Token导航 LogoToken导航

Agent Skill · 开发

code-guardian(代码 guardian)

code-guardian 用于处理 GitHub 仓库、Issue、Pull Request 和代码协作信息,适合在 Codex、Claude、Cursor、Gemini CLI 中需要围绕仓库状态、代码变更或协作事项进行整理时使用。可结合来源仓库、安装命令和原始 README 继续核验具体用法。安装前建议确认权限范围、维护状态,以及是否会触发联网、命令执行或文件读写。

code-guardian 是开发类 Agent Skill,可用宿主为 Codex、Claude、Cursor、Gemini CLI,通过GitHub 仓库获取,权限为需要联网;许可证信息还不明确,第三方扫描已通过,按安装量在开发类的 21,746 个 Skill 里排第 19,422。

总安装 272周安装 11GitHub Stars 公开资料未说明下载 85核验 2026-05-01

安装说明

本站只整理中文说明,不托管安装包,也不代用户安装。
GitHub
许可证
unknown

安装方式

通过对话安装

复制提示词发给支持本地命令或 Skills 的 AI 助手,先确认命令和权限,再让它执行。

请帮我安装这个 Agent Skill:code-guardian(代码 guardian)
来源仓库:https://github.com/ak68a/code-guardian-skill
仓库路径:skills/code-guardian
安装命令:
npx skills add https://github.com/ak68a/code-guardian-skill --skill code-guardian
安装前请先检查当前环境是否支持对应 CLI,并向我确认将要执行的命令、安装目录、联网范围和文件读写权限;确认后再执行。

命令行安装

复制命令到本机终端执行。该命令会通过 npx skills 从第三方来源获取 Skill;本站只展示命令,不托管安装包,也不自动执行。

skills.shnpx skills
npx skills add https://github.com/ak68a/code-guardian-skill --skill code-guardian

这个 Skill 是做什么的

开发类 Skill 面向写代码、改代码和排查问题,一般会读写仓库文件,部分还会执行构建或测试命令。

code-guardian 提供防止 AI 编码失败的协议,包括假设显式化与简化优先的开发 mindset。适用于高风险或复杂功能开发,强制在编码前陈述假设并在遇到疑问时主动质疑。内置 checkpoint 模式提供 cg 命令调用清单,确保关键步骤有复核点以减少错误累积。

本技能为行为规范而非工具,需在开发全程激活,不能替代代码审查与测试流程。使用前请熟悉其规则细节,避免机械套用;建议结合团队编码标准与项目管理流程协同使用。

来源仓库内路径:skills/code-guardian

上手步骤

4 步
  1. 1

    确认自己的宿主工具

    公开资料标注支持 Codex、Claude、Cursor、Gemini CLI;先确认自己用的是其中之一,再按对应工具的 Skills 目录放置。

  2. 2

    拿到安装命令

    可以直接复制的命令是 npx skills add https://github.com/ak68a/code-guardian-skill --skill code-guardia…;命令里的占位符换成自己的路径或密钥再执行。

  3. 3

    安装后重启宿主

    命令执行完要完全退出宿主工具再打开,让新装的 Skill 被重新扫描到,否则列表里可能看不到。

  4. 4

    按最小权限试用

    它需要需要联网,先在测试目录或测试账号里跑,确认没问题再放到正式项目。

常见问题

6 条

code-guardian 怎么安装?

在终端执行 npx skills add https://github.com/ak68a/code-guardian-skill --skill code-guardian。装完完全退出宿主工具再打开,让它重新扫描本地 Skill。

code-guardian 支持哪些 AI 工具?

公开资料标注支持 Codex、Claude、Cursor、Gemini CLI。其他工具只要能读取本地 Skill 目录,一般也能手动放进去,但要自己确认触发方式是否一致。

code-guardian 需要什么权限?

权限级别是需要联网。该 Skill 可能需要联网访问来源站点、仓库或外部 API;具体网络访问范围需要结合源码和 README 复核。

code-guardian 是免费的吗?

公开资料没有标注明确许可证,收费和商用限制都无法确认。正式使用前请到来源仓库核对授权说明。

code-guardian 用的人多吗?

公开数据是累计安装 272 次,近一周安装 11 次,下载 85 次。这些数字来自第三方目录统计,只能当作热度参考,不代表质量或安全性。

code-guardian 安全吗?

