close is: "Built it: want me to QA it? (I can verify it end-to-end, and add tests / do it test-first if you want.)"

Mode 1: Verify (the default; always do this)

1. Write the checklist before you look. Turn the expected behavior into a short list of criteria that are each individually checkable: one line per criterion, phrased so the answer is met or not met, never "looks fine". Write it before exercising anything, because a list written while looking is a list that describes what you found. If you can't say what correct looks like, you can't verify it.

   Prefer criteria a person could count or observe over ones needing an opinion: "a second submit while pending is rejected" beats "handles concurrency well." This list is what you report against in step 5, and what makes "it works" mean something.

2. Pick the lightest real check. Drive the actual thing over reasoning about it: run the app and click the flow, hit the endpoint, run the script/CLI, render the component. Reuse the project's run/dev command; reach for harness the project already has.
3. Happy path, then the edges that matter: empty, null, error, loading, zero/one/many, unauthorized, malformed, offline/slow (~/.mastermind/engineering/core/rigor.md). Observe actual output and state.
4. Check the invisible: typecheck, lint, build; console/network for errors; for UI, keyboard + focus, contrast, no layout shift/regression nearby.
5. Say what must NOT happen, not only what must. Half of a real check is a negative: no request was sent, the tokens were not cleared, the user was not bounced to login, no second write landed. A test that only asserts the visible message passes while the damage happens behind it.
6. Report with evidence: what you ran and what you observed (command output, response, screenshot). State confidence plainly. Couldn't run a check? Say so; never present unrun work as verified.

Driving a browser, when the thing under test is a page

"Click the flow" is the step most often skipped, because most sessions have no way to click anything. Work down this list and use the first rung you actually have:

1. No browser needed. A route test, an API call, a rendered component in a unit test. Fastest, and it is what most UI claims really rest on. Exhaust this before reaching further.
2. A deterministic driver the project already has: Playwright, Cypress, or the same exposed to your tool over MCP. Use this for anything that should be repeatable, because a written selector fails the same way twice and that is what makes it a test.
3. An agentic driver, browser-use being the current example: it opens the page, clicks and types from a natural-language task. Reach for it when writing selectors is the bottleneck and the question is exploratory: does a person get through this flow. It is MIT, it needs Python and its own model call, and it is never a requirement: it is a rung, not a dependency.
4. Nothing. Then say so. Report exactly which claims are unverified and why, and never let a screenshot you did not take stand in for a check you did not run.

An agentic driver reads a page the way a person would, so it finds what a selector-based test cannot: the button nobody can find, the flow that technically works. It also varies between runs, so it belongs in exploration and not in a gate.

Found a bug? Fix the root cause (or route to debug if it's not obvious), never suppress a symptom to make a check pass. Verify against the requirement, not the code you just wrote (a hostile eye).

Mode 2: Test-first / TDD (thorough by default; writing the files is the opt-in)

Writing a test suite is a heavy optional step: offer it, get a yes, then write. You're in Mode 2 only because the user asked for tests / test-first, or said yes to the offer after a build. Inside it, test the product fully and in detail: writing tests (even test-first) to prove behavior across the happy path and every edge is good QA, not overreach.

A test written for a bug must be proven against it: the revert-proof in debug phase 6 (~/.mastermind/skills/debug/SKILL.md) is the rule; it applies unchanged here.

Get the yes before the first file lands. Name where the tests would live and what they'd cover, "Want these as tests? They'd go in <path> and cover <happy path + the edges that matter>.": and write nothing to disk before it. Proportionality, not ceremony: if the project already has a suite and the change belongs in it, adding the case is the normal way to work: do it and say so; the gate is for starting a suite, or putting test files in a repo that has none.

Then Red → Green → Refactor:

1. Red: one small failing test stating the next behavior (for a bug: a test that reproduces it). Watch it fail for the right reason.
2. Green: the minimum code to pass. No gold-plating.
3. Refactor: improve the design while green; loop.

Test behavior/contracts, not internals (brittle implementation tests are worse than none). Test names read like the spec. Struggling to make a test pass cleanly? The design is probably wrong, listen to it.

Two checks for a test you wrote. Name the specific break it catches, then make that break by hand and watch it fail. A test that still passes against the broken code tests nothing. And a test that only fails when the code changes, not when it breaks, such as one asserting exact copy or a snapshot of internals, is a change detector: it costs every future edit and catches nothing.

The two ways a green suite lies. A tautological test recomputes the expected value the way the code does: expect(add(a, b)).toBe(a + b) passes whatever add does, so the expected value has to come from somewhere independent: a hand-worked example, the spec, a known-good output. A horizontally sliced suite tests a layer in bulk before anything runs end to end, which verifies imagined behavior, one thin vertical slice that actually executes is worth more than a wall of green.

And don't silently leave test files in the user's repo. They were agreed before they were written, so nothing arrives as a surprise, but close the loop: say what now exists and confirm it stays, "Tested it thoroughly; keeping these tests in <path> as coverage unless you'd rather I remove them." Match the project: if it already has a suite, add to it; if it has none, don't impose one: the files are the user's call, even though the testing wasn't optional.

Output

A plain verdict: works / doesn't: with the evidence and the edge cases exercised, plus any gaps you couldn't cover (and why). If tests were written: note what they cover and confirm they stay, never write or persist a test suite the user didn't agree to first. If the project's cycle-report preference is on (.mastermind/prefs.md), also run the report skill to save a durable write-up; default off, so by default the verdict stays in chat.