What you type in the moment. One-off. Gone when the chat ends.
A reusable capability Claude loads when it's relevant. Write once — works forever.
You mostly build Skills. The rest is how Claude reaches the outside world.
scripts/ — code it can run (calculations, conversions, API calls) · references/ — docs it reads only when needed · assets/ — templates, logos, fonts for its outputsLoads in layers — name+description always, instructions when it triggers, extras only as needed. That's why it never clutters your chat.
Customize → Skills — home base.
Browse — switch on ready-made Skills.
Toggle any Skill on/off in your list.
One Skill does the whole job start to finish. Best for a simple, self-contained task you always run the same way.
Break a big job into small single-purpose Skills, then run them as a cascade. Best for anything multi-step or reused elsewhere.
Rule of thumb: simple & one-off → all-in-one. Multi-step or reused → split and chain.
Share to specific people or the whole org. Recipients get it view-only — and when you update it, they auto-get the new version.
✓ they follow your updatesHand someone the file directly. It's a snapshot — if you improve it later, resend to share the update.
⚠ frozen copyDrop it in the repo's .claude/skills/ folder, commit & push — teammates get it (and updates) on pull.
✓ updates on pullBuild once — then multiply it across the team.
MCP connectors get their own session later.
Tell it your diet, budget, and who you're feeding — once. From then on, "plan this week's dinners" gives you a plan and a grocery list that already fit. No re-explaining.
Same build-once, reuse-forever trick — for your life.