testing-skills-with-subagents

You test AI agent skills by applying Test-Driven Development (TDD) principles, defining pressure scenarios, capturing agent failures verbatim, and iteratively refining documentation to close loopholes where the agent rationalizes non-compliance.

The TDD cycle for prompt engineering involves running baseline scenarios to expose failures (RED), writing minimal skill documentation to address those failures (GREEN), and iteratively refining to close rationalization loopholes (REFACTOR).

You validate agent documentation by pressure testing it with realistic scenarios involving time constraints, sunk costs, and authority pressures to ensure the skill consistently enforces desired behaviors and prevents deviations.

You should use TDD for skill testing when your skills enforce discipline, have compliance costs, or could be rationalized away, ensuring they remain robust against agent excuses before deployment.

Testing skills with subagents requires designing realistic pressure scenarios and capturing failures verbatim, which means limitations include the overhead of iteratively refactoring documentation to close every loophole exposed during baseline scenario runs.