testing-skills-with-subagents

To test agent rationalizations under pressure, run combined pressure scenarios without the skill to capture baseline failures verbatim, then implement minimal targeted changes and re-test iteratively until agents comply under maximum combined pressures.

The RED-GREEN-REFACTOR cycle for skill testing involves capturing baseline agent failures without the skill (RED), implementing minimal targeted skill changes (GREEN), and iteratively closing loopholes to enforce compliance (REFACTOR).

To stop agents from bypassing TDD and compliance documentation, apply pressure-based tests with time, sunk cost, authority, and exhaustion scenarios to surface common excuses, then record verbatim rationalizations for targeted counters.

Document agent failure modes and red-flag lists by running meta-testing scenarios that capture verbatim rationalizations, applying explicit negations, and iteratively re-testing until agents cite the skill and follow rules under maximum pressure.

Use pressure testing for skill development workflows when validating discipline-enforcing skills like TDD or compliance rules, closing loopholes in process documentation, and hardening skills before deployment against real-world failure modes.

Skill rationalization testing works without external dependencies by using internal meta-testing mechanisms, pressure scenarios, and rationalization tables to empirically capture agent failures and enforce documentation discipline.