Claude Code skills: the packaging unit I didn't know I needed
Claude Code's skill system is the right abstraction for packaging team workflows. Treat them like internal CLI commands: small, composable, with clear inputs and outputs.
Skills are the piece of Claude Code I underrated for the longest. On paper they're just a folder with a SKILL.md describing when to invoke it and what tools it needs. In practice they're how you stop reinventing the same prompt every Tuesday.
The pattern that worked for me: one skill per repeated workflow. A `/ship` skill that runs lint, type-check, tests, then drafts a PR title and description. A `/refresh-data` skill that pulls fresh fixtures and regenerates the GraphQL codegen. A `/triage` skill that reads the GitHub issue queue and proposes labels. None of these are smart on their own; they're checklists with a model on top, and that's the point.
What makes skills better than raw slash commands: they're discoverable, and the model decides when to invoke them based on the description. You don't have to remember the trigger. The SKILL.md frontmatter is the agent's index. Get the description sharp and stop writing brittle decision logic in your prompt.
Where they fall down: trying to make one skill do too much. A skill that does 'ship and deploy and notify Slack and update Linear' is unreadable for the model. Split it. Skills compose; monoliths don't.