If you’re using Agent Skills the same way you did six months ago, you’re missing out on a lot. In this video I walk through what quietly changed in Claude Code Skills—and why it basically upgrades your whole agentic coding workflow. I start with the boring-but-important part: installing skills. You can do it directly in Claude Code via the plugin command (and make sure the official marketplace repo is added), or you can use skills.sh to install skills in a way that works across a bunch of agentic coding tools—not just Claude Code. Then I show how skills are structured (the skill.md front matter, name + description) and why I mostly don’t bother building them manually anymore: the Skill Creator skill now does the heavy lifting. It can generate skills, but more importantly it can test them, create evals, and run blind A/B comparisons against the baseline model. I also break down the two skill types Anthropic talks about—capability uplift vs encoded preferences—so you know when you’re adding a real new capability versus just enforcing a workflow. From there we get into the workflows that actually matter: subagents with background mode for long-running tasks, context fork to keep your main context window clean, and dynamic context injection (pre-processing commands) so your skill can pull in live project data like file trees, package.json, and recent git commits before the agent even reads the prompt.