mintuz/.dotfiles/create-verification-skill

Create a verification skill

Create one shared verify-<project> recipe another agent can execute cold. Authoring success and reproducibility are separate gates.

1. Discover the target. Read project instructions to find the canonical skill source and registration, supported platforms, startup/readiness, identity checks, fixtures, authentication and existing interaction harness. Use that source and the host's available tools; maintain one recipe rather than divergent harness copies. Select one current critical user journey. When refreshing, inventory and preserve every supported journey; repair stale commands, locators and prerequisites without silently reducing coverage.
2. Establish identity and isolation. Record revision/build, environment and fixture. Identify owned instances and scratch state; avoid driving a shared user instance. For mobile targets, resolve scheme, bundle, runtime/device, signing and relevant entitlements. Use existing platform recipes when available: xcode-dev-loop for Apple build/install/run identity, swift-testing for test authoring, and eyes for web visual capture. A missing tool or launch prerequisite leaves a narrow diagnosis and draft, not permission to repair unrelated product or environment problems.
3. Save an executable recipe. Use real project commands and stable selectors, with runnable helper invocations if needed. Include:

• Launch: documented prerequisites, isolated fixture setup, usable readiness and target identity.
• Doctor: read-only checks of build, ownership, readiness and authentication when a run looks wrong.
• Drive: the actual public user journey, its preconditions, actions and expected result. An internal setter cannot replace the claimed boundary. Name device-only or external boundaries such as Siri, widgets, notifications, health data and purchases.
• Evidence: action, resulting state and material side effects, with identity and durable paths outside disposable scratch state. Observe what dry-run/test modes skip; their names do not prove safety. Use mocks only at an established isolated boundary and qualify the resulting proof.
• Cleanup: on success and failure, release only resources created by this run. Verify current resource ownership before termination; a reused PID or process name is insufficient. Preserve and confirm proof artifacts after cleanup.

1. Execute, then cold-replay. Run launch → doctor → journey → evidence → cleanup. Save corrections. Reset only owned fixtures, then replay the saved recipe from a fresh agent context using only its documented prerequisites. The authoring run cannot substitute for this replay. Resolve missing setup in the recipe and replay again; failed or unavailable replay leaves reproducibility unproven.
2. Hand over the proof. Link the saved recipe and retained evidence, report both execution and cold-replay outcomes, and identify uncovered journeys and unavailable boundaries. Keep build, unit, simulator, physical-device, browser and deployed evidence distinct. A stand-in leaves target-specific setup unverified. Verify representative replay separately in each claimed harness before claiming cross-harness portability.