Author: Anthropic
Claude Skill Creator — Author and evaluate new Agent Skills
Design, package, and eval a new Agent Skill (SKILL.md + scripts/resources) so Claude loads durable workflow guidance — instead of pasting the same mega-prompt into every chat.
- What it does
- Walks Claude through skill authoring: clear trigger description, scoped instructions, optional scripts, and evaluation so the skill fires when it should and stays quiet when it should not.
- Best for
- Teams turning a repeated playbook (brand rules, release checklist, internal API patterns) into a reusable Claude skill folder.
- How to install / use
- Enable skill-creator from anthropics/skills (example-skills / Apache-2.0). State the job-to-be-done, trigger phrases, and non-goals before drafting SKILL.md.
Use this when the same prompt keeps getting pasted — graduate it into a skill folder.
Copy this
- Write triggers as when-to-use + when-not-to-use before long instruction prose
- Keep SKILL.md lean; put heavy reference docs in linked files
- Add eval prompts that prove the skill activates and stays out of unrelated tasks
When to use
Use Anthropic’s skill-creator skill when you are authoring a new Agent Skill — a folder with SKILL.md (and optional scripts/resources) that Claude should load for a repeatable job — not when you only need a one-off answer in the current thread.
Good fits:
- Codifying a team playbook (release notes, support macros, design constraints).
- Packaging scripts + instructions so the agent runs a checklist the same way every time.
- Tightening an existing skill’s description so it triggers on the right tasks only.
- Adding lightweight evals that prove activation and non-activation cases.
Skip it when you should just use an existing public skill (docx/pdf/frontend-design/etc.), or when a Cursor project rule / plain prompt is enough for a personal habit.
Why
Mega-prompts rot: people forget to paste them, they balloon, and they never get evals. Anthropic’s public skill-creator skill (in anthropics/skills) pushes you toward a proper skill package — description, scoped body, resources, and evaluation — aligned with the Agent Skills model.
Baonat marks this verified. Free to clone (Apache-2.0 example skill). Running authoring inside Claude still follows your host plan.
Install
- Get the skill from
skills/skill-creator, or install Anthropic’s example-skills plugin where available. - Define the job: one sentence outcome, trigger phrases, explicit non-goals.
- Draft
SKILL.mdwith a crisp frontmatter description (this is what decides loading). - Add only necessary scripts/references — keep the main file scannable.
- Write eval cases: should-trigger and must-not-trigger prompts.
- Install/attach the new skill in Claude Code / your skills path and run the evals.
Copy this
Kickoff prompt:
Attach the skill-creator skill. I want a new Agent Skill named [name].
Job-to-be-done: [one sentence].
Should trigger when: [phrases/situations].
Must NOT trigger when: [situations].
Audience: [team]. Output artifacts: [files/format].
First: propose description + outline of SKILL.md sections + eval prompt list (activate + stay-quiet).
After approval: generate the skill folder. Keep SKILL.md lean; link heavy refs as separate files.
Concrete scenarios
- Brand skill. Encode colors, type, and banned patterns for marketing pages.
- Release checklist skill. Steps + script hooks for changelog and version bumps.
- Support macros skill. Tone + required fields for ticket replies — with stay-quiet evals for coding tasks.
Pitfalls / when NOT to use
- Kitchen-sink skills. One job per skill; split when triggers conflict.
- Vague descriptions. If the description is fuzzy, the skill loads on the wrong tasks.
- No evals. Ship at least a few activate / stay-quiet prompts before sharing widely.
FAQ
Q: Official Anthropic skill?
A: Yes — skill-creator under anthropics/skills / example-skills.
Q: License?
A: Apache-2.0 — free to clone and adapt.
Q: Relates to Baonat cards?
A: Baonat recommends skills others wrote. Use skill-creator when you are authoring the next one.