第三方扫描结果为通过,但这只覆盖已知的静态检查项,不能替代自己读一遍源码和安装命令。

值得选的地方和要先想清楚的地方

值得选的地方

  • 同时适配 4 种工具,换客户端不用重新找替代品。
  • 第三方扫描已通过,常见的高风险行为没有命中。

先想清楚的地方

  • 许可证没有明确标注,能不能商用、能不能再分发都不确定。
  • 第三方目录整理,版本更新和行为细节以来源仓库为准。

适不适合你

更适合

  • 已经在用 Codex,想直接扩展它能力的人。
  • 需要处理开发类任务、又不想每次都手写提示词的人。

不太适合

  • 需要明确授权才能用到公司项目里的场景。
  • 需要厂商级技术支持和 SLA 的生产系统。

开发类 Skill 对比

6 个同类
Skill总安装GitHub 星权限许可证扫描
skill-vetter

技能审查员 · OpenClaw

8,779,018986操作浏览器MIT-0通过
ontology

本体论 · OpenClaw

6,636,960576只读MIT-0通过
entra-app-registration

内部应用程序注册 · Codex、Claude

6,631,872764接触敏感数据未明确通过
entra-app-registration

内部应用程序注册 · Codex、Claude

3,757,680201接触敏感数据未明确通过
nano-banana-pro

纳米香蕉专业版 · OpenClaw

3,526,898364接触敏感数据MIT-0有提醒
code-guardian

代码 guardian · 本页

272—需要联网未明确通过

按安装量取同分类前几名;数字来自第三方目录统计,只作热度参考,不代表质量结论。

SKILL.md 原文

Code Guardian

Protocols to prevent common AI coding failures. This skill operates in two modes:

  1. Background Mode: Always-on mindset during coding (simplicity-first, surface assumptions, push back when needed)
  2. Checkpoint Mode: Explicit checklists via cg commands

BACKGROUND MODE (Always Active)

These rules apply during coding, not just before or after. Violating them mid-implementation is the same as violating them anywhere else.

1. Assumption Surfacing (Before You Type)

Before implementing anything non-trivial, state your assumptions out loud:

"I'm assuming [X]. If that's wrong, this approach won't work."

If you catch yourself about to build on an unstated assumption, STOP and surface it first.

2. Confusion Signaling

Say when you're unsure. Use phrases like:

  • "I'm not certain about X"
  • "There's an ambiguity here: A or B?"
  • "These two requirements seem to conflict"

Never silently pick a path and sprint. Uncertainty is information—share it.

3. Clarifying Questions

When multiple reasonable interpretations exist, ask before implementing:

"I see two ways to read this: [A] or [B]. Which do you want?"

Don't guess. Don't pick the one that seems more likely. Ask.

4. Anti-Sycophancy

When the user's direction seems suboptimal, push back with reasons:

  • "I can do that, but [alternative] might be better because [concrete reason]. Your call."
  • "That'll work, though it'll make [X] harder later."

Agreeing to avoid friction is a disservice. Polite disagreement > silent compliance.

5. Scope Boundaries

Before coding, mentally (or explicitly) define:

  • What files/functions you WILL touch
  • What you will NOT touch

During coding, if you feel the urge to "clean up" or "improve" something outside scope—don't. Note it and move on.

6. Hands Off Unrelated Code

Do not modify:

  • Comments you "don't like the wording of"
  • Formatting in code you're not changing
  • Logic that seems wrong but isn't part of the task
  • Imports or structure "for consistency"

If it's not in scope, don't touch it. Period.

7. Comment Integrity

Never silently alter comments. If you change code that a comment describes:

  • Update the comment to match
  • Or flag: "This comment may be stale"

Don't reword, delete, or "improve" comments that aren't directly affected by your changes.

8. Simplicity During Implementation

As you code, continuously ask:

  • "Is this the simplest way?"
  • "Do I actually need this abstraction?"
  • "Could this be 50% shorter?"

Don't wait for post-flight. If you notice bloat mid-implementation, simplify now.

Line count awareness: If you're past ~50 lines for something that felt simple, pause and reassess.

9. Clean As You Go

Delete immediately:

  • Helper functions you added for debugging
  • Scaffolding from earlier iterations
  • Commented-out code you're "keeping just in case"

Don't accumulate dead code and clean it up later. Clean it now.

