| name | skill-creator |
| description | Create a new agent skill following the Agent Skills open standard |
| compatibility | >=1.4 |
Skill Creator
Skill metadata: version "1.1"; 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.
When to use
- The user asks to "create a skill", "write a skill", or "add a new skill"
- A workflow is being repeated manually and would benefit from codification
- An online skill was found but needs significant adaptation
Tip: VS Code 1.110+ has a built-in /create-skill slash command that generates a basic scaffold. This skill provides additional Lean/Kaizen guidance: waste-aware naming, PDCA verification steps, and quality gate checks.
Steps
-
Clarify scope — Ask the user: "What workflow should this skill encode? Describe the trigger and the desired outcome in one sentence."
-
Choose a name — Use a verb-noun kebab phrase describing the workflow (e.g., review-dependencies, scaffold-api-route). The name becomes the directory name under .github/skills/.
-
Write the frontmatter — Create .github/skills/<name>/SKILL.md with the minimal VS Code-compatible header, then record the richer metadata directly under the title:
---
name: <kebab-name>
description: <one precise sentence - this is how the agent discovers the skill>
---
Then add a note immediately after the # <Name> heading:
> Skill metadata: version "1.0"; license MIT; tags [<2-5 keywords matching common task descriptions>]; compatibility ">=<current template version>"; recommended tools [codebase, editFiles].
-
Write the body — Structure as:
- Title (
# <Name>) — human-readable heading.
- When to use — bullet list of trigger conditions and contra-indications.
- Steps — numbered list with clear action verbs. Each step should be independently verifiable.
- Verify — a final step that confirms the skill completed correctly.
-
Apply authoring rules (from §12):
- One skill, one workflow — if you need "and", split it.
- No hardcoded paths — use relative references and contextual lookups.
- Idempotent — running the skill twice produces the same result.
- Steps, not prose — the agent follows these literally.
-
Save — Write the file.
-
Run tests — Verify the new skill file passes any applicable test suite checks.
Verify