Skills Repos Now Own GitHub Trending
Look at today's GitHub Trending and count: mattpocock/skills gained 2,180 stars today (208k total), addyosmani/agent-skills gained 1,131 (83k), obra/superpowers gained 794 (268k), google/skills gained 305. Four of the top spots are repos containing no runnable software at all — just structured instructions for AI coding agents. Markdown is outcompeting code on a site built for code.
The one spiking hardest relative to its age is Addy Osmani's agent-skills (github.com/addyosmani/agent-skills). It packages 24 skills across six lifecycle phases — define, plan, build, verify, review, ship — plus specialist personas like a security auditor and a web performance auditor, and setup guides for ten-plus agents from Claude Code to Cursor to Copilot. The detail that stands out: every skill ships with an anti-rationalization table, pre-written counters to the excuses agents (and humans) make to skip writing tests or cut review corners. Someone finally noticed that the failure mode of coding agents is not ignorance, it is rationalized laziness, and wrote defenses against it directly into the prompt layer.
The economics of this trend are worth a pause. Engineering judgment used to be transmitted through books, code review, and years of apprenticeship. It is now being serialized into markdown files that any agent can load in seconds, forkable and diffable like code. When a Chrome engineering leader's entire quality playbook costs one git clone, the scarce asset stops being knowledge of best practices and becomes taste in choosing which pack to trust. The star counts say tens of thousands of developers are making that choice right now.
← Back to all articles
The one spiking hardest relative to its age is Addy Osmani's agent-skills (github.com/addyosmani/agent-skills). It packages 24 skills across six lifecycle phases — define, plan, build, verify, review, ship — plus specialist personas like a security auditor and a web performance auditor, and setup guides for ten-plus agents from Claude Code to Cursor to Copilot. The detail that stands out: every skill ships with an anti-rationalization table, pre-written counters to the excuses agents (and humans) make to skip writing tests or cut review corners. Someone finally noticed that the failure mode of coding agents is not ignorance, it is rationalized laziness, and wrote defenses against it directly into the prompt layer.
The economics of this trend are worth a pause. Engineering judgment used to be transmitted through books, code review, and years of apprenticeship. It is now being serialized into markdown files that any agent can load in seconds, forkable and diffable like code. When a Chrome engineering leader's entire quality playbook costs one git clone, the scarce asset stops being knowledge of best practices and becomes taste in choosing which pack to trust. The star counts say tens of thousands of developers are making that choice right now.
Comments