Meta Skill Creator
Overview
Agent Skills are portable folders of instructions, scripts, and resources that agents discover and load on demand. They follow the Agent Skills open standard and work across 27+ compatible agents including Claude Code, Cursor, Gemini CLI, OpenAI Codex, VS Code, GitHub Copilot, Windsurf, Goose, and Roo Code. This skill covers the complete workflow for creating, structuring, and validating skills.
Quick Reference
| Concept | Summary | Key Rule |
|---|---|---|
| Directory structure | skill-name/SKILL.md plus optional references/, scripts/, assets/ | SKILL.md is the only required file |
| Frontmatter (required) | YAML block with name and description | name must match directory, max 64 chars |
| Frontmatter (optional) | license, compatibility, metadata, allowed-tools | Defined by open standard, portable across agents |
| Agent-specific fields | model, context, agent, hooks, user-invocable | Claude Code extensions, not portable |
| Description triggers | Include "Use when..." or "Use for..." phrases | Keyword-rich, third-person voice |
| Progressive disclosure | Metadata -> Instructions -> Resources (three tiers) | SKILL.md loaded in full; references on demand |
| SKILL.md size | Target 100-150 lines, max 500 | No code examples in SKILL.md |
| SKILL.md sections | Overview, Quick Reference, Common Mistakes, Delegation, References | All five sections required |
| Reference files | references/[topic].md with frontmatter | Max 500 lines each, self-contained |
| Reference frontmatter | title, description, tags fields | Required for all reference files |
| Scripts | scripts/ directory, executable code | Agents run scripts, not read source |
| Name format | Lowercase letters, numbers, hyphens | No --, no leading/trailing -, min 4 chars |
| Excluded filenames | README.md, metadata.json, _* files | Not installed by distribution CLI |
| Validation | skills-ref validate or project-specific validators | Always validate before distributing |
Skill Creation Workflow
- Create directory at the appropriate location with a
SKILL.md - Write frontmatter with
name(matching directory) and trigger-richdescription - Write required sections: Overview, Quick Reference, Common Mistakes, Delegation, References
- Extract code examples into
references/files (no code in SKILL.md) - Validate with your project validator or
skills-ref validate
Common Mistakes
| Mistake | Correct Pattern |
|---|---|
| Missing "Use when" or "Use for" trigger | Include trigger phrases in description for agent discovery |
| Code examples in SKILL.md | Move all code to references/ files |
| SKILL.md over 500 lines | Split detailed content into reference files |
| Reference files cross-referencing | Keep references self-contained, one level deep from SKILL.md |
| Name does not match directory | name field must exactly match parent directory name |
| Name contains uppercase or spaces | Use only lowercase letters, numbers, and hyphens |
| Vague description without keywords | Pack triggers with synonyms, abbreviations, library names |
| Code blocks without language specifier | Always specify language on fenced code blocks |
| Supporting files not linked in SKILL.md | Agents discover files through explicit links in SKILL.md |
| Using excluded filenames | Avoid README.md, metadata.json, files starting with _ |
| Blank line before frontmatter | --- must be on line 1, no preceding blank lines |
| Description in first-person voice | Use third-person: "Extracts text", not "I help you" |
| Deeply nested reference chains | Keep one level deep: SKILL.md links to references directly |
| Missing required SKILL.md sections | Include all five: Overview, Quick Ref, Mistakes, Delegation, Refs |
| Reference file missing frontmatter | Every reference needs title, description, and tags |
Skill Locations
Skills can be installed at different scopes depending on visibility needs:
| Location | Path | Scope |
|---|---|---|
| Personal | ~/.claude/skills/ | You, all projects |
| Project | .claude/skills/ | Anyone in repository |
| Distributed | Via skills CLI | Plugin users |
Delegation
- Specification questions: Refer to the Agent Skills open standard
- Pattern discovery: Use
Exploreagent to analyze existing skills for structure examples - Validation: Run project validator or
skills-ref validatebefore committing