审计插件
安装方式
先用 Prompt 核对来源,也可切换到安装命令或下载可用的安装包。
用 AI 助手安装
复制提示词,让助手先核对固定版本和依赖。
查看 Prompt收起 Prompt
先检查当前 Agent 是否已安装 SkillPI Skill Finder。如果尚未安装,请按官方仓库的说明安装并启用:https://github.com/Findskills/skillpi。然后使用 Skill Finder 打开此 SkillPI 详情页并定位 "audit-plugin":https://skillpi.com/zh/skills/deepskill-167094638749fef50d74238b459b4aeaeade49eaf33bfcbb4efe8e94ca707d7d。安装前核对上游来源和版本:https://github.com/jeremylongshore/tons-of-skills-marketplace/blob/6ea46b08dce892944db7c16f28d0b5308d402402/plugins/productivity/claude-workflow-skills/skills/audit-plugin/SKILL.md。来源不一致时停止并说明差异。核对通过后,用 Skill Finder 安装目标 Skill,并报告安装路径和来源版本。
此命令从 GitHub 来源仓库获取指定技能。运行前请先查看源文件。
npx skills add https://github.com/jeremylongshore/tons-of-skills-marketplace/tree/6ea46b08dce892944db7c16f28d0b5308d402402/plugins/productivity/claude-workflow-skills/skills/audit-plugin --skill audit-plugin来源与配置
查看 GitHub 源码jeremylongshore/tons-of-skills-marketplace固定提交 6ea46b08dce892944db7c16f28d0b5308d402402
配置方式待核实
请查看原始来源中的适用工具、依赖和权限要求。
安装实测待补充
本条目介绍 审计插件 的公开用途。DeepSkill 当前公开目录说明:“对当前项目中定义的 Claude Code 插件、技能或子代理进行深入审查,依据官方最佳实践。将发现的问题记录为 GitHub 议题,并编写一份优先级排序的修复计划,写入项目 CLAUDE.md 文件。当用户说“审计此插件”、“审查此技能”、“检查此代理”或“审计此附加组件”时使用。” 适用场景包括写代码与代码审查、测试与安全校验。本页根据公开目录资料和固定版本的上游仓库整理,建议使用前核对该版本的环境、工具、权限及数据处理要求。Skillπ 提供站方用途介绍与原站链接;原始 SKILL.md 和附件仅在对应许可允许时开放。
文件浏览
---
name: audit-plugin
description: Performs a deep review of the Claude Code plugin, skill, or sub-agent defined in the current project against official best practices. Documents findings as GitHub issues and writes a prioritised fix plan to the project CLAUDE.md. Use when the user says audit this plugin, review this skill, check this agent, or audit addon.
allowed-tools: Read Glob Grep Bash WebSearch WebFetch
disable-model-invocation: true
---
# Audit Plugin
Project: !`basename $(git rev-parse --show-toplevel 2>/dev/null) 2>/dev/null || basename $PWD`
Branch: !`git branch --show-current 2>/dev/null || echo "unknown"`
Reviews the Claude Code addon in the current project (plugin, skill, sub-agent, or a combination)
against official Claude Code best practices. Generates actionable GitHub issues and a prioritised
fix plan.
## Step 0: Pre-flight check
```bash
gh auth status 2>&1 || { echo "ERROR: gh is not authenticated. Run: gh auth login"; exit 1; }
```
## Step 1: Identify what kind of addon this project defines
Scan for known Claude Code addon files:
```bash
find . -not -path './.git/*' \( \
-name 'plugin.json' -path '*/.claude-plugin/*' \
-o -name 'SKILL.md' -path '*/skills/*' \
-o -name '*.md' -path '*/agents/*' \
\) 2>/dev/null
```
Read each file found. Build a mental model of:
- Plugin manifest (`.claude-plugin/plugin.json`) — name, version, declared agents/skills
- Skills (`skills/<name>/SKILL.md`) — frontmatter fields, body structure, tool declarations
- Agents (`agents/<name>.md`) — frontmatter fields, description examples, tool restrictions
## Step 2: Fetch current best-practice documentation
Use WebFetch to retrieve up-to-date guidance from these known URLs:
- Skills: `https://code.claude.com/docs/en/skills`
- Sub-agents: `https://code.claude.com/docs/en/agents`
- Plugins: `https://code.claude.com/docs/en/plugins`
If any URL returns an error, use WebSearch to find the current equivalent under `code.claude.com`.
