Skills as Saved Workflows

A skill is a saved workflow that Claude can invoke on demand. Under the hood, it is a markdown file with a YAML front matter section and a body of instructions. The front matter contains the skill name, a description, and trigger words — the phrases that tell Claude when to activate this skill. The body contains the procedural steps, decision logic, and any linked files that the skill needs. When you say a trigger phrase, Claude loads the skill into context and follows the instructions. It is the difference between giving directions from memory every time and handing someone a printed route.

Skills persist across sessions. Once created, a skill is available every time you launch Claude Code in that project. This is what makes them fundamentally different from conversation instructions. "Always run tests after making changes" in a conversation dies when the session ends. The same instruction encoded as a skill is permanent infrastructure. It fires every time, for every developer, without anyone remembering to say it. Skills are how you turn ad-hoc behavior into institutional process.

Creating a skill follows the meta-prompt pattern. Do the task once with Claude manually. Watch what works — the sequence of steps, the decisions made, the error handling that emerged naturally. Then tell Claude: "Crystallize what we just did into a skill called deploy-staging. The trigger phrase should be 'deploy to staging.'" Claude captures the process, strips the noise, and saves it as a reusable skill file. You went from doing a thing to owning a process in sixty seconds.