代理定义审核员
- 作者仓库星标 0
- 作者仓库 skills-registry
- 领域
- 安全
- 作者声明的智能体
- 未找到明确声明;不据此推断已兼容或已测试
- 静态检查
- 88 / 100 · 启发式扫描,不代表运行安全
- 运行环境证据
- 未执行;当前没有隔离运行测试证据
- 作者 / 版本 / 许可证
- @tomevault-io · 未声明许可证
- 上下文用量评级
- 低消耗
- 接入复杂程度
- 即装即用
- 是否需要外部接口密钥
- 不需要
- 兼容的系统
- 未声明(默认跨平台)
- 运行环境要求
- 未检出明确要求
- 检测到的文件与系统行为
-
- 只读
- 允许写入 / 修改
- 网络行为
- 仅限本地
- 安装命令数
- 无(仅作为资料)
需要注意: 未声明工具权限范围;实际权限由运行环境决定。
# Agent Definition Auditor
Use this skill when the job is auditing one agent-definition file, not rewriting it.
## When to use
- The user wants to audit or score an `AGENTS.md`, `CLAUDE.md`, `SKILL.md`, `SOUL.md`, `.cursorrules`, system prompt, or other agent-definition markdown.
- The user wants a cold-reader review of agent-contract quality.
- The user wants a deploy / revise / do-not-ship judgment with concrete findings and fastest improvements.
- The user wants a reusable rubric-based critique of an agent instruction file rather than an output benchmark.
## When not to use
- The user wants to write, rewrite, or refactor the agent file rather than score it.
- The user wants to evaluate agent outputs, benchmark runtime behavior, or compare model responses rather than audit the instruction artifact itself.
- The job is code review, product-doc review, or general technical writing feedback unrelated to an agent-definition file.
- No target artifact or pasted instruction text is available.
## Non-negotiables
- Default to a single-artifact, read-only cold read.
- Judge the document as an operational contract, not as domain strategy or business policy.
- Distinguish precise layered contracts from phantom context.
- Apply the scoring model, hard caps, and report shape from the reference prompt literally.
- Cite line numbers when available; otherwise cite exact headings or short anchors.
- Do not widen to repo-wide inspection unless the user explicitly asks for multi-file or repo-context evaluation.
- Do not rewrite the file unless the user explicitly asks for fixes after the audit.
## First move
1. Resolve one target artifact path or pasted instruction text.
2. Read `references/agent-definition-judge-prompt.md`.
3. Read the full target artifact once before scoring.
4. If the user also wants revisions, finish the audit first unless they explicitly want authoring instead of judging.
## Workflow
1. Confirm the artifact class and whether the input is a full file or an excerpt.
2. Stay on the single document by default.
3. Inventory the artifact's hard rules:
- mission and scope
- output rules
- tool rules
- approval and escalation rules
- failure handling
- external dependencies
4. Apply the reference prompt's scoring model and hard caps.
5. Return the exact report sections required by the reference prompt.
6. If the user then wants fixes, use the audit as the basis for the follow-up instead of recomputing the diagnosis from scratch.
## Output expectations
- Return the exact top-level sections from the reference prompt.
- Include overall and uncapped score, verdict, confidence, cap logic, findings, strengths, fastest score gains, and limits.
- Keep findings ordered by severity and tied to concrete evidence from the artifact.
## Reference map
- `references/agent-definition-judge-prompt.md` - full judging prompt, scoring model, hard caps, report contract, and examples
---
> Source: [aelaguiz/arch_skill](https://github.com/aelaguiz/arch_skill) — distributed by [TomeVault](https://tomevault.io).
<!-- tomevault:4.0:skill_md:2026-05-23 -->
章节与文件索引
按说明文件中的顺序排列。
何时使用
何时不使用
不容谈判
第一步
工作流程
文件/命令 -> AGENTS.md · CLAUDE.md · SOUL.md · .cursorrules · references/agent-definition-judge-prompt.md
内容校验值 -> 08fea40981c6
当前原文的自动摘译,尚未人工审校。
何时使用
用户想要审核或评分
AGENTS.md、CLAUDE.md、SKILL.md、SOUL.md、.cursorrules、系统提示或其他代理定义标记。用户希望对代理合同质量进行冷读审查。用户希望做出部署/修改/不发货的判断,并提供具体的发现和最快的改进。用户想要对代理指令文件进行可重用的基于评分标准的批评,而不是输出基准。何时不使用
用户想要编写、重写或重构代理文件而不是对其进行评分。用户想要评估代理输出、基准运行时行为或比较模型响应,而不是审核指令工件本身。这项工作是代码审查、产品文档审查或与代理定义文件无关的一般技术写作反馈。没有可用的目标工件或粘贴的说明文本。
不容谈判
默认为单个工件、只读冷读。将文档视为运营合同,而不是域策略或业务策略。将精确的分层合约与虚拟上下文区分开来。从字面上应用参考提示中的评分模型、硬上限和报告形状。引用可用的行号;否则引用准确的标题或短锚点。
第一步
解析一个目标工件路径或粘贴的说明文本。阅读
references/agent-definition-judge-prompt.md。在评分之前阅读完整的目标工件一次。如果用户也想要修改,请先完成审核,除非他们明确想要创作而不是判断。工作流程
确认工件类别以及输入是完整文件还是摘录。默认情况下保留单个文档。盘点工件的硬规则:任务和范围输出规则