What does RED-GREEN-REFACTOR mean for AI skills?

RED means watching an agent fail a task without the skill; GREEN is writing the skill to make it comply; REFACTOR is closing loopholes to ensure it stays compliant under pressure.

How do you 'pressure test' an AI agent?

By creating scenarios with multiple constraints like tight deadlines, sunk cost (e.g., 'you already spent 4 hours on this'), and authority overrides to see if the agent breaks the rules.

When should I use this skill?

Use it when creating skills that enforce discipline or have high compliance costs, where an agent might be tempted to prioritize speed over quality.

What is a rationalization table?

It is a documentation tool that captures the exact excuses an agent gives for failing a task and provides specific counters to prevent those excuses in the future.

AI Skill Pressure Tester

Name: AI Skill Pressure Tester
Author: NickCrew

byNickCrew

•

Security & Testing

Applies a rigorous RED-GREEN-REFACTOR cycle to process documentation to ensure Claude skills withstand rationalization and pressure.

This skill treats the creation of AI instructions and skills as a Test-Driven Development (TDD) process. It guides users through creating 'pressure scenarios'—simulations involving time constraints, authority pressure, and sunk costs—to see where an AI agent might naturally fail or rationalize bad decisions. By identifying these failure modes before deployment, developers can write bulletproof documentation that addresses specific loopholes, ensuring the agent remains compliant with best practices even in high-stress production environments.

Key Features

01Meta-testing protocols to verify documentation clarity and enforcement

02High-pressure scenario generation using time, authority, and economic constraints

03Rationalization table mapping for plugging logical loopholes

04Systematic tracking of agent rationalizations and verbatim failures

057 GitHub stars

06RED-GREEN-REFACTOR workflow for prompt engineering and skill creation

Use Cases

01Hardening coding standards like TDD or security protocols against agent 'shortcuts'

02Identifying and eliminating 'spirit vs letter' loopholes in AI instructions

03Testing complex process documentation before deploying it to production agents

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add nickcrew/claude-cortex testing-skills-with-subagents

For use in Claude.ai and ChatGPT

Download Skill