What is skill testing with subagents?

It is the application of Test-Driven Development (TDD) principles to AI instructions, using a subagent to verify if a skill successfully prevents specific failures or bad habits.

Why do I need to run a 'RED' phase for a skill?

The RED phase establishes a baseline by watching the agent fail without the skill. Without this, you cannot know if the skill actually prevents the targeted behavior or if the agent was naturally complying.

How do pressure scenarios work in AI testing?

Pressure scenarios combine constraints like time limits, authority figures, and sunk costs to create realistic situations where an agent would be tempted to bypass its instructions.

What happens in the REFACTOR phase of skill testing?

During refactoring, you identify any remaining 'rationalizations' the AI uses to ignore the skill and update the documentation with explicit negations and red flags to close those loopholes.

Skill Testing with Subagents

Name: Skill Testing with Subagents
Author: ed3dai

byed3dai

•

セキュリティとテスト

Verifies Claude skills and process documentation using a Red-Green-Refactor cycle to ensure compliance under pressure.

This skill applies the rigorous Red-Green-Refactor cycle of Test-Driven Development (TDD) to AI process documentation and custom Claude skills. By running baseline 'Red' tests to observe where agents fail naturally, iteratively writing 'Green' instructions to address those specific failures, and 'Refactoring' to close logic loopholes and rationalizations, it ensures that your AI workflows remain bulletproof even under high-pressure scenarios like tight deadlines or complex constraints. It provides a structured framework for creating 'pressure scenarios' that force agents to choose between following the rules or taking easy shortcuts.

主な機能

015 GitHub stars

02Red-Green-Refactor workflow for AI instructions

03Multi-model verification (Sonnet/Haiku cross-testing)

04Rationalization capture and loophole plugging

05Pressure-test scenarios (time, authority, sunk cost)

06Meta-testing for instruction clarity

ユースケース

01Validating and debugging custom Claude Code skills before deployment

02Enforcing strict adherence to TDD or security protocols in AI agents

03Hardening mission-critical process documentation against AI rationalization

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add ed3dai/ed3d-plugins-testing testing-skills-with-subagents

For use in Claude.ai and ChatGPT

Download Skill