skill-auditor

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Skill Auditor

Workflow

工作流程

Step 1: Discover All Skills

步骤1：发现所有Skill

Scan these locations for

SKILL.md

files:

Level	Path
Global	`~/.claude/skills/*/SKILL.md`
Project	`.claude/skills/*/SKILL.md`
Plugin	`plugins//skills//SKILL.md`

For each skill found, collect:

Name: from frontmatter
```
name:
```
field (fall back to directory name)
Level: global / project / plugin (include plugin name for plugin skills)
Lines: total line count of
```
SKILL.md
```
Has allowed-tools: yes/no
Description length: character count of
```
description:
```
field

扫描以下位置的

SKILL.md

文件：

层级	路径
全局	`~/.claude/skills/*/SKILL.md`
项目	`.claude/skills/*/SKILL.md`
插件	`plugins//skills//SKILL.md`

针对每个找到的Skill，收集以下信息：

名称：从前置元数据的
```
name:
```
字段获取（若不存在则回退到目录名称）
层级：全局 / 项目 / 插件（插件Skill需包含插件名称）
行数：
```
SKILL.md
```
的总行数
是否包含allowed-tools：是/否
描述长度：
```
description:
```
字段的字符数

Step 2: Present Summary Table

步骤2：展示汇总表格

Sort by line count descending. Format:

| # | Level | Plugin | Skill Name | Lines | Status |

Status indicators:

```
OVER
```
— exceeds 500-line limit
```
HEAVY
```
— 300-499 lines (approaching limit)
```
OK
```
— under 300 lines

Include totals:

Total skills per level
Total lines across all skills
Average lines per skill
Skills exceeding limits

按行数降序排列，格式如下：

| # | Level | Plugin | Skill Name | Lines | Status |

状态标识：

```
OVER
```
— 超过500行限制
```
HEAVY
```
— 300-499行（接近限制）
```
OK
```
— 少于300行

汇总信息包含：

各层级的Skill总数
所有Skill的总行数
平均每个Skill的行数
超出限制的Skill数量

Step 3: Ask User for Selection

步骤3：请求用户选择

Use

AskUserQuestion

to ask which skills to review. Suggest:

All skills marked OVER or HEAVY
Top 5 largest skills by line count
Option to review by plugin name
Option to review all skills at a specific level

使用

AskUserQuestion

询问用户要审核哪些Skill，建议选项：

所有标记为OVER或HEAVY的Skill
行数最多的前5个Skill
按插件名称筛选审核
审核特定层级的所有Skill

Step 4: Deep Review (per selected skill)

步骤4：深度审核（针对每个选中的Skill）

Read each selected skill fully. Evaluate against these criteria:

Conciseness (token efficiency)

Lines that don't change LLM behavior (fluff, attribution, personas)
Redundant explanations of concepts Claude already knows
Verbose examples that could be compressed
Sections that repeat CLAUDE.md rules

Clarity

Ambiguous instructions that could be interpreted multiple ways
Missing context that forces Claude to guess
Inconsistent terminology within the skill

Scope Overlap

Compare skill's purpose against other skills at same level
Flag skills that cover substantially similar ground
Identify candidates for merging or splitting

Structure

Frontmatter completeness (name, description, allowed-tools)
Description quality (too short = undiscoverable, too long = wasteful)
Section organization (follows marketplace conventions?)

完整读取每个选中的Skill，根据以下标准进行评估：

简洁性（Token效率）

不影响LLM行为的行（冗余内容、署名、角色设定）
Claude已熟知概念的重复解释
可压缩的冗长示例
重复CLAUDE.md规则的章节

清晰度

可能产生多种解读的模糊指令
导致Claude需要猜测的缺失上下文
Skill内部术语不一致的问题

范围重叠

对比同层级其他Skill的用途
标记功能大幅重叠的Skill
识别可合并或拆分的候选Skill

结构

前置元数据完整性（名称、描述、allowed-tools）
描述质量（过短=难以被发现，过长=浪费Token）
章节组织（是否遵循市场规范）

Step 5: Report Findings

步骤5：输出审核结果

For each reviewed skill, output:

undefined

针对每个被审核的Skill，输出以下格式内容：

undefined

{skill-name} ({lines} lines)

Verdict: {TRIM | RESTRUCTURE | MERGE | OK}

Issues:

[CONCISENESS] {specific finding with line reference}
[CLARITY] {specific finding}
[OVERLAP] overlaps with {other-skill}: {shared scope}

Suggested savings: ~{N} lines ({percentage}% reduction) Recommended actions:

{specific action}
{specific action}

undefined

结论: {TRIM | RESTRUCTURE | MERGE | OK}

问题:

[CONCISENESS] {带行号的具体问题}
[CLARITY] {具体问题}
[OVERLAP] 与{other-skill}重叠：{共享范围}

预计可节省行数: ~{N}行（减少{percentage}%） 建议操作:

{具体操作}
{具体操作}

undefined

Step 6: Summary

步骤6：总结

After all reviews, provide:

Total potential line savings
Skills recommended for merging (with rationale)
Priority order for improvements (highest token savings first)

所有审核完成后，提供以下内容：

预计总行数节省量
建议合并的Skill及理由
优化优先级（按Token节省量从高到低排序）