Show HN: An AI agent skill repo built around evals, not demos (github.com)

A new initiative has emerged within the AI community: a repository for reusable AI agent skills centered around rigorous evaluation rather than mere demonstrations. This platform, exemplified by the "world-cup-picks-report" skill, is designed to track regression metrics and job trial histories comprehensively. With focused dashboards showcasing top-line metrics and a complete historical record, users can thoroughly evaluate the performance and development of each skill before integrating it into their projects. This model emphasizes the importance of evaluation artifacts, promoting a higher standard for AI development and deployment. Significantly, the use of Anthropic-backed LLM judges in the regression verification process enhances the credibility and accuracy of the evaluations. Developers can utilize various agents, such as Codex or Claude-Code, to run the regression evals, allowing for flexibility based on user preference and resource availability. The workflow not only enables continuous integration of improvements through comparisons against Git-tracked baselines but also promotes a disciplined approach to tracking changes and updates in skills. This initiative reflects a growing trend in the AI/ML community towards accountability and transparency, encouraging developers to build and refine AI systems based on concrete evidence of performance.