How do I stop an AI from rationalizing away rules?

By documenting their specific excuses (e.g., 'it's more pragmatic this way') and adding explicit negations in your skill documentation that address those excuses directly.

Why is the RED phase important in skill testing?

The RED phase establishes a baseline by watching an agent fail without the skill. If you don't see the agent fail first, you can't be sure the skill is actually what's preventing the failure.

What are pressure scenarios?

Pressure scenarios are realistic prompts that combine constraints like tight deadlines, sunk cost (hours of work), and authority pressure to force an AI to choose between following a skill or taking a shortcut.

What does it mean to test a skill with subagents?

It means using one AI agent to run 'pressure scenarios' against a specific skill to see if the agent will ignore the skill's rules in favor of speed or convenience, allowing you to find and fix loopholes.

Subagent Skill Testing

Name: Subagent Skill Testing
Author: dsifry

bydsifry

•

安全与测试

Validates and stress-tests AI skill documentation using a TDD-inspired Red-Green-Refactor cycle to ensure agent compliance under pressure.

关于

This skill applies Test-Driven Development (TDD) principles to the creation of process documentation and AI behaviors. By running 'pressure scenarios' without a skill (RED), writing minimal documentation to address specific failures (GREEN), and iteratively closing rationalization loopholes (REFACTOR), it ensures that AI agents actually follow guidelines instead of bypassing them. It is essential for developing 'bulletproof' skills that enforce discipline, manage compliance costs, or contradict immediate goals like speed over quality.

主要功能

Implements a RED-GREEN-REFACTOR cycle for process documentation.
Generates multi-pressure test scenarios including time, sunk cost, and authority constraints.
Includes a rationalization table and 'Red Flag' detection for compliance monitoring.
29 GitHub stars
Provides framework for Meta-Testing to clarify documentation ambiguity.
Identifies and documents AI rationalization patterns verbatim to plug loopholes.

使用场景

Hardening TDD enforcement skills to prevent 'testing after code' behavior in agents.
Verifying that safety or quality guidelines are followed during simulated high-pressure deadlines.
Debugging and refining ambiguous documentation that leads to inconsistent AI agent behavior.

关于

主要功能

Implements a RED-GREEN-REFACTOR cycle for process documentation.
Generates multi-pressure test scenarios including time, sunk cost, and authority constraints.
Includes a rationalization table and 'Red Flag' detection for compliance monitoring.
29 GitHub stars
Provides framework for Meta-Testing to clarify documentation ambiguity.
Identifies and documents AI rationalization patterns verbatim to plug loopholes.

使用场景

Hardening TDD enforcement skills to prevent 'testing after code' behavior in agents.
Verifying that safety or quality guidelines are followed during simulated high-pressure deadlines.
Debugging and refining ambiguous documentation that leads to inconsistent AI agent behavior.