writing-skills

Test-driven development for AI agent documentation involves running baseline pressure scenarios without the skill, observing specific agent failures, and writing minimal process documentation that addresses those exact failures. This approach ensures AI agents follow instructions even under time, sunk cost, and authority pressures.

Test-driven development for skill authoring applies RED-GREEN-REFACTOR cycles to process documentation instead of code. It uses pressure scenarios and rationalization tables to test that AI agent skills resist loopholes, ensuring documentation enforces compliance when agents face stress conditions like time pressure or sunk cost bias.

You test AI agent skills against rationalization by using subagent pressure scenarios that simulate time constraints, sunk cost bias, and authority pressure. The writing-skills approach baselines agent behavior without the skill, then verifies that documentation prevents common excuses like skipping tests or claiming manual testing sufficed.

Yes, you can use TDD principles to create enforcement skills for AI agent systems by applying RED-GREEN-REFACTOR cycles to documentation. This approach suits discipline-enforcing skills, technique guides, pattern references, and API documentation requiring compliance verification under stress conditions like deadline pressure or authority override attempts.

The best way to optimize AI agent skill discovery is through structured documentation writing that applies discovery optimization techniques. By crafting descriptions and skill structures specifically for agent comprehension, you ensure AI systems find and correctly use process documentation when following test-driven development workflows.

AI agents ignore process documentation under pressure when skills lack testing against rationalization scenarios. You prevent this by applying test-driven development to authoring, using pressure tests for time and authority biases, and building rationalization tables that address common excuses like deferring tests or claiming manual verification is sufficient.