create-tests

Imported from stoftot/RealJapanese (.agents/skills/create-tests/SKILL.md). Install upstream with npx skills add stoftot/RealJapanese --skill create-tests. Copyright stays with the author.

Create durable regression coverage

Consume the temporary handoff from plan-tests, or an equally clear request specifying behavior and expected results. This skill owns coverage implementation, not independent requirements or test-design decisions.

Create and check coverage

1. Inspect existing test organization, neighboring cases, frameworks, fixtures, helpers and conventions through PROJECT_MAP. Reuse established executable checks as well as framework-based tests; do not add a test framework without a concrete need.
2. Implement the planned automated cases at the selected level. Update existing tests where appropriate instead of duplicating them. Assert observable outcomes with deterministic, isolated data; preserve real catalogs and user progress.
3. Run created/modified tests where practical using canonical repository commands and applicable adapters. Inspect results and correct test-implementation errors. Report unavailable execution with its reason; writing a test does not prove it passes. Broader completion validation belongs to validate-change.
4. Create/update the planned manual specifications in the established catalog. Search for equivalent cases before creating files. This project's location and format are owned by tests/manual; group by coherent feature/domain and preserve useful existing cases and stable IDs.
5. Return changed test paths/case IDs, actual execution results and unresolved limitations to the calling workflow for completed-change validation.

Production-code boundary

Do not change production code merely to force a planned test to pass. If evidence suggests incorrect production behavior, conflicting requirements, ambiguity in the plan or architecture that prevents reasonable testing, return the evidence to the capable primary model. Use plan-tests for design ambiguity, debug-problem for unexplained failures and implement-change for an understood production fix. Do not weaken assertions to hide a failure or invent expected behavior.

Model routing

When delegation is available and worthwhile, prefer cheaper/faster subagents for clearly specified cases, variants following established patterns, structured manual descriptions and repetitive scaffolding. Give each a bounded file scope, the plan, relevant conventions and expected execution evidence; review the result.

Escalate substantial test architecture, unclear requirements, conflicts between expectations and implementation, or important cross-component reasoning to the capable primary model. If delegation is unavailable or costs more than it saves, continue locally with the same handoff and production-code boundary.