skill-creator

skill-creator is Anthropic's reference skill for authoring other skills. It walks an agent through the process of turning a recorded workflow or written procedure into a SKILL.md that conforms to the agentskills.io standard. This is the bridge between a one-off prompt workflow and a durable, redistributable skill.

If you've used OpenAI Codex's Record & Replay to capture a session, skill-creator is the equivalent downstream: take the artifact and emit a standards-compliant skill.

When to use

• You've done a multi-step task in a session more than once and want to package the procedure
• You maintain an internal skills catalog and need a consistent template
• You want a skill that codifies an existing runbook (incident response, code review, deploy checklist)

When not to use

• The task is genuinely one-off — just keep the prompt
• The procedure has decision points that an agent can't reliably navigate (escalation paths, ambiguous failure modes)

How it works

skill-creator ships with a structured template for emitting SKILL.md frontmatter (name, description, license, compatibility, metadata, allowed-tools) and a set of progressive-disclosure references for the most common authoring pitfalls (over-broad descriptions, scope creep, missing allowed-tools declarations). The agent is guided to write a skill that activates only on the intended trigger and stays within its declared capability scope.

Install

npx skills add anthropics/skills/skill-creator

Full SKILL.md

Full source at the canonical repo: github.com/anthropics/skills. Embedded copy omitted at launch; link above tracks the latest authoritative copy.

Verify it yourself

Quick static check — confirm allowed-tools in the SKILL.md frontmatter is bounded to what skill authoring actually needs (file read/write in a working directory; no network). The meta-skill pattern is attractive to attackers because a compromised skill-creator would emit malicious skills downstream — pay extra attention here.

Two-model review — paste the full SKILL.md into two different models with the prompt from the Vetting Guide. This skill is unusually important to vet because it generates other skills; injection here would propagate.

Review transparency

Related

• Skills Collection Overview
• Vetting Guide
• Skill Packs — the pack concept this skill feeds into
• Building Antigravity + OpenCode Skills