writing-skills

Evaluation-driven development for skill writing enforces test creation and verification before finalizing changes. It uses a structured RED, GREEN, and REFACTOR workflow with subagents to validate correctness, ensuring skill updates cannot bypass reliability checks.

To enforce test-driven skill writing, initialize the skill with objective evaluation criteria, run a baseline RED test to confirm failure without the skill, implement a minimal GREEN version, then iterate through REFACTOR and multi-model tests using subagents.

Yes, multi-model testing validates skill behavior across multiple models and scenarios to ensure robustness. It operates during the verification phase, using subagents to confirm that skill changes perform correctly under varied conditions.

Yes, subagents are used throughout the phase-driven workflow to verify correctness. During the REFACTOR pass, they execute explicit checks alongside multi-model testing to confirm the skill maintains its expected behavior after modifications.

Yes, explicit frontmatter and a clear Quick Start are required to enable reproducible skill creation and verification. This structured documentation ensures discoverability and maintains auditability for any evaluation-driven skill updates.

Skill modifications bypass verification when lacking a codified evaluation-driven workflow. Enforcing a structured baseline RED, GREEN, and REFACTOR cycle with multi-model testing prevents unverified changes from reaching production-ready skills.