メニュー
Create a new agent skill following the Agent Skills open standard
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Health check procedures D1–D14 for the Audit agent — structural validation, attention budget, version checks, workspace integrity, and static audit
Configure and manage Model Context Protocol servers for external tool access
Review a UI for accessibility — WCAG 2.1 AA compliance, semantic HTML, ARIA usage, keyboard navigation, focus management, colour contrast, and screen reader compatibility
Design or review a REST or GraphQL API — resource modeling, versioning strategy, error contract, OpenAPI/schema-first workflow, and security baseline
Generate a CHANGELOG.md entry from staged changes, a commit range, or a PR diff — following Keep a Changelog format with conventional commit classification
Set up and audit environment variable management — create .env.example, add startup validation, separate secrets from config, and document every variable
| name | skill-creator |
| description | Create a new agent skill following the Agent Skills open standard |
| compatibility | >=1.4 |
Skill metadata: version "1.2"; license MIT; tags [meta, authoring, skill, scaffold]; compatibility ">=1.4"; recommended tools [codebase, editFiles, runCommands].
Create a new agent skill that follows the Agent Skills open standard and the project's §12 Skill Protocol.
Tip: VS Code 1.110+ has
/create-skillfor basic scaffolds. This skill adds Lean/Kaizen guidance: waste-aware naming, PDCA verification, quality gates.
Clarify scope — Ask the user: "What workflow should this skill encode? Describe the trigger and the desired outcome in one sentence."
Choose a name — Verb-noun kebab phrase (review-dependencies, scaffold-api-route). Becomes directory name under .github/skills/. Rules: 1–64 chars, lowercase alphanumeric + hyphens, no leading/trailing/consecutive hyphens, must match directory.
Write the frontmatter — Create .github/skills/<name>/SKILL.md:
---
name: <kebab-name>
description: <one precise sentence - this is how the agent discovers the skill>
compatibility: ">=<current template version>"
---
Optional frontmatter fields (add only when relevant):
| Field | When to include |
|---|---|
user-invocable | Set to false to hide from the / menu while still allowing auto-load |
disable-model-invocation | Set to true to require manual / invocation only |
allowed-tools | Space-delimited pre-approved tools (experimental) |
compatibility | Environment requirements (e.g., ">=3.2", "Requires Python 3.14+") |
Do not add unsupported top-level keys like stacks. Put stack hints in description or the metadata note.
Add a metadata note after the # <Name> heading:
> Skill metadata: version "1.0"; license MIT; tags [<2-5 keywords>]; compatibility ">={{version}}"; recommended tools [codebase, editFiles].
Write the body — Title, When to use (triggers + contra-indications), Steps (numbered, action verbs, independently verifiable), Verify.
Authoring rules (§12): one skill = one workflow; no hardcoded paths; idempotent; steps not prose.
Progressive disclosure: metadata ~100 tokens at startup; instructions <5000 tokens on activation; resources on demand in scripts/, references/, assets/ subdirs. Keep under 500 lines.
Save the file.
Validate — Run skills-ref validate .github/skills/<name> if available, otherwise verify manually.
Run tests — Verify new skill passes applicable test suite checks.
.github/skills/<name>/SKILL.md exists with name and description frontmattername matches directory name (lowercase, hyphens)