apw-write-testlisted

Use this skill when the deliverable is *new test code* — whether you are writing tests first (TDD) or covering an already-written change (post-hoc). It exists to fight one specific failure: left alone, an agent writes tests that **codify whatever the code currently does** — assertions reverse-engineered from the implementation instead of derived from intended behavior. Those tests pass by construction, are never seen to fail, exercise only the happy path, and mock away the very logic under test. Green coverage that proves nothing. This is the testing-specific face of the model's **bias toward confirming its own output**. The antidote is structural: derive the expected behavior from the *requirement* before reasoning from the code, and trust no test until it has been *seen to fail*. That same spine works in both modes — in post-hoc mode it makes you behave as if the tests came first. This is a different job from deciding *what work to do* (`apw-plan`), proving a *finished* implementation meets its requirements (`apw-verify`, which consumes tests rather than authoring them), or general adversarial review (`apw-review`, of which Step 6 here is a narrow, test-only echo). Follow the steps below strictly and in order. ## Output style When this skill asks for a plan, report, summary, status update, or other chat-facing output, prefer bullet points or numbered lists over plain paragraphs. Keep each item concise and content-bearing. Omit language that only smooths tone, repeats con