testing-skills-with-subagents

Test-driven development for skills uses RED-GREEN-REFACTOR cycles with subagent-based testing to expose and close rationalization loopholes. RED tests run without skill access to establish baseline behavior, GREEN tests verify skill effectiveness, and REFACTOR iterations tighten prompt wording to eliminate bypass routes.

Pressure scenarios combine realistic stressors—time constraints, sunk costs, exhaustion—to rigorously test whether agents rationalize away skill instructions under stress. They systematically reveal behavioral gaps that benign testing misses.

Yes. Define baseline and skill-enabled test prompts from your scenario specifications, run sandboxed subagent evaluations with deny-path restrictions, log results in structured format, and iterate until prompts are bulletproof against agent workarounds.

Extract Instructions for Sub-Agent into baseline-prompt.md and green-prompt.md files, create scenario.md-driven test specifications, configure fixture worktrees for isolated execution, and set up per-skill jsonl logging directories to capture compliance results.

Isolate baseline and green prompts in separate files, use sandboxed execution environments with deny-path restrictions, and maintain independent logging per test phase to prevent cross-contamination and preserve reproducibility.

RED tests establish baseline agent behavior without skill access, exposing natural rationalization patterns. GREEN tests measure skill effectiveness by running the same scenarios with skill instructions enabled, quantifying the behavioral delta.