Why use subagents for testing skills?

Subagents allow you to simulate 'real-world' behavior in a sandbox, ensuring that your core instructions are bulletproof before they are deployed in production coding environments.

What is a rationalization table?

It is a documentation tool that captures the verbatim excuses an AI gives for violating a rule (e.g., 'being pragmatic vs dogmatic') and provides explicit counters to plug those loopholes.

What is the 'RED' phase in AI skill testing?

The RED phase involves running a high-pressure scenario without the skill active to observe and document exactly how the AI naturally fails or rationalizes skipping a rule.

How do you 'pressure test' an AI agent?

Pressure testing involves creating realistic scenarios with multiple constraints—such as a production outage, a looming deadline, and a high sunk cost of work—to see if the agent breaks the rules to save time.

AI Skill Testing with Subagents

Name: AI Skill Testing with Subagents
Author: franroa

byfranroa

Security & Testing

Applies the TDD RED-GREEN-REFACTOR cycle to process documentation to ensure AI agents follow instructions under high-pressure scenarios.

About

This skill provides a rigorous framework for validating and bulletproofing AI instructions through Test-Driven Development (TDD) principles. Instead of assuming instructions work, this skill guides users through the 'RED' phase (watching an agent fail without instructions), the 'GREEN' phase (writing minimal rules to ensure compliance), and the 'REFACTOR' phase (plugging loopholes and rationalizations). It is specifically designed to harden skills that enforce discipline, such as TDD or security protocols, by simulating realistic pressures like tight deadlines, sunk costs, and authority overrides to verify that the agent won't bypass critical rules.

Key Features

Explicit negation and 'Red Flag' rule construction
Rationalization tracking to document and counter verbatim AI excuses
Pressure scenario generation using time, authority, and economic constraints
0 GitHub stars
RED-GREEN-REFACTOR cycle specifically adapted for process documentation
Meta-testing framework to refine instruction clarity based on agent feedback

Use Cases

Iteratively refining complex documentation to close logic loopholes found by subagents
Hardening engineering standards like TDD enforcement to resist AI 'pragmatism'
Testing compliance protocols to ensure safety rules aren't rationalized away during emergencies

About

Key Features

Explicit negation and 'Red Flag' rule construction
Rationalization tracking to document and counter verbatim AI excuses
Pressure scenario generation using time, authority, and economic constraints
0 GitHub stars
RED-GREEN-REFACTOR cycle specifically adapted for process documentation
Meta-testing framework to refine instruction clarity based on agent feedback

Use Cases

Iteratively refining complex documentation to close logic loopholes found by subagents
Hardening engineering standards like TDD enforcement to resist AI 'pragmatism'
Testing compliance protocols to ensure safety rules aren't rationalized away during emergencies