独立代码审计与风控评测实验室

Agent 技能:深度解剖、安全管控与实测选型。

不收录泛滥的 Prompt 包装壳。我们通过双遍源码静态分析解剖开源 Agent 技能,测定 L1~L5 宿主安全边界,并发布真实可复现的生产级测评推荐。

564 564 款全量源码审计
100% 100% R2 确定性镜像
L1–L5 L1~L5 宿主权限矩阵
Zero 纯静态构建 / 零数据外发

代表性技能深度解剖与安全管控账本

基于 GitHub 真实开源数据、静态 AST 审计结论与宿主安全边界的深度评测卡。
8 Benchmark Cases
guizang-social-card-skill @op7418/guizang-social-card-skill · Commit cf4b810
★ 7,045 GitHub Stars AGPL-3.0 L4 高权限
guizang-social-card-skill Verified Output
Architectural Analysis & Mechanism Verdict

把版式做成可测量工程:种子模板 + 编号配方(M01-M16/S01-S12)+ 自带 Playwright 校验器(R1-R9),比凭感觉排版的图形类 skill 可复现得多。

Host Security Governance & Runtime Boundaries
Host Privilege: L4 · L4 高权限
Scripts Executed: 5 个辅助子进程
Network Egress: 4 个外呼网络端点
Runtime Policy: 分级授权与审计
baoyu-infographic @JimLiu/baoyu-skills · Commit 1567581
★ 25,926 GitHub Stars MIT L4 高权限
baoyu-infographic Verified Output
Architectural Analysis & Mechanism Verdict

把信息图拆成『结构(布局)× 审美(风格)』两个正交维度并给出 21×22 份可拼装定义,组合自由度远高于固定模板类工具。

Host Security Governance & Runtime Boundaries
Host Privilege: L4 · L4 高权限
Scripts Executed: 1 个辅助子进程
Network Egress: 2 个外呼网络端点
Runtime Policy: 分级授权与审计
khazix-writer @KKKKhazix/khazix-skills · Commit 48e8ba5
★ 20,687 GitHub Stars MIT L1 基础
khazix-writer Verified Output
Architectural Analysis & Mechanism Verdict

风格规格可执行:把「像人写的」拆成节奏/故意打破/知识随手掏/句式断裂/回环呼应等具体手法,并配 12 类示例库。

Host Security Governance & Runtime Boundaries
Host Privilege: L1 · L1 基础
Scripts Executed: 0 个辅助子进程
Network Egress: 0 个外呼网络端点
Runtime Policy: 登记来源即可
design-taste-frontend @leonxlnx/taste-skill · Commit ccbc156
★ 84,793 GitHub Stars MIT L1 基础
design-taste-frontend Verified Output
Architectural Analysis & Mechanism Verdict

反 slop 机制可操作:不是空泛『要有品味』,而是逐条可数的禁令(em-dash 全禁、eyebrow 配额公式、zigzag/hero 结构上限)与机械式 Pre-Flight 矩阵。

Host Security Governance & Runtime Boundaries
Host Privilege: L1 · L1 基础
Scripts Executed: 0 个辅助子进程
Network Egress: 0 个外呼网络端点
Runtime Policy: 登记来源即可
scenes-gathered-zine @Zeejay0/gathered-scenes-zine-skill · Commit b9edb83
★ 4,918 GitHub Stars Gathered Scenes Zine Pers... L3 联网
scenes-gathered-zine Verified Output
Architectural Analysis & Mechanism Verdict

把『好看』拆成了可执行的决策管线:Scene Card → Abstraction Map → 五种插画语法 → 色彩集成模式 → 纸边工艺,每步输入输出与硬禁项明确。

Host Security Governance & Runtime Boundaries
Host Privilege: L3 · L3 联网
Scripts Executed: 0 个辅助子进程
Network Egress: 2 个外呼网络端点
Runtime Policy: 声明域名与密钥范围
youtube-downloader @daymade/claude-code-skills · Commit 63a65c1
★ 1,378 GitHub Stars MIT L5 严管
youtube-downloader Verified Output
Architectural Analysis & Mechanism Verdict

反爬故障知识高度工程化:nsig/PO token/403/'Only images'/SSL EOF 等坑被编码成错误分类与自动重试链,脚本自愈能力强。

Host Security Governance & Runtime Boundaries
Host Privilege: L5 · L5 严管
Scripts Executed: 3 个辅助子进程
Network Egress: 6 个外呼网络端点
Runtime Policy: 隔离环境与严格审计
grill-me @mattpocock/skills · Commit 3cca18b
★ 253,948 GitHub Stars MIT L1 基础
grill-me Verified Output
Architectural Analysis & Mechanism Verdict

可组合的薄壳设计:grill-me 用一行正文把 grilling 原语变成用户入口,同一原语被多个工程 skill 复用,行为单一来源。

Host Security Governance & Runtime Boundaries
Host Privilege: L1 · L1 基础
Scripts Executed: 0 个辅助子进程
Network Egress: 0 个外呼网络端点
Runtime Policy: 登记来源即可
baoyu-comic @JimLiu/baoyu-skills · Commit 1567581
★ 25,926 GitHub Stars MIT L4 高权限
baoyu-comic Verified Output
Architectural Analysis & Mechanism Verdict

风格空间大且可控:6 art × 7 tone × 7 layout × 5 preset,且 preset 带额外叙事规则(ohmsha 的隐喻要求、four-panel 的起承转合)。

Host Security Governance & Runtime Boundaries
Host Privilege: L4 · L4 高权限
Scripts Executed: 1 个辅助子进程
Network Egress: 0 个外呼网络端点
Runtime Policy: 分级授权与审计

五档安全管控矩阵 (L1–L5 Governance)

开源 AI Agent 技能宿主执行权限与风控隔离标准。
了解风控评测规范 →
L1 Pure Instruction 164 Skills

Prompt-only specifications with zero scripts, zero network egress, and zero file access. Completely safe for any host runtime.

Standard registration · Safe for all
L2 Workspace I/O 45 Skills

Reads or writes files within local project directory. Sandboxed workspace execution recommended.

Confined project boundary
L3 Network Read 47 Skills

Queries public HTTP/HTTPS documentation, search endpoints, or APIs. Requires egress monitoring for privacy.

Outbound network audit required
L4 Script Execution 277 Skills

Executes Python, Node.js, or Bash automation scripts. All helper processes verified by static AST code audits.

Subprocess review & authorization
L5 High-Privilege 31 Skills

Interacts with system credentials, tokens, or reverse-engineered services. Requires explicit administrative approval.

Mandatory administrative audit
进入目录查看全量 564 款技能 →