wtf is agent-skills?

This project offers a pair of portable skills for AI coding assistants that are meant to fix a common problem: assistants like Claude Code often tell you something looks good in a vague way, without a clear standard for what good actually means. These skills give that judgment a fixed reference point instead. The main skill, called production bar, audits a feature or system you built and compares it against named, well known peer products that are considered top tier in that space. It then gives a single production readiness score based on a calibrated scale showing how far the work is from that peer level, and separately ranks every gap it finds by how much fixing it would actually matter. The goal is not to make the assistant more critical for its own sake, but to make its approval something that is earned and specific rather than another version of good enough. A second skill, common sense check, rechecks that the assistant's work still lines up with the original goal as it proceeds, catches situations where it is repeating the same small fix over and over without real progress, and provides a structured way to run an explicit sanity check partway through a task. Both skills follow an open specification called Agent Skills, which means they are not limited to Claude Code and can also be installed for other compatible assistants such as Codex, GitHub Copilot, Cursor, and Gemini CLI using a GitHub command line installer. You can preview a skill before installing it. The repository also includes eight test cases, including one that specifically checks that the skill correctly refuses to invent a score when there is not enough evidence to give one. The project is released under the MIT License.