testing-skills-with-subagents

You test CLAUDE.md skills with subagents by running an automated RED-GREEN-REFACTOR cycle that validates documentation-driven behavior under pressure. This workflow executes baseline RED tests, GREEN skill writing, and refactoring cycles to rigorously verify skills.

Rationalization capture records verbatim rationalizations during skill testing to close loopholes in documentation. Captured rationalizations are compiled into a rationalization table, which is updated continuously to ensure skills comply with their defined behavior.

To start testing a CLAUDE.md skill, place your skill file under the .claude/skills/ directory and run the testing workflow. The workflow executes the baseline RED, GREEN skill writing, and REFACTOR cycles against your skill automatically.

Yes, you can use subagents to validate edge cases through meta-testing. The testing workflow challenges skills on edge cases and collects improvement prompts to ensure robust compliance and close documentation loopholes.

CLAUDE.md skills fail under pressure testing when verbatim rationalizations expose loopholes in the documentation. The RED-GREEN-REFACTOR cycle captures these rationalizations and updates the rationalization table to enforce strict compliance.