Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
Skill Conflict Auditor
A cross-platform Agent Skill — runs in Claude Code · Codex · Cursor (and Copilot CLI / Gemini CLI) — that audits a set of AI skills/rules for conflicts, overlaps, redundancy, unclear priority, and safety risks: the problems that only appear when multiple skills coexist.
Supported tools: Claude Code · Codex · Cursor · Copilot CLI · Gemini CLI Languages: English · 中文
English
What it is
skill-conflict-auditor is a cross-platform Agent Skill. It runs natively in Claude Code, Codex, Copilot CLI, and Gemini CLI via SKILL.md, and ships as a Cursor rule (.mdc) for Cursor. Most skill problems are not inside one skill — they live in the relationships between skills. A skill that is perfect alone can become unsafe or unstable next to another one. This skill evaluates those interactions.
Give it two or more SKILL.md files (or descriptions, or folder trees) and it produces a structured audit: a skill inventory, severity-ranked conflict findings with cited evidence, merge suggestions, priority rules to add, rewritten descriptions, and a final recommendation.
When to use it
- You have multiple
SKILL.mdfiles and want to know "do these conflict?" - You are designing a group of skills meant to work together.
- Installed skills behave unpredictably (wrong skill triggers, inconsistent output, contradictory actions).
- You want to merge, split, rename, or reorganize skills.
- You are building a personal or team skill library and want a coherence/safety pass.
Not for: auditing a single skill alone, reviewing application source code, writing prose, or optimizing one standalone prompt.
What it checks (conflict categories)
| # | Category | Example |
|---|---|---|
| 1 | Trigger overlap | Two skills fire on the same task with no tiebreaker |
| 2 | Goal conflict | Speed vs. conservative/thorough review |
| 3 | Priority conflict | No rule for which skill wins when they disagree |
| 4 | Workflow conflict | One outputs first, another requires verification first |
| 5 | Output format conflict | Competing mandated output structures |
| 6 | Tool / script conflict | One formats the whole repo, another forbids it |
| 7 | Safety boundary conflict | One auto-submits/deletes, another requires confirmation |
| 8 | Evidence standard conflict | One allows speculation, another requires evidence |
| 9 | Scope conflict | A broad skill absorbs a more specific one |
| 10 | Redundancy | Two skills do nearly the same thing |
| 11 | Naming ambiguity | Names too vague or too similar |
| 12 | Hidden assumption conflict | Different assumed platforms/environments/users |
| 13 | Maintenance risk | Duplicated rules that will drift over time |
Severity levels
- High — can cause unsafe action, data loss, incorrect submission/sending/deletion, or wrong technical judgment.
- Medium — unstable output, workflow confusion, inconsistent formatting, duplicate or inefficient work.
- Low — naming, wording, mild overlap, or minor redundancy.
Output format
The audit always uses this fixed structure:
## Overall assessment
## Skill inventory (table)
## Conflict findings (severity-ranked table with cited evidence)
## Redundancy and merge suggestions
## Priority rules to add
## Revised descriptions
## Recommended directory or naming changes
## Final recommendationInstallation
One repo, two formats: a SKILL.md (Claude Code, Codex, Copilot CLI, Gemini CLI) and a Cursor .mdc rule. Use the lines for your tool.
#### Claude Code
# user-level (all projects)
git clone https://github.com/lawrence3699/skill-conflict-auditor.git \
~/.claude/skills/skill-conflict-auditor
# or project-level (current repo only)
git clone https://github.com/lawrence3699/skill-conflict-auditor.git \
.claude/skills/skill-conflict-auditor#### Codex
Codex loads SKILL.md natively. Pick one location (most → least portable):
# Cross-runtime (recommended) — also discovered by Copilot CLI and Gemini CLI
git clone https://github.com/lawrence3699/skill-conflict-auditor.git \
~/.agents/skills/skill-conflict-auditor
# Codex user-level home (Codex-specific)
git clone https://github.com/lawrence3699/skill-conflict-auditor.git \
~/.codex/skills/skill-conflict-auditor
# Codex project-level — scanned from the cwd up to the repo root
git clone https://github.com/lawrence3699/skill-conflict-auditor.git \
<your-repo>/.agents/skills/skill-conflict-auditor~/.codex/skills/ is Codex's documented user-level home; ~/.agents/skills/ is the shared cross-runtime convention. See the Codex skills docs.
