testing-skills-with-subagents_skill

Overview

This skill formalizes test-driven development for skills by applying a RED-GREEN-REFACTOR cycle to process documentation. It helps you discover how agents naturally fail, write minimal rules to prevent those failures, and iteratively close loopholes until compliance is reliable. Use it to ensure rules survive realistic pressure and rationalization attempts.

How this skill works

Run realistic pressure scenarios without the skill (RED) to capture exact failures and rationalizations. Implement a minimal skill addressing those failures (GREEN) and pressure-test it with weaker models to ensure instructions are explicit. When agents invent new rationalizations, refactor the skill to negate each excuse and re-test until agents consistently follow the rule.

When to use it

• Creating or editing discipline-enforcing skills (TDD, compliance, safety).
• Before deploying skills that impose time or workflow costs on users.
• When agents could rationalize bypassing rules under pressure.
• To validate that documentation prevents real-world violations.
• When you need reproducible tests that prove a skill is followed.

Best practices

• Always run RED first: observe and document agent failures verbatim before writing the skill.
• Design pressure scenarios combining 3+ pressures (time, sunk cost, authority, exhaustion).
• Test GREEN on a weaker model tier so instructions must be explicit to pass.
• For each new rationalization, add explicit negation, a rationalization table entry, and a red-flag.
• Keep the skill minimal—address observed failures first, then iterate to close loopholes.

Example use cases

• Validating a TDD-enforcing skill so agents delete pre-written code and restart with tests.
• Ensuring a release-checklist skill prevents last-minute unsafe hotfixes under management pressure.
• Bulletproofing a safety rule that agents might rationalize away in emergency scenarios.
• Hardening a code-review policy skill so agents cannot bypass required approvals.
• Testing documentation that enforces a strict experiment protocol against pragmatic workarounds.