Is this skill only for testing TDD instructions?

No, while it uses TDD principles as a framework, it can be used to test any skill that enforces discipline, follows specific coding patterns, or might be bypassed by an agent for the sake of speed.

What is a pressure scenario in skill testing?

A pressure scenario is a test case that includes stressors like deadlines, budget constraints, or authority figures to see if the AI will break its rules to achieve a goal faster.

Why do I need to test skills without the skill enabled first?

The 'RED' phase establishes a baseline. You must confirm that the AI actually fails or takes the wrong path without the instruction to ensure your skill is addressing a real behavior rather than a hypothetical one.

How do I handle an agent that keeps finding excuses to ignore a skill?

Use the REFACTOR phase to document the specific 'rationalization' verbatim and add explicit negations or a rationalization table to the skill to block those specific excuses.

Skill Testing with Subagents

Name: Skill Testing with Subagents
Author: seanGSISG

byseanGSISG

0•

安全与测试

Applies a rigorous TDD cycle to verify and harden Claude Code skills against pressure and rationalization.

This skill provides a systematic framework for hardening Claude Code instructions through a rigorous RED-GREEN-REFACTOR cycle. By using subagents to run realistic pressure scenarios—incorporating factors like time constraints, sunk costs, and authority overrides—it identifies where an agent might bypass rules and provides specific patterns for plugging those loopholes. It transforms skill creation from a guesswork-based process into a verifiable engineering discipline, ensuring that your custom instructions actually hold up under the stress of real-world development and prevent agents from 'rationalizing' their way out of following protocols.

主要功能

01Meta-testing protocols to refine instruction clarity and prominent placement

02Framework for documenting and counteracting AI rationalizations verbatim

03Standardized RED-GREEN-REFACTOR cycle for process documentation

04Realistic pressure scenario templates involving time, sunk cost, and authority

05Comprehensive verification checklist for bulletproofing agent behavior

060 GitHub stars

使用场景

01Hardening compliance-heavy skills like TDD enforcement or security protocols

02Verifying that agents follow specific architectural patterns under tight deadlines

03Debugging why an agent is ignoring custom instructions or bypassing established rules

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add seangsisg/crispy-claude testing-skills-with-subagents

For use in Claude.ai and ChatGPT

Download Skill