How to Install and Use Skill Creator with Claude Code

Skill Creator is the skill that builds skills. If you've used Claude Code for a while and noticed yourself writing the same instructions over and over — "format the output as markdown, then save to /docs, then run the linter" — that's a skill waiting to happen. Skill Creator does the writing.

This guide walks through installing it, building your first skill in 15 minutes, and running evals so you know it works before relying on it.

Prerequisites

You need:

• Claude Code installed (the CLI, not Claude.ai web)
• A working ~/.claude/skills/ directory (Claude Code creates this automatically)
• Git installed
• Basic comfort with the command line

If you're missing any of these, install Claude Code first from claude.com/code, then come back.

Step 1 — Install Skill Creator

Skill Creator ships from the official Anthropic repo. Clone it into your skills folder:

``bash cd ~/.claude/skills

# Clone the official skills repo git clone https://github.com/anthropics/skills anthropic-skills-temp

# Move just skill-creator into place mv anthropic-skills-temp/skill-creator ./

# Clean up rm -rf anthropic-skills-temp `

Verify it landed correctly:

`bash ls ~/.claude/skills/skill-creator/ # Should show: SKILL.md, references/, scripts/, evaluations/ `

That's it. Claude Code auto-discovers skills in this folder. Next time you start a session, Skill Creator is available.

Step 2 — Confirm it's loaded

Open a fresh Claude Code session and ask:

> What skills do you have available?

You should see Skill Creator in the list. If you don't, the most common causes are:

• The folder isn't named exactly skill-creator (case matters)
• SKILL.md isn't at the root of the folder
• Claude Code is using a different skills directory (check your ~/.claude/settings.json for any skills_path override)

Step 3 — Read SKILL.md before using it

Always read the SKILL.md of any skill before running it. Skill Creator's is short and worth knowing:

`bash less ~/.claude/skills/skill-creator/SKILL.md `

Key things to note:

• It can write files to ~/.claude/skills/ (that's its job)
• It runs scripts from its scripts/ folder for evals
• It doesn't make outbound network calls

This is a low-risk skill — it operates inside your skills folder and runs locally. Compared to skills that hit external APIs, Skill Creator is one of the safer ones in the official repo.

Step 4 — Build your first skill

Let's build something small but useful: a skill that takes a meeting transcript and produces a clean action-items doc.

Start a Claude Code session in any directory and ask:

> Use Skill Creator to build a new skill called meeting-notes-cleaner. It should: take a raw meeting transcript as input, extract action items with assignee and due date, format the output as a markdown action-items doc, and save it to a meeting-notes/ folder. Include 3 evaluation cases.

Claude (with Skill Creator) will:

1. Ask clarifying questions about scope and edge cases 2. Draft ~/.claude/skills/meeting-notes-cleaner/SKILL.md 3. Add reference files for prompt structure 4. Generate 3 sample meeting transcripts as evaluation inputs 5. Write expected outputs for each 6. Run the evals and report pass/fail

The whole process takes 5-15 minutes depending on how clearly you described what you wanted.

Step 5 — Inspect what got built

After Skill Creator finishes, check the output:

`bash ls ~/.claude/skills/meeting-notes-cleaner/ # Expected: # SKILL.md # references/ # evaluations/ # scripts/run-evals.sh `

Open SKILL.md. It should have:

• A clear description field that explains when Claude should use this skill
• Step-by-step instructions for the workflow
• Examples of input and output
• References to any helper files

If anything is vague or missing, ask Claude (with Skill Creator) to revise. The first draft is rarely the final draft.

Step 6 — Run the evals

Skill Creator generates evals so you can verify the skill works before depending on it. Run them:

`bash bash ~/.claude/skills/meeting-notes-cleaner/scripts/run-evals.sh `

Each eval feeds a sample input through the skill and compares the output to expected. You'll see something like:

` Eval 1: pass Eval 2: fail (missing due date for action item 3) Eval 3: pass 2/3 passed `

Failures aren't bad — they're data. They tell you exactly which edge cases the skill misses, so you can revise it. Ask Skill Creator to fix the failing eval, re-run, and repeat until 100%.

