← Blog
·5 min read

Claude Code skills: teach your agent your workflows

Skills are folders of instructions Claude Code loads on demand — write down how your team deploys, reviews, and ships once, and every future session just knows. What they are and how to write good ones.

Skills are how you stop re-explaining your workflows to Claude Code. A skill is a folder with a SKILL.md file — instructions, optionally scripts and templates — that Claude Code loads when a task matches. Write down how your team does deploys, reviews, or release notes once, and every future session just knows.

The problem they solve

Out of the box, an agent is a brilliant new hire on day one: capable, but ignorant of your conventions. You find yourself typing the same context every session — “we use conventional commits,” “always run the smoke tests before pushing,” “PDFs get generated with our template, not from scratch.” A CLAUDE.md file covers project facts; skills cover procedures. The difference is that skills load on demand — the deploy skill only enters context when you’re deploying — so you can have dozens without drowning the model in instructions.

Anatomy of a skill

~/.claude/skills/release-notes/
  SKILL.md        # when to use it + step-by-step instructions
  template.md     # the format your team expects
  changelog.sh    # a script it can run instead of improvising

The SKILL.md frontmatter carries a name and a description; the description is what Claude Code matches against when deciding a skill is relevant. The body is instructions in plain markdown. Bundled files let the skill carry working code — deterministic scripts beat regenerated ones.

What good skills look like

  • Procedural, not encyclopedic. “How we cut a release,” not “everything about our codebase.”
  • Trigger-worded. Write descriptions with the phrases you’d actually say: “use when asked to prep a release or write a changelog.”
  • Carrying tools. If a step is a command, ship the command in the skill instead of describing it.

Skills + a task queue = a workflow that compounds

Skills get sharpest when the agent works from real tasks. Our loop: scoped tasks live in Lume, Claude Code picks them up over MCP, and skills tell it how we do the work — commit style, test gates, review handoff. The task says what; the skill says how; the human reviews. New teammates (human or AI) inherit the whole system on day one. If you’re not dispatching tasks to Claude Code yet, here’s how that works — and if you’re brand new, start with the install guide.

Want a list your agents can pull from?

Lume gives every task an API, an MCP server, and an assignee. Free to start.