#### Cursor
Cursor uses project rules, not SKILL.md. Drop the .mdc rule into your project's .cursor/rules/:
mkdir -p .cursor/rules
curl -fsSL https://raw.githubusercontent.com/lawrence3699/skill-conflict-auditor/main/.cursor/rules/skill-conflict-auditor.mdc \
-o .cursor/rules/skill-conflict-auditor.mdcIt is an Agent Requested rule: description only, no globs. Cursor's agent loads it by description when you ask to audit skills/rules. (A rule that sets globs is instead treated as Auto-Attach and only loads when a matching file is open — which is why this rule deliberately omits them.) Check Cursor Settings → Rules to confirm it loaded.
After installing on any tool, restart the agent / start a new session so it discovers the skill.
Usage
Once installed, just ask in natural language and paste your skills:
Audit these two skills for conflicts:
<paste SKILL.md #1>
<paste SKILL.md #2>The agent invokes skill-conflict-auditor and returns the structured audit. See examples/ for two example skills (pr-fast-writer and pr-careful-reviewer) that demonstrate a High-severity safety-boundary conflict.
Repository layout
skill-conflict-auditor/
├── SKILL.md # Canonical skill (Claude Code, Codex, Copilot CLI, Gemini CLI)
├── .cursor/
│ └── rules/
│ └── skill-conflict-auditor.mdc # Cursor rule — GENERATED from SKILL.md (do not edit by hand)
├── AGENTS.md # Notes for agents working on this repo
├── README.md # This file (EN + 中文)
├── LICENSE # MIT
├── .gitignore
├── Makefile # make build · make check · make eval
├── .github/workflows/ci.yml # Static checks on every push/PR
├── scripts/ # Build + regression tooling (stdlib Python, no deps)
│ ├── build_cursor_rule.py # regenerate the .mdc from SKILL.md
│ ├── check_sync.py # fail if the .mdc drifted from SKILL.md
│ ├── validate_frontmatter.py # frontmatter / spec checks
│ ├── check_examples.py # sample-output shape contract
│ └── eval_audit.py # optional live LLM eval (API-key gated)
└── examples/
├── pr-fast-writer.md # Example skill A (speed-oriented)
├── pr-careful-reviewer.md # Example skill B (conservative)
└── sample-audit-output.md # Illustrative audit of A + B (shape-checked in CI)The Cursor.mdcis generated from `SKILL.md` (make build) — same body, Cursor frontmatter. EditSKILL.md, never the.mdc; CI fails if they drift (scripts/check_sync.py).
Development / tests
No app, no build toolchain — just the skill and a small stdlib-Python harness (zero dependencies).
make build # regenerate .cursor/rules/*.mdc from SKILL.md (run after editing SKILL.md)
make check # CI suite: frontmatter validation, SKILL.md↔.mdc sync, sample-output contract
make eval # optional live eval; no-ops unless ANTHROPIC_API_KEY or OPENAI_API_KEY is setmake check runs in CI on every push/PR and is fully offline. make eval actually runs the auditor over the example pair and asserts the same shape invariants (set EVAL_MODEL to choose the model).
Contributing
Issues and PRs welcome. If you find a conflict category the auditor misses, open an issue with a concrete two-skill example that triggers it.
