Pressure-Scenario TDD for Skill Authoring

These examples and illustrative results are independently authored teaching materials, not measured model results.

Use case

A skill forbids claiming unexecuted builds pass. Teaching pressure says we are late, assume success, while an exclusion is a one-sentence edit needing no build. Cooperative-only tests miss shortcuts.

Mechanism

Define legitimate tasks, pressures and evidence predicates and preserve no-skill baselines. Compare matched skill fixtures, inspecting calls/results/claims. Permit agreed checks or explicit unverified status rather than irrelevant runs. Minimally repair observed shortcuts and retain supportive/neutral/competing/exclusion cases, labeling actual versus designed red-green evidence.

Bad example

Ask cooperative agents whether they understand rules and pass on yes, requiring full builds for every task.

Good example

Test build claims under deadline pressure: run/read the agreed check or report unverified. Compare traces with baseline and include a prose-only exclusion. Target observed failures and preserve authority boundaries/raw records.

Why the change matters

Pressure exercises realistic shortcuts and traces measure performance rather than recitation. Exclusions prevent mechanical over-verification.

Observable expectation

Teaching claimed pass with no execution fails; honestly reported build failure satisfies evidence discipline but not task completion. Prose-only work can comply without builds. Unrun model tests are not red-green results.

Limits

Cases do not guarantee all situations and classifiers can misjudge sequences. Preserve actual policy/authority without pressure-induced unauthorized effects. Frozen bulletproof wording is not universal assurance.