Summarise the key quality criteria from each source.
## Step 3: Evaluate against best practices
For **plugin.json**, check:
- Required fields present: `name`, `version`, `description`, `author`, `license`
- `minVersion` set to a current compatible value
- Keywords are relevant and searchable
- `description` is concise and accurate
For **each SKILL.md**, check:
- `description` is a single unbroken line under 1,536 characters (combined with `when_to_use`)
- `allowed-tools` is set and follows least-privilege (only tools the skill actually needs)
- Body is under 500 lines; large reference content moved to separate files
- Shell injection blocks (`` !`command` ``) are used where live context would help
- Step numbering is clear and actionable
- Code blocks specify a language
- Prose lines are ≤ 120 characters
- No personalised language in formal content (no "you", "your" in instructions)
For **each agent `.md`**, check:
- `description` is a quoted single-line string with proper `<example>` blocks for auto-delegation
- `model`, `color`, `maxTurns`, `memory`, `tools` and `initialPrompt` fields are present where appropriate
- Agent body is clear, focused, and actionable
- Description examples use current delegation language (no "Task tool" narration)
- `tools` list follows least-privilege
## Step 4: Generate GitHub issues
For each distinct finding, create a GitHub issue:
```bash
gh issue create \
--title "<type>: <brief description>" \
--body "$(cat <<'EOF'
## Finding
<description of the problem>
## Expected
<what best practice requires>
## Current
<what the file actually has>
## Suggested fix
<concrete change to make>
EOF
)" \
--label "enhancement"
```
Group closely related findings into a single issue where it makes sense.
Use `--label "bug"` for broken or non-compliant fields, `--label "enhancement"` for improvements.
Note the issue numbers as you go.
## Step 5: Write prioritised fix plan to CLAUDE.md
Append or update a section in the project `CLAUDE.md` under the heading
`## Audit Findings — <today's date>`:
```markdown
## Audit Findings — YYYY-MM-DD
Issues generated from `/audit-plugin` review. Suggested fix order:
### Group 1 — Correctness (fix first)
- #N: <title>
- #N: <title>
### Group 2 — Best-Practice Compliance
- #N: <title>
- #N: <title>
### Group 3 — Quality Improvements
- #N: <title>
- #N: <title>
```
Order groups by: correctness blockers first, then compliance, then polish.
## Step 6: Report summary
Output a brief summary:
- Total issues created (with links)
- Top-priority fix
- Link to the CLAUDE.md section added
当前收录版本
0.0.1
质量评测
满分 100。前四项来自 DeepSkill 收录的质量报告。
评分依据
高价值/高频刚需。直接解决用户明确的核心任务痛点,显著提升日常高频场景的效率,自动化审计 Claude Code 插件、技能或子代理,符合官方最佳实践
评分依据
触发词精准完整,覆盖多种变体,边界清晰
评分依据
Phase/步骤清晰,有出口条件,含端到端示例,frontmatter完整,单文件≤500行
评分依据
工具来源明确,参数正确,覆盖主要异常,不可逆操作有确认机制
计算方式
取前四项原始分值的算术平均,四舍五入取整。可通过收录报告核对各项评分与说明,运行记录需单独查证。
权限与使用前检查
根据当前主文档自动整理,命中内容可能是示例或防护说明;不是实际权限声明,也未覆盖附件、外部依赖或运行行为。
未从主文档识别到上述常见操作,不能据此判断无需权限或没有风险。
文档引用的域名或地址
这些地址可能用于文档链接或示例,不是已批准的联网清单。使用前核对实际目的地及发送的数据。
- code.claude.com
安全评级与访问范围
risk · 静态证据核验,不代表运行安全保证。
使用前请由宿主核对并授权实际文件读写、外部服务和联网目的地;本站的核验不会授予任何运行权限。
用户讨论
还没有已核实的点评。