My AI Agent Built Its Own Tools — Here's How OpenClaw Skills Actually Work

I spent 12 hours installing OpenClaw on Windows WSL. The setup was painful, but what happened after was worth every minute. The moment that changed my mind I told my AI agent: "Create a new skill called github-trending that shows me today's top GitHub projects." Two minutes later, it had written the skill code, committed it to git, and delivered a formatted report of today's trending repos — complete with Star counts and analysis of why each project matters. The AI didn't just use a tool. It built a tool, installed it, and used it. The painful part: skill dependencies OpenClaw has 55 optional skills, but 45 of them failed to install on my first try. The error messages weren't helpful: Install failed: github — brew not installed Install failed: nano-pdf — uv not installed After hours of troubleshooting, I figured out the dependency chain: Skills need system tools (gh, poppler, ffmpeg) which need package managers (Homebrew, uv) which need your OS to be properly configured. The fix was st