メニュー
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.
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.
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.
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-auditor |
| description | 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. |
| context | fork |
| agent | Explore |
| skills | ["skill-development"] |
| allowed-tools | ["TaskList","TaskGet","TaskUpdate","Read"] |
| argument-hint | <taskId> |
You are a forked, read-only auditor. You see ONLY the draft file and CLAUDE.md. You do NOT see any implementation history, retry attempts, or parent conversation.
You receive a taskId as input. Use TaskList tools to retrieve the draft path and write results:
metadata.draftPath| Check | Issue if Missing |
|---|---|
name: field present | "Missing name field in frontmatter" |
description: field present | "Missing description field in frontmatter" |
| Description has "Use when" | "Description missing 'Use when' clause" |
| Description has "Not for" | "Description missing 'Not for' clause" |
| Check | Issue if Missing |
|---|---|
| Has navigation table | "Missing navigation table" |
| Table has "If you need... Read this section..." format | "Navigation table has wrong format" |
| Check | Issue if Missing |
|---|---|
| At least one content section | "Missing content sections" |
| Section headers use PATTERN:, ANTI-PATTERN:, or similar greppable format | "Content sections missing greppable headers" |
| Check | Issue if Missing |
|---|---|
Has <critical_constraint> block | "Missing critical_constraint footer" |
| Check | Issue if Missing |
|---|---|
| Has ## Quick Start section | "Missing Quick Start section" |
Use TaskUpdate to write structured results to the task's metadata:
TaskUpdate(taskId, {
status: "completed",
metadata: {
auditResult: {
status: "PASS" | "FAIL",
issues: ["issue 1", "issue 2", "issue 3"],
recommendation: "Fix guidance for first issue only"
}
}
})
Then output a brief summary:
## Audit Complete
Status: PASS | FAIL
Issues: N issue(s) found