Looking at nine weeks of the star-growth ranking, the category that held attention the longest was not a model or a framework. It was repositories of skills for coding agents. mattpocock/skills was on the list for seven consecutive weeks; addyosmani/agent-skills, Tencent/BrowserSkill and K-Dense-AI/scientific-agent-skills have all appeared.
What they are, for anyone who has not looked: a skill is a folder with a SKILL.md file, a short description of when it applies and a set of instructions for the agent, plus optional scripts or reference files. The agent reads the description, and when a task matches it loads the full instructions. It is prompt engineering with a file format: versioned, shareable, testable. Anthropic published the format for Claude in October 2025 and the repositories above are collections of them, some tool-specific, some meant to be portable.
Why they trend: they are cheap to write, immediately useful, and anyone who has got an agent to do something reliably wants to keep the recipe. A skills repository is a cookbook.
What we are less sure about and would like first-hand views on: do skills from a public repository transfer to your project without editing, or do they mostly serve as examples? And how do people keep a skill from going stale when the tool it describes changes?