testing-skills-with-subagents

Testing agent skills with a RED-GREEN-REFACTOR cycle involves running baseline failure tests without the skill, validating the implementation against those failures, and refactoring to close loopholes with explicit negations.

The RED phase in skills testing is the baseline testing stage that reveals agent failures and vulnerabilities without the skill applied, exposing exactly where the documentation needs enforcement.

You validate agent documentation before deployment by running structured testing steps and meta-testing scenarios that exercise baseline failures, skill implementation, and regression checks to ensure robust behavior.

Yes, testing skills with subagents works with CLAUDE.md by running a full RED-GREEN-REFACTOR cycle on your skill examples to verify agent behavior before deployment.

You need to capture rationalization during skill testing to document new agent workarounds, which enables targeted refactor loops that close loopholes with explicit negations and counters.