Is skill-creator official and free?

Yes. It's written and shipped by Anthropic in the anthropics/skills repository and the official Claude Code plugin directory, and the skill's folder carries an Apache-2.0 license. It's a reference implementation, though — the repo disclaimer says it's for demonstration and educational purposes.

Do I need skill-creator to create a Claude skill?

No. A skill is just a folder containing a SKILL.md file with YAML frontmatter and instructions, and you can write one by hand in minutes. Skill-creator adds the parts people skip: an intent interview, eval runs against a no-skill baseline, benchmarking, and description tuning for triggering accuracy.

Does it work in Claude.ai or only Claude Code?

Both. Claude Code is its primary home, installed as a plugin or copied into ~/.claude/skills/. On Claude.ai, Anthropic's example skills including skill-creator are already available on paid plans once Skills is enabled in Settings, and the API accepts uploaded skills too.

What are the four modes?

Create drafts a new SKILL.md from an interview about intent, triggers, and output format. Eval runs your test prompts through paired subagents — with and without the skill — and shows results in an HTML viewer. Improve uses Grader, Comparator, and Analyzer agents to score outputs and suggest rewrites, and Benchmark repeats runs for variance analysis.

Skill review · Anthropic official

Claude Skill Creator

Search for a Claude skill creator and every road leads to the same place: skill-creator, Anthropic's official skill for authoring other skills, shipped inside the anthropics/skills repository (172,953 stars) and as a standalone plugin in the official Claude Code directory (385,083 installs at last check). It covers the full lifecycle in four modes — Create interviews you and drafts the SKILL.md, Eval runs your test prompts with and without the skill attached, Improve rewrites based on graded results, Benchmark repeats runs for variance analysis — plus a script that tunes the frontmatter description so the finished skill actually triggers. One thing to know before installing: Anthropic labels the whole repo "for demonstration and educational purposes," and the eval loop spawns two subagent runs per test case, so a full iteration cycle burns real token budget on metered plans. You do not need it to make a skill — a skill is a folder with one markdown file — but hand-rolled skills rarely get tested, and testing is most of what this one does.

View the sourceApache-2.0 licensed · 172,953★ · facts checked September 1, 2026

By Robert MacKelfresh

Founder, Automation Squad · September 1, 2026

Install

1Claude Code, quickest route: run /plugin install skill-creator@claude-plugins-official — the standalone plugin from Anthropic's official directory. If that marketplace isn't registered yet, add it first with /plugin marketplace add anthropics/claude-plugins-official.
2Or install from the skills mono-repo: /plugin marketplace add anthropics/skills, then /plugin install example-skills@anthropic-agent-skills — this bundles skill-creator with the other 18 example skills.
3Manual option: clone anthropics/skills and copy the skills/skill-creator/ folder into ~/.claude/skills/ (available in all projects) or <project>/.claude/skills/ (one project).
4Claude.ai: Anthropic's example skills, skill-creator included, are already available on paid plans — turn on Skills under Settings > Capabilities; the same screen accepts a zipped custom skill if you want to upload a modified copy.
5Invoke it with /skill-creator, or just say "create a skill that does X" — the description is written to trigger on skill-creation requests.

What it does

Create: Interviews you to capture intent — what the skill should do, when it should trigger, expected output format — then drafts a SKILL.md with YAML frontmatter and a progressive-disclosure structure (metadata always in context, body under 500 lines, bundled scripts/references/assets loaded only as needed).
Eval: Writes 2-3 realistic test prompts to evals/evals.json, then spawns paired subagent runs per test case — one with the skill, one baseline without — and renders the results in an HTML review page via eval-viewer/generate_review.py so you judge outputs side by side.
Improve: Three bundled agent definitions do the scoring: a Grader evaluates outputs against assertions, a Comparator runs blind A/B comparisons between skill versions, and an Analyzer reads results and suggests targeted rewrites for the next iteration.
Benchmark: Repeats eval runs across multiple passes and aggregates quantitative metrics with variance analysis (scripts/aggregate_benchmark.py), so you can tell a real improvement from run-to-run noise.
Description tuning and packaging: A dedicated improve_description.py script optimizes the frontmatter description for triggering accuracy, plus quick_validate.py for structural checks and package_skill.py to zip the finished skill for upload.

Common questions

Is skill-creator official and free?: Yes. It's written and shipped by Anthropic in the anthropics/skills repository and the official Claude Code plugin directory, and the skill's folder carries an Apache-2.0 license. It's a reference implementation, though — the repo disclaimer says it's for demonstration and educational purposes.
Do I need skill-creator to create a Claude skill?: No. A skill is just a folder containing a SKILL.md file with YAML frontmatter and instructions, and you can write one by hand in minutes. Skill-creator adds the parts people skip: an intent interview, eval runs against a no-skill baseline, benchmarking, and description tuning for triggering accuracy.
Does it work in Claude.ai or only Claude Code?: Both. Claude Code is its primary home, installed as a plugin or copied into ~/.claude/skills/. On Claude.ai, Anthropic's example skills including skill-creator are already available on paid plans once Skills is enabled in Settings, and the API accepts uploaded skills too.
What are the four modes?: Create drafts a new SKILL.md from an interview about intent, triggers, and output format. Eval runs your test prompts through paired subagents — with and without the skill — and shows results in an HTML viewer. Improve uses Grader, Comparator, and Analyzer agents to score outputs and suggest rewrites, and Benchmark repeats runs for variance analysis.

Before you install

Anthropic's own repo README says the skills are "provided for demonstration and educational purposes only" and that behavior in production Claude may differ — this is a reference implementation, not a supported product.
The anthropics/skills repo has no repository-level license; skill-creator carries its own Apache-2.0 LICENSE.txt in its folder, while sibling document skills (docx, pdf, pptx, xlsx) are source-available, not open source. Check the folder you're copying.
The last commit touching skills/skill-creator landed 2026-04-20, and it was license housekeeping — the skill's actual content has gone months without a substantive update.
Eval and Benchmark modes spawn two subagent runs per test case (with-skill plus baseline) and Benchmark repeats runs for variance — a full iteration loop is token-hungry on metered API or usage-limited plans.
The SKILL.md explicitly instructs Claude to write "pushy" trigger descriptions to combat under-triggering; skills it authors can over-trigger until you tune the description back down yourself.
The eval/benchmark tooling is a set of Python scripts — you need a working Python environment in the session for anything beyond drafting the SKILL.md.
It ships from two Anthropic repos (anthropics/skills and anthropics/claude-plugins-official), and the copies can drift; the plugin descriptions already differ slightly between the two.

Facts come from the skill's own repository and docs on the date above, via the GitHub API — not from memory, and never from a sponsor.

Worth comparing

Hand-authoring from the anthropics/skills template — A skill is a folder with a SKILL.md; the repo's template/ and spec/ folders are enough for simple skills with zero install.
Superpowers (obra/superpowers) — Community skills framework (280,319 stars, actively pushed) with its own opinionated skill-writing workflow — heavier methodology, not Anthropic-maintained.
Claude Code skills documentation — Anthropic's official reference for where skills live and how they load — the read-first option if you want understanding rather than tooling.

← All MCP servers & skills