Step 7 — Use the skill in real work

Once evals pass, the skill is available in any future Claude Code session. Drop a transcript file into a working directory and ask:

> Clean up the meeting notes in transcript.txt and save the action items.

Claude pattern-matches your request to the meeting-notes-cleaner skill description, loads the SKILL.md, and runs the workflow.

Common patterns I've built with Skill Creator

After a few months with it, here are the patterns I keep coming back to:

• Format converters — markdown to docx, docx to markdown with frontmatter preserved, CSV to JSON with type inference
• Repetitive document workflows — daily standup notes formatter, customer feedback triage, code review summary
• Style enforcers — internal style guide checker that runs across drafts before publish
• Project setup — new project bootstrapper that creates the folder structure I use, drops in templates, sets up README

The pattern is the same: anything you'd write the same instructions for twice is a skill candidate.

Mistakes to avoid

Don't skip the evals. A skill without evals is a skill you can't trust. The 5 minutes to write 3 sample inputs saves you debugging time later.

Don't make skills too broad. A skill that "handles everything about meetings" is harder to use than three small skills (notes-cleaner, agenda-generator, follow-up-emails). Smaller skills are easier to evaluate and easier for Claude to pick up.

Don't include API keys in SKILL.md. Skills are committed to git in many setups. Use environment variables and reference them from the skill: process.env.MY_API_KEY.

Don't trust outputs without spot-checks for the first week. Even with passing evals, edge cases bite. Run the new skill on real work for a few days and verify the output before assuming it's reliable.

When to use Skill Creator vs. write a skill manually

Skill Creator is the right tool when:

• You're new to writing skills
• The skill is medium-complexity (10-50 lines of SKILL.md)
• You want evals and don't want to write them by hand

Skip Skill Creator when:

• The skill is trivial (< 5 lines of instructions) — just write it
• You need very specific control over the prompt structure
• You're modifying an existing skill (open the SKILL.md and edit directly)

For deeper customization, see Anthropic's docs at github.com/anthropics/skills. They include a writing-style guide for SKILL.md that's useful even if you're not using Skill Creator.

Security checklist

Before installing any third-party skill (not just Skill Creator):

• [ ] Repo is from a known author or has 50+ stars
• [ ] SKILL.md doesn't have unexplained shell commands
• [ ] Scripts in scripts/` only operate within the skill's expected scope
• [ ] No network calls to unfamiliar domains
• [ ] No environment variables read that shouldn't be needed

For a longer version, see [how to audit a Claude skill before installing](/blog/how-to-audit-a-claude-skill-before-installing-2026).

FAQ

Does Skill Creator work with the Claude.ai web app? No. It's a Claude Code CLI tool. The web app has a different skill system.

Can I share a skill I built with Skill Creator? Yes. Push the skill folder to a GitHub repo and others can clone it the same way you cloned Skill Creator. Include a README with install instructions.

How does Skill Creator differ from MCP Builder? [MCP Builder](/skills/mcp-builder) is for building MCP servers (network services with tool definitions). Skill Creator builds skills (filesystem-based instruction bundles). Use Skill Creator for "Claude does X with files." Use MCP Builder for "Claude talks to service Y." See our [comparison guide](/blog/mcp-servers-vs-agent-skills-difference-2026).

Can I edit a skill after Skill Creator builds it? Yes. SKILL.md is a markdown file. Edit it directly, or ask Skill Creator to revise. Re-run the evals after any change.

What if my skill has dependencies? Document them in SKILL.md. If the skill needs Python or Node packages, list them. If it needs a CLI tool installed, say so. Claude reads the prerequisites and warns the user if something is missing.

How many skills can I have installed? No hard limit, but each skill description loads into Claude's context when relevant. 10-20 active skills is the sweet spot for most users. Beyond that, audit and remove what you don't use.

Bottom line

Skill Creator is the most useful meta-skill in the Anthropic repo. It turns the "I keep telling Claude the same thing" problem into a 15-minute fix that pays off every time you do that task.

Install it, build something small in your first session, and let the evals catch your bugs. By the third skill, you'll have a clear sense of when something is worth turning into a skill versus when it's a one-off ask.