Skills: your SOPs, encoded

The core move of this whole course

Three checkpoints mark this lesson: you have run real skills, read one from the inside, and ranked the five chores you will encode this week. A skill is a procedure written down in a folder, in a format Claude Code knows how to find and follow. Write the weekly-report procedure once as a skill, and from then on /weekly-report runs it - for you and for anyone you share the folder with.

If your org has SOPs, you already know how to think about skills. The difference is that an SOP hopes a human follows it, while a skill is executed by an agent that actually reads the whole thing, every time.

The law has a path attached. We call it the Ladder, and you climb it for the rest of the course:

• Do it once - do it in chat. A prompt is the right tool for a one-off.
• Do it twice - find or build a skill. Check the library first, write your own when only your company knows the procedure. Repetition is the signal; never type the same instructions a third time.
• It runs on a schedule - make it an automation: a hook when it must fire around Claude's work, a headless job or Routine when it runs on the clock (lesson three).
• It needs judgment mid-flow - that is an agent: a system that decides between steps, built from the same skills. The vertical tracks build these.

How a skill works

Mechanically, a skill is a folder containing a SKILL.md file - instructions in markdown with a small header - plus any supporting files: templates, reference docs, scripts. Skills follow the Agent Skills open standard (agentskills.io), which Anthropic started and opened up, so the format works across products, not just Claude Code.

The clever part is progressive disclosure, and it connects straight to week two's context lesson: only each skill's name and short description sit in context all the time. The full instructions load when the skill is actually used, and supporting files load only if needed. You can have fifty skills installed for the context cost of a paragraph.

~/.claude/skills/        # personal - all your projects
.claude/skills/          # project - shared with everyone via git
~/.claude/skills/synced/ # skills enabled on your claude.ai account,
                         # synced automatically (anthropic-skills: namespace)
# plus skills shipped inside plugins, and org-managed skills

Project skills are discovered up the folder tree and in nested folders in bigger repos, and they live-reload - edit a SKILL.md and the change applies immediately, no restart. If .claude/skills/ itself did not exist when the session started, run /reload-skills once. Skills you enable on claude.ai sync down on their own; set syncClaudeAiSkills: false if a project must stay isolated.

Borrow before you build: the open skills libraries

Because a skill is just a SKILL.md file (plus optional scripts) in a portable open standard, there are huge free libraries of ready-made skills you can borrow instead of writing your own. That sets up the core choice you will make for every recurring task from here on: borrow a ready skill from the library, or build your own.

• SkillsMP (skillsmp.com) - a free, independent community catalog that indexes 3,000,000+ open-source skills in SKILL.md format from public GitHub (as of September 2026), browsable by job and category, with a free REST API and MCP server so Claude can search it from inside a session.
• Anthropic's skills repo (github.com/anthropics/skills) - the reference set: document skills (docx, pdf, pptx, xlsx), creative and enterprise examples, the skill-creator skill that writes skills, the spec, and a template. Installs as a plugin marketplace.
• The official plugin marketplace - pre-registered in every install as claude-plugins-official, browsable at claude.com/marketplace/plugins. Most of it comes from vendors and partners, plus a smaller set Anthropic maintains; the web catalog marks some plugins 'Anthropic verified'. The community marketplace sits beside it.
• agentskills.io - the open standard itself, adopted by Gemini CLI, OpenCode, OpenHands and JetBrains Junie, so a skill you write for Claude Code travels.

# Anthropic's reference skills (includes skill-creator):
/plugin marketplace add anthropics/skills
/plugin install example-skills@anthropic-agent-skills
/plugin install document-skills@anthropic-agent-skills

# The official and community marketplaces:
/plugin install <name>@claude-plugins-official     # pre-registered
/plugin marketplace add anthropics/claude-plugins-community
/plugin install <name>@claude-community            # note: not the repo name

# From GitHub or SkillsMP: read the skill first, then copy its folder into:
.claude/skills/<skill-name>/     # project - shared via git
~/.claude/skills/<skill-name>/   # personal - all your projects

Do both. Install the generic ones - a PDF filler, a CSV cleaner, a commit-message writer - and hand-build only what is genuinely yours: your voice, your ICP, your gates, your SOPs. Build from scratch is for the stuff no library can know about your business. The next lesson opens with this exact check, then teaches the build path in full.

Run what is already there

Before writing anything, spend fifteen minutes as a user of skills that already exist - it sets your quality bar and shows you the format from the consumer side.

1. In a session, type / and pause. The menu lists every command and skill available right now.
2. Run /skills to see the full inventory and where each one comes from - personal, project, plugin, or bundled.
3. Claude Code ships with bundled skills - try /code-review after editing something, or look at /loop (run a prompt on an interval), /debug, /verify, /batch and /deep-research.
4. Read one: find a bundled or installed skill and ask Claude to show you its SKILL.md. Notice it is just markdown anyone can read.

One change to know about, so old tutorials don't confuse you

Older guides talk about custom slash commands - markdown files in .claude/commands/ - as a separate thing from skills. They have since been merged: both a commands file and a skill folder named deploy create /deploy, old commands files still work, and skills are the recommended form going forward. If a 2025 tutorial says "create a command", read it as "create a skill".

Practical consequence: if you inherit a project with a .claude/commands/ folder, nothing is broken and nothing is urgent. Migrate a command to a skill folder when you next touch it - the skill form gets you the frontmatter controls from the next lesson.

The extension map: which tool for which job

Skills are one of several ways to extend Claude Code, and the framing below is the canonical one. You will meet the rest over the next two weeks; for now, just be able to place them.

• CLAUDE.md and rules - always-loaded facts. Who you are, what is always true.
• Skills - on-demand procedures. How to do a specific recurring thing.
• Subagents - delegated workers with isolated context, restricted tools and their own pinned model, for research, bulk tagging and review. You met one in the models lesson; the vertical tracks build real ones.
• Hooks - guaranteed actions at fixed moments. Not advisory: they always fire (lesson three this week).
• MCP - connections to external tools: your CRM, Slack, Notion (week four).
• Workflows - saved orchestration scripts in .claude/workflows/ that run many agents at once (models lesson). For work too big for one session.
• Plugins - the packaging format that bundles all of the above for distribution (lesson four this week).

Do this now

Sources and further reading

• SkillsMP - community catalog of 3M+ open-source skills
• Anthropic's skills repo - reference skills and skill-creator
• Official plugin marketplace
• Agent Skills - the open standard
• Skills - Claude Code docs