"Github stars don't guarantee safety," Alan wrote at the bottom of his own list of the ten most bookmarked skill repos for coding agents, right after telling people to save the list. Superpowers leads at roughly 290,000 stars. Ponytail follows at 147,000. Nine more names round out a list that's been reposted all week, and almost nobody installing from it is going to open the files first, caveat or not. Anthropic published Agent Skills as an open spec in December, about fourteen months after Claude Skills first shipped, and close to 40 products now claim compatibility, including OpenAI's Codex, GitHub Copilot, Cursor and Gemini CLI. The scale behind that adoption curve is enormous: one live tracker counts over 62,000 individual skills and 10,000 marketplaces, more than 172 million installs combined. Skills.sh, backed by Vercel and barely ten months old, already lists hundreds of thousands on its own. Separately, more than 2,500 Claude Code plugin marketplaces are now registered at claudemarketplaces.com. Nobody built review at the same speed distribution grew. SkillsBench ran a quality pass on over 47,000 public skills and landed an average score of 6.2 out of 12. A separate audit of 22,500 skills found roughly 140,000 issues between them, about six per skill. Snyk's ToxicSkills research found prompt injection present in 36 percent of the skills it tested. That mismatch, not the model underneath any of these skills, is where the next real incident is going to come from. xitter.cf/alannnfx/status/210920…