Skip to content

Claude Code skill

Hand someone's Claude a new ability — a routine it runs on demand, not a document they read. Turn a workflow you keep re-explaining (a house style, a grant procedure, an onboarding checklist) into a small folder. Their agent picks it up and uses it across all their work — they never read or run the code.

Reach for it when you want a repeatable capability spread, not a one-off file. Skip it when someone just needs to read or run a finished thing — that's a repo or a doc.

Last verified: 2026-06-07 · Confidence: high on what a skill is, the share paths, and the access model — all checked against Anthropic's skills/plugins docs.

It allows you to

  • Teach their agent your procedure. Their Claude runs the routine itself — spotting when to use it, then following your steps. No app, no file to read.
  • Make it reusable. It fires across every future project, not just the task you shared it for.
  • Encode judgement, not just text. It carries the how — the order, the gotchas, the format — a static document can only describe.
  • Fix it once for everyone. Ship it as a one-command install and your later changes reach everyone who has it. → Good to know.
  • Keep the surface tiny. One short file (SKILL.md) plus any helpers — readable in a minute, so people trust it before installing.

Ideal for

  • A shared procedure across an org — your grants team's "turn a project brief into our funder's proposal format" routine, so every officer's Claude writes proposals the same way without re-pasting the template.
  • A house data-cleaning standard — "normalise a messy survey export: dedupe, fix the country codes, flag PII columns" — so the whole research team cleans to one spec, not five.
  • An onboarding ability for new hires — a /our-stack skill that walks a newcomer's agent through your repos, conventions, and where things live, day one.
  • A domain method too fiddly to re-explain — the exact way your forecasting team scores a question, baked in so the steps never drift between people.
  • A field-wide library — like K-Dense's Scientific Agent Skills: 142 science skills that 160,000+ researchers install into their own agent, one repo handing a whole field a shared toolkit.

Who can get in

A skill lives inside a repo (a marketplace is just a repo too) — so its reach is exactly the repo's reach: private for people you invite, public for anyone, nothing in between. [confirmed]

  • The skill's own twist: dropping someone from the private repo stops their updates, but a skill they already installed sits in their ~/.claude/skills/ and keeps running. (Revoking never reaches a copy.) [confirmed]

Which rungs it can hold. Just you / named people / org-only / the whole internet — every rung except anyone with the link, inherited from the repo. → Who can see it? [confirmed]

Handing data to the host. Two companies touch this: GitHub stores the folder (private repos stay private) and Anthropic makes the Claude that runs it. The skill is instructions you meant to share — the data question is about the recipient's own chats, not your skill. → Can you trust the company?

What you do to set it up

  • Ask: tell Claude Code "turn this workflow into a Claude Code skill and put it in a repo I can share." Your agent writes the SKILL.md, tests it loads, and puts the folder online. Want a one-command install instead of a folder to copy? Add: "make it a plugin with a marketplace repo." Every change after: one sentence, ~0 effort.
  • One-time, in order:
    1. Set up Claude Code — what builds and tests the skill, ~10 min once.
    2. Create a GitHub account + connect the GitHub CLI — so your agent can put the skill in a repo, ~13 min once total.

Full walkthrough, both paths (including by hand): Create and share a skill.

What the other person does

The recipient needs their own Claude Code — this ability lives in their agent, so there's no path for someone without one. (New to it? Set up Claude Code, ~10 min once.) [confirmed]

  • From a repo folder: point their agent at the link — "install the skill in this repo into my Claude" — and it drops the folder into their ~/.claude/skills/. Private repo? Accept your invite first, ~2 min. [confirmed]

  • From a one-command install (a marketplace): two lines and it's live — and they get your updates when you publish them. [confirmed]

    /plugin marketplace add your-github-user/your-repo
/plugin install your-plugin@your-marketplace
    - Glance at the SKILL.md first. It runs with their agent's tools on their machine, so only install from someone you trust — Anthropic's docs flag this. [confirmed] - Pay: nothing — skills, repos, and the marketplace are all free.

Full recipient walkthrough: What the recipient does.

Other ways to share

  • Connecting their agent to a live service — an API, a database, your internal data? → an MCP server: a skill is instructions, an MCP server is a live connection.
  • It's really one command a person runs, not a routine their agent absorbs? → a one-command tool they type by name, or — on a Mac — a Raycast extension they fire from a hotkey. Neither needs Claude Code on their side.
  • Want to hand over the whole project the skill lives in? → a GitHub repository gives them every file and the full history to build on.

Sources

Good to know

  • Updates only auto-reach the one-command path. A copied folder is a snapshot — they re-copy to get changes; the one-command install pulls your latest on a refresh. [confirmed]
  • Not locked to Claude Code, and installs can't collide. A skill follows the open Agent Skills standard, so the same folder works in Cursor, Codex, and other agents; plugin skills are namespaced (/my-tools:review, not /review). [confirmed] Both, plus the by-hand path: the fine print.