Takes the RED-GREEN-REFACTOR cycle from TDD and applies it to writing prompts for Claude. You spin up subagents to test scenarios without your prompt first, watch exactly how they fail, then write instructions that fix those specific failures. The approach covers everything from git commands to discipline-enforcing skills that need to resist rationalization under pressure. What's smart here is the insistence on documented baseline failures. If you didn't watch an agent mess up without the prompt, you're just guessing at what needs fixing. Ships with scenario templates for different prompt types and a clear mapping between test-driven development phases and prompt iteration.
npx skills add https://github.com/neolabhq/context-engineering-kit --skill test-prompt