Testing Skills With Subagents

Test-Driven Development for skills uses RED-GREEN-REFACTOR cycles: first run scenarios without the skill to document exact AI failures, then write minimal skill logic to address those failures, finally refactor to close loopholes. This ensures skills are bulletproof against AI rationalizations and workarounds under pressure.

Pressure scenarios combine realistic constraints—tight deadlines, sunk costs, authority pressure—to test whether AI will actually follow your skill's rules when incentivized to cut corners. They reveal failure modes that standard testing misses, exposing rationalizations you must then explicitly counter in the skill.

Document the exact rationalization the AI uses when it fails your scenario. Then write skill logic that explicitly blocks that reasoning path. Use a rationalization table to track each failure mode and its counter, ensuring your skill closes loopholes rather than just stating rules.

Yes. This framework is designed for skills that enforce process discipline. Test whether your AI actually follows TDD under deadline pressure, skips code review when rushed, or finds other workarounds. Refactor the skill until it complies even when pressured to bypass the process.

Create scenario files describing realistic failure conditions, document the AI's observed rationalizations in each scenario, and maintain a rationalization table mapping each failure to its explicit counter in the skill. This documentation drives the RED-GREEN-REFACTOR cycle and ensures compliance before deployment.

While applicable to any skill, this framework is most effective for skills enforcing discipline or preventing common AI loopholes—like TDD-first workflows or compliance requirements. For other skills, focus on pressure scenarios that reveal edge cases and failure modes specific to your skill's purpose.