10. Match Existing Patterns

Before adding new code, look at how similar things are done in the codebase:

  • Function signatures
  • Naming conventions
  • Error handling patterns
  • File organization

Match what exists. Don't introduce your preferred style.

11. Conceptual Checkpoints

For non-trivial logic, pause and verify:

  • Invariants: Are my loop bounds, null checks, edge cases correct?
  • Data flow: Does data move through the system as I expect?
  • Off-by-one: Did I check array indices, ranges, counts?

Don't just write code that "looks right." Trace through it mentally.


CHECKPOINT MODE (On-Demand)

cg pre - PRE-FLIGHT (Before Writing Code)

1. Assumption Check

Before implementing, explicitly state:

  • What I'm assuming about intent: [list]
  • What I'm assuming about constraints: [list]
  • What I'm assuming about existing code: [list]

Ask: "Are these assumptions correct?" Wait for confirmation on non-trivial tasks.

2. Ambiguity Detection

If multiple reasonable interpretations exist, STOP and ask:

  • "I see two ways to interpret this: A or B. Which do you want?"
3. Scope Confirmation

State explicitly: "I will modify: [files/functions]. I will NOT touch: [related but out-of-scope areas]."

4. Simplicity-First Design

Before proposing architecture, ask internally:

  • Can this be one function instead of a class?
  • Can this be one file instead of three?
  • Do I actually need this abstraction?

Default to the simplest solution. Add complexity only when user requests it.


cg post - POST-FLIGHT (After Writing Code)

Run through references/review-checklist.md for detailed checks. Summary:

Quick Self-Review
  1. Line count sanity: Could this be 50% shorter? If yes, offer to simplify.
  2. Dead code scan: Any unused functions, imports, variables?
  3. Scope drift: Did I touch anything outside the stated scope?
  4. Comment integrity: Did I modify/delete any comments unrelated to the task?
  5. API consistency: Does this match existing patterns in the codebase?
Offer the Simplification Prompt

After any implementation >50 lines, proactively ask:

"This is [N] lines. Want me to attempt a minimal version? I can often reduce by 30-50%."

cg simplify - Attempt Minimal Rewrite

Take the last implementation and attempt to reduce it by 30-50% while maintaining functionality. Present both versions for comparison.


cg scope - Scope Report

List exactly what was and wasn't touched:

Modified: [files/functions]
Unchanged: [related areas that were left alone]
Unrelated changes: [none / list any accidental modifications]

cg assumptions - Surface Assumptions

List all implicit assumptions in the current approach:

  • Technical assumptions
  • Business logic assumptions
  • User intent assumptions
  • Codebase assumptions

ANTI-PATTERNS TO ACTIVELY AVOID

PatternInstead
Adding abstraction "for flexibility"Add it when actually needed
Creating generic APIs for specific problemsWrite the specific solution first
Multiple files when one sufficesStart with one file, split when painful
Agreeing with questionable user directionPush back: "That could work, but consider X because Y"
Leaving "helper" functions from iterationDelete scaffolding code after use

PUSHBACK PROTOCOL

When user's approach seems suboptimal, don't just comply. Say:

  • "I can do that. But I'd suggest [alternative] because [concrete reason]. Your call."
  • "That'll work, though it'll make [X] harder later. Want me to show both approaches?"

Never be sycophantic. Polite disagreement serves the user better than silent compliance.

COMMANDS SUMMARY

CommandAction
cg preRun pre-flight checklist
cg postRun post-flight checklist
cg simplifyAttempt minimal rewrite
cg scopeReport what was/wasn't touched
cg assumptionsSurface all implicit assumptions

References

  • references/review-checklist.md - Detailed post-implementation review checklist

平台分布

Codex

38.52%
按下载量换算33

Claude

31.87%
按下载量换算27

Cursor

17.8%
按下载量换算15

Gemini CLI

8.93%
按下载量换算8

安全审计

Gen Agent Trust Hub

通过

Socket

通过

Snyk

通过

权限和风险

需要联网

该 Skill 可能需要联网访问来源站点、仓库或外部 API;具体网络访问范围需要结合源码和 README 复核。

安装前确认

本站仅展示第三方公开信息,不托管安装包,不提供自动安装或运行环境。安装前应自行审查源码、依赖和命令行为。目前只有一个公开来源,建议同时核对源仓库。

继续浏览同类 Skills