testing-skills-with-subagents

Apply RED-GREEN-REFACTOR testing by first running baseline tests without your skill to capture exact failures, then iteratively implement minimal updates and validate through test cycles. This TDD methodology reveals whether your skill maintains discipline when agents face time constraints, sunk costs, or authority pressure.

RED-GREEN-REFACTOR is a testing pattern: RED watches your skill fail without implementation, GREEN adds minimal code to pass, REFACTOR improves while maintaining test success. For skills, this means documenting agent failures under pressure, building the skill constraint, then refining to close loopholes.

Combine time constraints, sunk costs, and authority pressure into realistic test cases. Run these without your skill to observe baseline rationalizations agents use to rationalize skipping compliance, then implement your skill and validate it prevents those shortcuts through subagent test harness execution.

Yes. This methodology targets skills designed to enforce discipline and prevent rationalization under pressure. It's particularly effective for compliance-heavy skills like code review, security checks, or approval gates where agents might skip steps when rushed.

Record verbatim rationalizations—the exact reasoning agents use to justify circumventing your skill's requirements. These become your RED test cases that your implementation must prevent, ensuring the skill closes specific loopholes rather than addressing vague concerns.

Yes. The Skill uses a subagent test harness to enforce RED-GREEN-REFACTOR cycles, generating 3+ pressure scenarios and validating your skill implementation against them. This automated validation ensures your skill performs consistently across realistic stress conditions.