메뉴
Create portable skills with SKILL.md. Use when building new skills or documenting patterns. Includes frontmatter syntax, quality standards, and primitives. Not for commands or agents.
Create single-file commands with dynamic content injection (@path and !command). Use when building commands that need filesystem access, git state, runtime context, or argument handling with $1 placeholders. Includes @file injection, !shell execution, <injected_content> wrappers, and single-file structure patterns. Not for skills (use skill-development) or audit workflows.
Create and refine single-file commands with @/! injection. Use when authoring new commands or improving existing ones. Includes namespacing rules, thin interface patterns, and delegation best practices. Not for skills or agents.
Master Claude Code's context primitives (Skill, Task, Command) and orchestration patterns. Use when designing agent workflows, understanding why some patterns work while others fail (e.g., Task→Task is forbidden but Skill→Skill is allowed), or implementing Chain of Experts architectures. Not for writing prompts or content generation.
Verify completion with evidence using 6-phase gates and three-way audits. Use when claiming task completion, committing code, or validating components. Not for skipping verification or bypassing quality gates.
Audit skill files against quality standards. Use ONLY within Manager Pattern workflow via TaskList. Receives taskId, reads draftPath from task metadata, validates against skill-development rules, writes auditResult to task metadata. Not for manual use.
Creates skills with automatic quality validation. Use when you need a new skill created with all quality gates enforced (frontmatter, navigation, critical_constraint). Uses TaskList for stateful retry loops with skill-auditor. Not for quick prototypes or when manual skill creation is preferred.
| name | skill-development |
| description | Create portable skills with SKILL.md. Use when building new skills or documenting patterns. Includes frontmatter syntax, quality standards, and primitives. Not for commands or agents. |
Put everything useful in SKILL.md. AI agents are smart (can infer from examples) and lazy (won't hunt for scattered references).
---
name: skill-name
description: "What it does. Use when {trigger + keywords}. Not for {exclusions}."
---
Description format: What it does, when to use it (with trigger phrases), what it's NOT for.
Optional fields: context: fork, agent:, skills:, disable-model-invocation, user-invocable
| Primitive | Context | Use When |
|---|---|---|
Task() | Forked | Heavy multi-step work |
Skill() | Shared | Adding expertise |
Skill(fork) | Isolated | Biased-free execution |
Key constraint: Task→Task is forbidden. Skill(fork)→Skill(fork) is allowed.
.claude/skills/skill-name/
└── SKILL.md
Folder and file names use kebab-case. References folder is RARE—needs strong justification.
| Check | Requirement |
|---|---|
| Frontmatter | name + description (What-When-Not format) |
| Quick Start | Scenario-based entry point |
| Navigation | "If you need X → Read Y" table |
| Core knowledge | In SKILL.md, not in references/ |
| Footer | critical_constraint for non-negotiable rules |
helper, utils)<critical_constraint> Portability Invariant: Every skill MUST work in isolation (zero external dependencies).
Gold Standards: