testing-skills-with-subagents

Run realistic pressure scenarios without the skill first and record the agent's exact failures, then write the skill to counter those failures and re-run the same scenarios. Keep iterating until the agent complies under combined pressures like time limits, sunk cost, and authority.

It maps TDD onto documentation: RED runs scenarios without the skill to watch the agent fail, GREEN writes the minimal skill that makes the agent comply, and REFACTOR closes newly discovered rationalization loopholes with explicit negations and red-flag lists.

Good scenarios combine three or more pressures such as deadlines, sunk cost, authority, exhaustion, and social pressure, and force a concrete A/B/C choice with real constraints and file paths. Academic open-ended prompts let agents recite the skill instead of acting.

Skip pure reference skills like API docs or syntax guides, skills with no rules to violate, and skills agents have no incentive to bypass. Pressure testing only matters for discipline-enforcing skills with compliance costs.

Agents generate new rationalizations such as following the spirit not the letter or keeping code as reference. Use meta-testing to ask the agent how the skill could be clearer, then add explicit counters for each loophole and re-test.