License
中文
这是什么
skill-conflict-auditor 是一个跨平台的 Agent 技能（Skill）。它通过 SKILL.md 在 Claude Code、Codex、Copilot CLI、Gemini CLI 中原生运行，并以 Cursor 规则（.mdc）的形式适配 Cursor。技能的大多数问题并不藏在单个技能内部，而是出现在技能与技能之间的关系里——一个单独看完美的技能，放到另一个技能旁边可能变得不安全或不稳定。这个技能专门审查这些“相互作用”。
把两个或更多 SKILL.md 文件（或描述、或目录结构）交给它，它会输出一份结构化审查报告：技能清单、按严重程度排序并附带证据引用的冲突清单、合并建议、需要补充的优先级规则、重写后的描述，以及最终建议。
何时使用
- 你有多个
SKILL.md文件，想知道“它们会不会冲突？” - 你正在设计一组需要协同工作的技能。
- 安装多个技能后行为不稳定（触发了错误的技能、输出不一致、动作互相矛盾）。
- 你想合并、拆分、重命名或重新组织技能。
- 你正在搭建个人或团队的技能库，需要做一次一致性 / 安全性检查。
不适用于： 单独审查一个技能、审查应用源代码、写文章、或优化单个独立 prompt。
检查哪些冲突（冲突类别）
| # | 类别 | 示例 |
|---|---|---|
| 1 | 触发条件重叠 | 两个技能在同一任务上都会被触发，且没有优先规则 |
| 2 | 目标冲突 | 追求速度 vs. 保守、严谨的审查 |
| 3 | 优先级冲突 | 两者规则矛盾时，没有规定谁优先 |
| 4 | 工作流冲突 | 一个先出结果，另一个要求先验证 |
| 5 | 输出格式冲突 | 对同一产物强制要求互斥的输出结构 |
| 6 | 工具 / 脚本冲突 | 一个格式化整个仓库，另一个禁止改动无关文件 |
| 7 | 安全边界冲突 | 一个自动提交 / 删除，另一个要求用户确认 |
| 8 | 证据标准冲突 | 一个允许推测，另一个要求有证据才能下结论 |
| 9 | 作用域冲突 | 范围过宽的技能吞掉了更专门的技能 |
| 10 | 冗余 | 两个技能几乎做同一件事 |
| 11 | 命名歧义 | 名称过于模糊或彼此过于相似 |
| 12 | 隐藏假设冲突 | 默认假设了不同的平台 / 环境 / 用户 |
| 13 | 维护风险 | 多个技能重复了同一条规则，日后会逐渐不一致 |
严重程度分级
- High（高） —— 可能导致不安全操作、数据丢失、错误提交 / 发送 / 删除，或错误的技术判断。
- Medium（中） —— 输出不稳定、工作流混乱、格式不一致、重复或低效的工作。
- Low（低） —— 命名、措辞、轻微重叠或次要冗余。
输出格式
审查报告固定使用以下结构：
## Overall assessment 总体评估
## Skill inventory 技能清单（表格）
## Conflict findings 冲突清单（按严重程度排序、附证据的表格）
## Redundancy and merge suggestions 冗余与合并建议
## Priority rules to add 需要补充的优先级规则
## Revised descriptions 重写后的描述
## Recommended directory or naming changes 建议的目录 / 命名调整
## Final recommendation 最终建议安装
同一个仓库提供两种格式：SKILL.md（Claude Code、Codex、Copilot CLI、Gemini CLI）和 Cursor 的 .mdc 规则。按你使用的工具选择对应命令。
#### Claude Code
# 用户级（对所有项目生效）
git clone https://github.com/lawrence3699/skill-conflict-auditor.git \
~/.claude/skills/skill-conflict-auditor
# 或项目级（仅当前仓库）
git clone https://github.com/lawrence3699/skill-conflict-auditor.git \
.claude/skills/skill-conflict-auditor#### Codex
Codex 原生加载 SKILL.md。可选位置（从最通用到最特定）：
# 跨运行时（推荐）—— Copilot CLI、Gemini CLI 也会发现
git clone https://github.com/lawrence3699/skill-conflict-auditor.git \
~/.agents/skills/skill-conflict-auditor
# Codex 用户级 home（Codex 专用）
git clone https://github.com/lawrence3699/skill-conflict-auditor.git \
~/.codex/skills/skill-conflict-auditor
