Skills

What skills are, how they differ from always-on memory, and a copy-paste demo skill you install locally.

Skills in Claude Code are folders with a SKILL.md file. They add a playbook Claude can load when it is relevant, or when you invoke them with a slash command. That keeps long procedures out of your main chat until you need them. See Anthropic’s guide: Extend Claude with skills.

How they differ from `CLAUDE.md`

CLAUDE.md (and imports like AGENTS.md) load every session and spend context up front.
Skills load on demand: Claude pulls them in when your request matches the skill’s description, or when you run /skill-name.

Use a skill when you would otherwise paste the same multi-step instructions over and over.

Where skills live (project)

For one repository, project skills go under:

.claude/skills/<skill-name>/SKILL.md

Pick the directory you open in Claude Code. Personal skills use ~/.claude/skills/ instead. Other locations (enterprise, plugins) are in the Where skills live table in the official skills doc linked above.

If you work on .NET and C#, the .NET Skills page covers Microsoft’s curated marketplace (dotnet/skills) for agents and IDEs.

To slash filler and output tokens while keeping technical content (a humorous “caveman” reply style, optional compression of memory files), see Caveman (JuliusBrussee/caveman).

For think-first, minimal-diff, test-driven behavior based on Andrej Karpathy’s take on common LLM coding failures, see Andrej Karpathy skills (multica-ai/andrej-karpathy-skills).

Install the demo skill (copy and paste)

Do this once per machine for the project where you want the skill:

Open a terminal in your project root.

Create the folder for the skill:

mkdir -p .claude/skills/workshop-three-checks

mkdir -p .claude/skills/workshop-three-checks

New-Item -ItemType Directory -Force -Path .claude/skills/workshop-three-checks | Out-Null

Create the file .claude/skills/workshop-three-checks/SKILL.md in your editor (on Unix you can touch the file first if you like).

Copy the entire block below (from the first --- through the last line) and paste it into SKILL.md. Save the file.

Start Claude Code from that same project root so it loads project skills.

Invoke the skill with /workshop-three-checks or ask naturally, for example: I am about to commit, run the quick workshop checklist.

Other natural openings that match the demo skill’s description:

/workshop-three-checks

I touched the docs sidebar and added a page, quick pre-commit pass before I commit.

About to push a branch that changes API handlers. Run the workshop three-checks skill and call out anything extra I should run beyond my usual test command.

Claude Code watches an existing .claude/skills/ directory, so new skill folders and edits normally load during the session. Restart only if the top-level skills directory did not exist when the session started (see Live change detection in the skills doc).

`SKILL.md` to copy

---
name: workshop-three-checks
description: Quick pre-commit sanity check for this workshop repo. Use when the user is about to commit, asks for a last check before commit, or wants a short review before pushing.
---

When this skill runs, answer in **short bullets** (stay under about 12 lines unless the user asks for more):

1. **Tests** Infer the stack from the repo (`package.json`, `.csproj`, `*.sln`, etc.). Name the most relevant check (for example `npm run build`, `dotnet test`, `dotnet build` for **.NET 10+** in this workshop) and say they should run it before committing if they have not.
2. **Scope** In one sentence, what changed and does it match what they said they were doing?
3. **Commit message** Propose a single-line subject in conventional style (for example `docs: add skills page` or `chore: tweak sidebar`).

Do not run destructive commands. This is a checklist only.

Make your own (minimal)

Create .claude/skills/<name>/.

Add SKILL.md with YAML frontmatter at the top:

For project and personal skills, the directory name becomes /<name>. Plugin skills are namespaced and can use frontmatter name for their final command segment.
description: when Claude should load it automatically.

Put the steps or rules in markdown below the frontmatter.

Keep the first skill small: one clear job, a few bullets, optional link to official docs for advanced options (supporting files, invocation control, subagents).

Claude Code plugins

Distribute skills with agents, hooks, MCP, and LSP integrations.

Skills in Codex

Compare .agents/skills and explicit $skill invocation.