How do I know if a Claude skill is bulletproof?

A skill is bulletproof when an agent follows the rules under maximum pressure, cites specific sections of the skill as justification, and explicitly rejects its own previous rationalizations.

What is the RED phase in skill testing?

The RED phase involves running a high-pressure scenario without the skill active to observe and document exactly how the agent fails or rationalizes ignoring best practices.

Why are pressure scenarios necessary for testing AI skills?

Agents often follow rules in academic tests but ignore them in real work. Pressure scenarios simulate constraints like deadlines or sunk costs to ensure the skill is followed when it matters most.

What is a rationalization table?

A rationalization table documents every specific excuse an agent provides for bypassing a rule, allowing you to refactor the skill with explicit counters for those specific loopholes.

Testing Skills with Subagents

Name: Testing Skills with Subagents
Author: NickCrew

byNickCrew

•

安全与测试

Applies Test-Driven Development (TDD) principles to Claude skill documentation through high-pressure scenario simulations.

This skill provides a rigorous framework for verifying that Claude skills and instructions actually work under real-world pressure. By employing a RED-GREEN-REFACTOR cycle—running baseline scenarios without the skill, writing instructions to address specific failures, and iteratively closing loopholes—it ensures that agents comply with rules even when faced with sunk costs, deadlines, or exhaustion. It transforms documentation from passive reference material into bulletproof operational guidance that resists rationalization.

主要功能

01Meta-testing framework to identify and fix unclear documentation through agent feedback

02Standardized RED-GREEN-REFACTOR cycle specifically for process documentation

03Rationalization tables for capturing and countering agent excuses verbatim

04Pressure scenario generation including Time, Authority, Economic, and Sunk Cost stressors

05Bulletproofing techniques including explicit negation, red flags, and foundational principles

067 GitHub stars

使用场景

01Testing if an agent will prioritize safety protocols during simulated high-pressure emergencies

02Verifying a new coding discipline skill before deploying it to a production workflow

03Strengthening documentation that agents frequently ignore or rationalize away during tasks

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