# Codex 项目级 —— 从当前目录向上扫描到仓库根
git clone https://github.com/lawrence3699/skill-conflict-auditor.git \
<你的仓库>/.agents/skills/skill-conflict-auditor~/.codex/skills/ 是 Codex 文档中的用户级 home；~/.agents/skills/ 是跨运行时共享约定。参见 Codex skills 文档。
#### Cursor
Cursor 使用项目规则（Project Rules），不读取 SKILL.md。把 .mdc 规则放进你项目的 .cursor/rules/：
mkdir -p .cursor/rules
curl -fsSL https://raw.githubusercontent.com/lawrence3699/skill-conflict-auditor/main/.cursor/rules/skill-conflict-auditor.mdc \
-o .cursor/rules/skill-conflict-auditor.mdc它是一条 Agent Requested 规则：只有 description、不设 globs。当你要求审查技能/规则时，Cursor 的 agent 会根据 description 主动加载它。（若设置了 globs，Cursor 会把它当作 Auto-Attach，只在命中文件打开时加载——所以本规则刻意不写 globs。）可在 Cursor Settings → Rules 中确认它已加载。
任意工具安装后，请重启 agent（或开启新会话）以加载该技能。
使用方法
安装后，用自然语言提问并粘贴你的技能即可：
帮我审查这两个技能是否冲突：
<粘贴 SKILL.md #1>
<粘贴 SKILL.md #2>Agent 会调用 skill-conflict-auditor 并返回结构化报告。可参考 examples/ 中的两个示例技能（pr-fast-writer 与 pr-careful-reviewer），它们演示了一个 High 级别的安全边界冲突。
目录结构
skill-conflict-auditor/
├── SKILL.md # 规范源（Claude Code / Codex / Copilot CLI / Gemini CLI）
├── .cursor/
│ └── rules/
│ └── skill-conflict-auditor.mdc # Cursor 规则 —— 由 SKILL.md 生成（请勿手改）
├── AGENTS.md # 给在本仓库内工作的 agent 的说明
├── README.md # 本文件（英文 + 中文）
├── LICENSE # MIT
├── .gitignore
├── Makefile # make build · make check · make eval
├── .github/workflows/ci.yml # 每次 push/PR 跑静态检查
├── scripts/ # 构建 + 回归工具（标准库 Python，无依赖）
│ ├── build_cursor_rule.py # 从 SKILL.md 生成 .mdc
│ ├── check_sync.py # .mdc 与 SKILL.md 不一致则失败
│ ├── validate_frontmatter.py # frontmatter / 规范校验
│ ├── check_examples.py # 示例输出的结构契约
│ └── eval_audit.py # 可选的在线 LLM 评测（需 API key）
└── examples/
├── pr-fast-writer.md # 示例技能 A（速度优先）
├── pr-careful-reviewer.md # 示例技能 B（保守严谨）
└── sample-audit-output.md # A + B 的示例审查（CI 校验其结构）Cursor 的.mdc由 `SKILL.md` 生成（make build）—— 同正文、Cursor frontmatter。只改SKILL.md，不要改.mdc；若两者漂移，CI 会失败（scripts/check_sync.py）。
开发 / 测试
没有应用、没有构建工具链——只有技能本身和一套标准库 Python 工具（零依赖）。
make build # 从 SKILL.md 重新生成 .cursor/rules/*.mdc（改完 SKILL.md 后运行）
make check # CI 套件：frontmatter 校验、SKILL.md↔.mdc 同步、示例输出契约
make eval # 可选在线评测；未设置 ANTHROPIC_API_KEY / OPENAI_API_KEY 时自动跳过make check 在 CI 中每次 push/PR 运行，完全离线。make eval 会真正用示例对运行审查器并断言同样的结构不变量（用 EVAL_MODEL 选择模型）。
参与贡献
欢迎提交 Issue 和 PR。如果你发现审查器漏掉了某种冲突类别，请用一个能复现该问题的“两技能”示例提交 Issue。