testing-skills-with-subagents

Test-Driven Development for skills involves running pressure scenarios without the skill (Red), writing the skill to address failures (Green), then closing loopholes agents might exploit (Refactor). This cycle validates agent compliance and prevents rationalization before deployment.

Rationalization prevention captures the exact excuses agents use to bypass skill instructions under stress—time pressure, sunk cost, authority conflicts—then explicitly counters them in the skill design. This makes skills bulletproof against agent workarounds.

Design realistic scenarios combining multiple pressures: time constraints, sunk-cost fallacy, authority conflict, or competing instructions. Run these against a subagent harness using scenario templates, observe baseline failures, and record agent rationalizations verbatim to refine skill instructions.

Use TDD when creating or editing any custom AI skill before deployment. It ensures your skill documentation is clear enough that agents cannot rationalize their way around it, and catches loopholes early through iterative subagent testing.

Yes. Apply pressure scenarios to existing skills using a subagent harness to identify where agents rationalize or deviate from instructions. Record these failures and rationalizations as baseline data, then refactor the skill to close each gap explicitly.

Baseline failure observation documents what agents do wrong and why they justify it without the skill (Red phase). Refactoring then strengthens the skill to block those specific rationalizations, ensuring agents cannot repeat the same excuse during Green phase validation.