How to Evaluate and Test Agent Skills

This video walks through a practical workflow for evaluating and testing agent skills so you can improve them without breaking what already works. I use the Hugging Face `$model-trainer` skill as the case study and show how to define success, trigger the skill manually, build a compact prompt set, and evolve your checks over time. If you're building workflows for Codex, Claude Code, or similar coding agents, this gives you a lightweight eval loop you can actually adopt. --- 🤓 Topics Covered - Agent skill eval workflow - Deterministic and rubric checks - Prompt sets for regression testing -- Links - https://developers.openai.com/blog/eval-skills - HF Skills: https://github.com/huggingface/skills (give us a star! ⭐️) ## Connect with me - X: https://x.com/_alejandroao - LinkedIn: https://www.linkedin.com/in/alejandro-ao/ --- ⏱️ Timestamps 0:00 Introduction 0:36 Source article and framing 0:58 What skills are 3:47 Demo setup with `$model-trainer` 6:14 Define success criteria 7:34 Version control your skills 8:04 Manually trigger the skill 11:16 Evolve tests and add LLM-as-judge 14:14 Final takeaway